<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Steven Muskal on Medium]]></title>
        <description><![CDATA[Stories by Steven Muskal on Medium]]></description>
        <link>https://medium.com/@smuskal?source=rss-7c2cfeb3b1e7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*eIH3Op0iRIthMYTo3di06w.png</url>
            <title>Stories by Steven Muskal on Medium</title>
            <link>https://medium.com/@smuskal?source=rss-7c2cfeb3b1e7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 15:22:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@smuskal/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[The AI Thesis Everyone Is Getting Wrong]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/illumination/the-ai-thesis-everyone-is-getting-wrong-6f2da9df5954?source=rss-7c2cfeb3b1e7------2"><img src="https://cdn-images-1.medium.com/max/1498/1*_kX4xkyKji0lRVlwHSk_og.png" width="1498"></a></p><p class="medium-feed-snippet">The models are good enough. The next gains come from orchestrating them, and running a trained model costs pennies on the dollar compared&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/illumination/the-ai-thesis-everyone-is-getting-wrong-6f2da9df5954?source=rss-7c2cfeb3b1e7------2">Continue reading on ILLUMINATION »</a></p></div>]]></description>
            <link>https://medium.com/illumination/the-ai-thesis-everyone-is-getting-wrong-6f2da9df5954?source=rss-7c2cfeb3b1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/6f2da9df5954</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[computer-science]]></category>
            <dc:creator><![CDATA[Steven Muskal]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 20:57:53 GMT</pubDate>
            <atom:updated>2026-11-12T20:57:53.223Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Who Splashed the Puddle?]]></title>
            <link>https://ai.plainenglish.io/who-splashed-the-puddle-356d16eb93ec?source=rss-7c2cfeb3b1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/356d16eb93ec</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[stock-market]]></category>
            <dc:creator><![CDATA[Steven Muskal]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 00:44:59 GMT</pubDate>
            <atom:updated>2026-11-11T06:24:04.977Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Three weeks after “The Moat Is a Puddle,” the trigger showed up wearing a security clearance problem.</em></p><p>On October 20 I published <a href="https://medium.com/generative-ai/the-moat-is-a-puddle-1e17f362d903">The Moat Is a Puddle</a>. The argument was simple: a frontier model is a file of numbers, distillation and open weights make that file cheap to reproduce, and a third of the S&amp;P 500 is leaning on a buildout thesis that assumes otherwise. I ended by saying I did not know the trigger or the quarter, only the direction.</p><p>Nineteen days later, NVIDIA’s board opened an investigation into its own chief executive.</p><h3>The Weekend It Broke</h3><p>On November 8, Jensen Huang told employees he had been in a relationship with a businesswoman whom U.S. officials had determined was working for Chinese intelligence. The New York Times reported that Secretary of State Marco Rubio had helped draft a memo about the entanglement, and that officials were tracking it by the spring. NVIDIA’s board says it has no reason to believe company secrets or national security were compromised. Congress wants to ask anyway.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H7tHylTK_wSseg7Z8JHI-A.png" /></figure><p>Whether anything left the building is for the investigators. For the market, the most important person in the AI supply chain is now a security question, and markets do not price security questions kindly.</p><h3>February Already Told Us</h3><p>This did not start in November. In February, Reuters reported that DeepSeek’s next flagship model had been trained on Blackwell, the NVIDIA chip that export controls exist to keep out of Chinese labs. Two days later Reuters reported that DeepSeek was withholding that model from U.S. chipmakers, breaking the usual practice of sharing new models with hardware partners for tuning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w501a-FVouihONxxwk9PMA.png" /></figure><p>Put February next to November. The restricted chips got there. The CEO’s personal life drew a State Department memo. Two separate stories describe the same leaky perimeter, from the silicon to the corner office.</p><p>And the perimeter is guarding the wrong object. Export controls meter the hardware used to make a model. The model itself, once made, is a weight file that fits on a thumb drive. A 70 billion parameter model at four-bit precision is about 40 GB. You can put that in a shirt pocket, upload it in an afternoon, or, as DeepSeek, Qwen and Kimi keep demonstrating, publish it to the world for free. I covered the distillation mechanics in the first piece. The point here is that every dollar spent defending the chip is a dollar spent defending a door the model can walk out of anyway.</p><h3>What the Tape Says</h3><p>NVDA closed at $186.64 on Friday, November 6. It closed at $178.67 on Monday and traded near $168 today, down roughly 10% in two sessions. Over the same two sessions the Nasdaq Composite rose about 1.6%. When a stock falls hard while its own index climbs, the market is pricing something specific to that company. From its October 6 close of $239.24, NVDA is down about 30%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xYTAGxLj9WLLVKbjOoha1g.png" /><figcaption>NVDA daily close, May to November 2026. Data: Yahoo Finance, pulled November 10, 2026.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7lzbgaql8SNtQxD5tIxzxg.png" /><figcaption>NVDA against three counterparties in the AI compute loop and the Nasdaq Composite, rebased to 100 six months ago. Data: Yahoo Finance.</figcaption></figure><h3>A Loop, Not a Market</h3><p>NVIDIA does not just sell chips anymore. It takes equity in the labs that buy its chips. The labs sign multi-year compute contracts with the clouds. The clouds book that as backlog and buy more chips. Oracle and Microsoft report revenue from the same labs NVIDIA has invested in. Draw the arrows and they close into a circle: vendor money funds customer purchases, customer purchases support the vendor’s valuation, and the valuation funds more vendor money.</p><p>I know that shape. In 2007 it was mortgages bundled into securities, the securities bundled again, every layer rated on the assumption that the layer beneath it would keep paying. Nobody had to commit fraud for that structure to fail. It only needed the underlying asset to be worth less than the paper said. Here the underlying asset is demand for ever-larger training runs, and the first piece made the case that the demand curve is flattening while the cost of copying the output falls to zero.</p><p>A loop like this does not need a balance-sheet event to unwind. It needs investors to start asking who owes whom. A spy story about the man at the center of the loop is a very good way to get them asking.</p><h3>Oppenheimer Was a Known Risk Too</h3><p>General Leslie Groves knew about J. Robert Oppenheimer’s Communist associations in 1943: his brother Frank, his wife Kitty, his former lover Jean Tatlock. Groves cleared him anyway, because the bomb needed him more than security needed the rules. Eleven years later, with the war won and the politics changed, the Atomic Energy Commission held a hearing and stripped his clearance. The facts had not changed much. The tolerance had.</p><p>The parallel is the pattern: a government tolerates a known risk around an indispensable technical figure, until a public moment turns tolerated into intolerable, and policy follows. A memo that tracked the risk “by this past spring” is the 1943 file. The congressional letters this week are the start of the hearing.</p><h3>Running a Model Is Not Training It</h3><p>The buildout case rests on several assumptions, and one of them collapses once you separate training from inference: that serving AI to the world takes the same kind of capacity it took to make it. This is not true. Training reads the corpus, runs a forward and a backward pass over every token, and updates the weights. Inference reuses the finished weights. It reads the prompt and generates the answer one token at a time, one forward pass per token, and it never touches the training data again.</p><p>Take Meta’s 405 billion parameter Llama 3, trained on 15.6 trillion tokens. Six times N times D gives about 38 trillion trillion operations, matching what Meta reported. A 1,000-token request costs 2 × 405 billion × 1,000, about 810 trillion. One training run equals 46.8 billion of those requests.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dohfdttCcL10Lx6_f5RzLQ.png" /><figcaption>Training versus one request for Llama 3 405B. Calculation from Meta’s disclosed inputs.</figcaption></figure><p>Meta serves that model on 16 H100s and trained it on as many as 16,384. Each request is cheap. Google measured its median Gemini text prompt at 0.24 Wh, and Microsoft researchers estimate 3.91 Wh even for a long reasoning query.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ekE6t41klOxDe5i0f3jrYg.png" /><figcaption>Requests and days needed for serving to match a 30 GWh training run.</figcaption></figure><p>Serving does overtake training, at modest power. At 1 Wh a request and a billion requests a day, it passes a 30 GWh training run in a month while drawing about 42 MW. To fill 10 GW with 10 billion tasks a day, each task has to average 24 Wh, a hundred times Google’s median prompt. That is agents and video generation, not people asking questions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2E1GW3_9AXVy8_D0c8nOvQ.png" /><figcaption>U.S. generative AI use, third quarter of 2026. Source: Real-Time Population Survey via FRED.</figcaption></figure><p><strong>The heavy usage is not here yet either</strong>. 60.8% of working-age Americans have used generative AI, but only 14.4% use it every day, and it assists 5.3% of all work hours. Adoption is broad, intensity is low, and intensity is what fills a data center.</p><h3>Behind the Brick Wall</h3><p>The pitch for hyperscale AI has always asked enterprises to send their most sensitive material into somebody else’s data center. Drug structures, patient records, legal strategy, source code. That trade was easier when the supply chain looked like a fortress. It looks different this week.</p><p>Meanwhile the alternative has quietly become ordinary. A Mac mini with enough unified memory runs a capable open model on a desk, with no network cable required. Inference, the part of AI that runs every time someone uses it, does not need a training cluster. In my own work, keeping proprietary chemistry and biology off shared infrastructure was always a requirement. It is about to become everybody’s requirement.</p><p>The shift is already measurable. In Broadcom’s 2026 survey of 1,800 IT leaders, 56% run or plan to run production inference on private cloud against 41% on public cloud, after a roughly even split a year earlier. Gartner expects more than half the PCs shipped this year to carry dedicated AI hardware. Apple now ships a 20 billion parameter model that runs on the device itself.</p><p>I expect three things to move faster now. Export enforcement will shift from chip specifications toward end-use verification and people. Enterprise and government buyers will put data sovereignty ahead of raw scale, and air-gapped or near-air-gapped deployments will stop sounding paranoid. And capital allocators will start discounting circular financing, rewarding inference efficiency and private deployment over the next gigawatt of training capacity.</p><p>None of that requires the models to get worse. They are already remarkable. It only requires the market to stop believing that training scale is a moat, at the moment a security scandal makes concentrating everything in one supply chain look like what it is.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XyuOGyonGLYmhcViiyvVLg.png" /></figure><h3>References</h3><ul><li>S. Muskal, “The Moat Is a Puddle,” Generative AI on Medium, October 20, 2026: <a href="https://medium.com/generative-ai/the-moat-is-a-puddle-1e17f362d903">medium.com</a></li><li>The New York Times, November 8, 2026 (two reports, linked above); BBC, November 9; CNBC, November 8; CBS News, November 10.</li><li>Reuters, February 24 and February 26, 2026 (linked above).</li><li>F. Oviedo et al., “Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute,” Joule: <a href="https://arxiv.org/abs/2509.20241">arXiv:2509.20241</a></li><li>Google, “Measuring the environmental impact of delivering AI at Google scale,” 2025: <a href="https://arxiv.org/abs/2508.15734">arXiv:2508.15734</a></li><li>Meta, “The Llama 3 Herd of Models,” 2024: <a href="https://arxiv.org/abs/2407.21783">arXiv:2407.21783</a>; Llama 3.1 model card: <a href="https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/MODEL_CARD.md">github.com</a></li><li>Real-Time Population Survey, generative AI series, Q3 2026: <a href="https://fred.stlouisfed.org/series/RPSGENAIUSAGESHAREALL">FRED</a></li><li>Apple Foundation Models, third generation, June 2026: <a href="https://machinelearning.apple.com/">machinelearning.apple.com</a></li><li>Broadcom, Private Cloud Outlook 2026: <a href="https://news.broadcom.com/releases/broadcom-private-cloud-outlook-2026">news.broadcom.com</a></li><li>Gartner, “AI PCs Will Represent 31% of Worldwide PC Market by the End of 2025,” August 28, 2025: <a href="https://www.gartner.com/en/newsroom/press-releases/2025-08-28-gartner-says-artificial-intelligence-pcs-will-represent-31-percent-of-worldwide-pc-market-by-the-end-of-2025">gartner.com</a></li><li>Price data: Yahoo Finance via yfinance, daily closes through November 10, 2026.</li><li>Oppenheimer security hearing, 1954: <a href="https://en.wikipedia.org/wiki/Oppenheimer_security_hearing">Wikipedia</a></li></ul><p>This piece is opinion and analysis, not investment advice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gcW0_c7FYu35aA9O.png" /></figure><p>Steven Muskal, Ph.D. is the CEO of <a href="https://www.eidogen-sertanty.com/">Eidogen-Sertanty, Inc.</a> — a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at <a href="http://stevenmuskal.com/">stevenmuskal.com</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=356d16eb93ec" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/who-splashed-the-puddle-356d16eb93ec">Who Splashed the Puddle?</a> was originally published in <a href="https://ai.plainenglish.io/">Artificial Intelligence in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Delta Is the Dose]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/illumination/the-delta-is-the-dose-fb1f01cc9ced?source=rss-7c2cfeb3b1e7------2"><img src="https://cdn-images-1.medium.com/max/1200/0*vetRmymMQsJMZUZU.png" width="1200"></a></p><p class="medium-feed-snippet">What a Year of Cold Plunging Taught Me as I Turn Sixty</p><p class="medium-feed-link"><a href="https://medium.com/illumination/the-delta-is-the-dose-fb1f01cc9ced?source=rss-7c2cfeb3b1e7------2">Continue reading on ILLUMINATION »</a></p></div>]]></description>
            <link>https://medium.com/illumination/the-delta-is-the-dose-fb1f01cc9ced?source=rss-7c2cfeb3b1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/fb1f01cc9ced</guid>
            <category><![CDATA[health]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[fitness]]></category>
            <dc:creator><![CDATA[Steven Muskal]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 00:32:52 GMT</pubDate>
            <atom:updated>2026-11-09T07:17:12.746Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Bring Them Forward]]></title>
            <link>https://medium.com/age-of-awareness/bring-them-forward-7b0f15981ee0?source=rss-7c2cfeb3b1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b0f15981ee0</guid>
            <category><![CDATA[relationships]]></category>
            <category><![CDATA[aging]]></category>
            <category><![CDATA[mentorship]]></category>
            <category><![CDATA[life-lessons]]></category>
            <category><![CDATA[education]]></category>
            <dc:creator><![CDATA[Steven Muskal]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 16:43:44 GMT</pubDate>
            <atom:updated>2026-11-06T16:43:44.296Z</atom:updated>
            <content:encoded><![CDATA[<p>Celebration is the frosting. Mentorship is the cake.</p><p>By Steven Muskal, Ph.D. | The Renaissance Circle | November 2026</p><p>I previously wrote about <a href="https://medium.com/@smuskal/274725d89646">Toast Our Friend</a>, a site I built so friends and family could post stories and photos for someone and let that person read them. It began with a simple observation at my sister Julie’s celebration of life: she would have loved hearing those stories. The first toasts went to friends facing the end, and a few of them arrived too late to be read.</p><p>That piece was about where the idea came from. This one is about where it is going.</p><h3>Circles</h3><p>The <a href="https://toastourfriend.com/">toastourfriend.com</a> (ToF) site is a “pay it forward,” no-cost site I created for my network of friends and family. I have shared it with others. It is not a business, and unlike the social media giants I have absolutely zero intention (or need) to monetize on other people’s celebratory content. ToF enables what we call circles: pages where anyone can gather toasts for one person. You can keep them private or open, 100% your choice. A birthday, a milestone, a reunion, or no reason at all. The idea is simple: make it easy to say something true and specific about a person, and easy for that person to see it while it still counts. <strong>This stays out of the hands of social media giants</strong>. I have way too much integrity for that — and wish one day those giants will pivot towards doing something good for society, but that is the subject for another time. I’ll keep ToF going as long as possible, but will probably GitHub the infrastructure so people can DIY their own if they like.</p><p>In ToF you can download and export your content at any time — which is something I always encourage. This was the key message in my <a href="https://medium.com/age-of-awareness/that-go-there-62a93ef456cd">That Go There</a> write-up. Content storage systems change over the years (yes, one day Facebook and YouTube will disappear), so be sure to <strong>always transfer your memories to the next “more durable” system</strong>…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/0*sTW7vSWXaD1DxuPz.jpeg" /><figcaption>That Go There - Why the memories worth keeping need a body that can move — <a href="https://medium.com/age-of-awareness/that-go-there-62a93ef456cd">https://medium.com/age-of-awareness/that-go-there-62a93ef456cd</a> and <a href="https://www.drstevenmuskal.com/p/that-go-there">https://www.drstevenmuskal.com/p/that-go-there</a></figcaption></figure><p>The real pivot with ToF has been toward celebrating people who are here to enjoy it, and it started with Sung-Hou Kim. <a href="https://toastshk.com/">Toast SHK</a> gathers stories from alumni of the Kim group lab, decades of chemistry graduate students, postdocs, and staff scientists, for the mentor who shaped so many of our careers. And he gets to read them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/374/1*70Sgmlt4oqTcfX9fUZaJqA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fNCALexXd_w4NtA_.png" /></figure><p>In July I created a ToF circle for my wife Kirsten’s 60th birthday. A circle full of stories from the people who know her. She read them as they were posted, as did others who had previously posted. Her birthday day became a birthday month. She was floating on air every day that month, and I’m sure she periodically returns to them for a quick smile or laugh. That is the whole point of the thing. Indeed, apparently her family and friends all think she is funny — which is why I often refer to her as “funny lady” (even with a sarcastic tone at times).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_czQgVYZieJiDy6t.png" /><figcaption><a href="https://toastkirsten.com/">ToastKirsten.com</a></figcaption></figure><p><a href="https://toaststeve.com/">Toast Steve</a> began as my testbed, a page I used to try features before anyone else saw them. Then Kirsten decided it should be real. She started filling it for my own 60th, which is coming up in a few weeks. In tune with past birthdays, this will be a gathering consistent with others involving friends and music in my studio. With <a href="https://toaststeve.com/">toaststeve.com</a> I am now on the receiving end of my own platform, and I recommend it :)</p><h3>If It Bleeds, It Leads</h3><p>Local news has run on a simple formula for decades: if it bleeds, it leads. Fear, outrage, and conflict pull attention more reliably than almost anything else, and the algorithms that choose what crosses our feeds learned the same lesson at far greater scale. Researchers call the underlying wiring negativity bias: bad news registers more strongly than good (<a href="https://www.nature.com/articles/s41562-023-01538-4">Nature Human Behaviour</a>).</p><p>A ToF circle runs on the opposite formula. Nobody creates one to complain about a person. It selects for the thing we are starved of: specific, positive attention, paid to someone while they can feel it land.</p><p>We are coming out of fifteen+ years that trained a generation to doom-scroll and to perform outrage for engagement. <strong>I think we are moving into something more connective, a true age of (in)person-to-people engagement.</strong> I see it in small places: book clubs filling up again, running groups that meet before anyone has coffee, dinners where nobody looks at a phone. All of it points toward rebuilding the ordinary, unplanned contact that third spaces used to provide for free.</p><h3>The Lab, and the Rooms in January</h3><p>This last summer, alumni of the Kim group lab came together for a session. It was 35 years ago for me, some more recent, and some longer ago. Some are professors now. Some run companies. We traded the old stories. It was extremely inspiring for me. I returned from that gathering super-charged and embarked on a campaign to <a href="https://medium.com/towards-artificial-intelligence/revisiting-old-work-8241cf3ba308">revisit old work</a> and turbocharge new projects.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DoHPRjpwGbWBiqb1.jpeg" /><figcaption>Melvin Calvin Laboratory, UC Berkeley</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xD9FFuU5IT82dfrj" /><figcaption>Three legends. One photo. One bench. At our recent reunion on August 1, 2026, in Berkeley, California: Joel Sussman, Sung-Hou Kim and George Church. From pioneering the structures of tRNA and acetylcholinesterase, to advancing structural genomics, DNA sequencing, synthetic biology and genome engineering, their collective contributions have helped shape modern molecular biology. Decades of discoveries that changed how we understand, read and engineer biology. Incredible to have this much scientific history sitting on one bench, and together again at the reunion!</figcaption></figure><p>That is the conversation I am taking on the road this December and late January 2027, with chemistry students at UC Berkeley and separately at Claremont McKenna / Harvey Mudd College. I’ll bring 40+ years of decisions, a few of them good, some of them lucky, and a willingness to say which was which in front of students entering a job market that looks nothing like the one I entered.</p><p>This is the start of a longer list. Any institution that shaped me or my family is fair game for the same offer: an afternoon with your students, a smorgasbord of appropriate stories and interactive discussion. Paying things back, helping to bring people forward. Being philanthropic with my time has been extremely enjoyable and energizing.</p><h3>The Epigenetics of Mentorship</h3><p>I think in chemical and biological terms, so here is the analogy. Transgenerational epigenetic inheritance is the way an organism’s environment and experience change gene expression in ways that can pass to offspring, without changing the DNA sequence itself (<a href="https://www.nature.com/articles/s41576-021-00438-5">Nature Reviews Genetics</a>). It is one of the ways a species carries information forward faster than mutation alone.</p><p>Culture is the version of that mechanism available to a species with language. <strong>A discussion, lecture, a mentorship, a story told at the right moment: these are how experience transmits outside of genes.</strong></p><p>Choosing not to mentor the generation behind you is counter-evolutionary. Most of us are sitting on specific, hard-won knowledge that somebody younger needs right now, and the obligation to hand it over grows as the gap between generations widens.</p><p>And the gap is widening fast. The students I will meet in December and January are entering a labor market being reshaped in real time by AI. Many came of age on platforms built for time on screen rather than connection, while third spaces (the diners, rec centers, and campus quads where people used to just be around each other) quietly shrank. Layer a K-shaped economy on top, where asset holders pull away from everyone else (<a href="https://www.cnbc.com/2026/01/30/wealth-inequality-k-shaped-economy-united-states-consumer-spending-trump.html">CNBC</a>), and you get a generation with real reason to be anxious about what comes next.</p><h3>What Rooster Gets Right</h3><p>One of my new favorite shows this year is Rooster, a comedy-drama that premiered in March 2026. Steve Carell plays a middle-aged author of commercially popular beach-read novels who ends up visiting and lecturing at the New England college where his daughter is a professor.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/494/0*whB7DoSCYreWYBlf" /></figure><p>Set the comedy and the family drama aside and look at what he is doing in those lecture scenes: an older person walking into a room of twenty-year-olds, trading real anecdotes and mistakes for their attention, and learning something back from them. <strong>Paying it forward runs in both directions. I think of it as bringing them forward.</strong></p><h3>The Frosting and the Cake</h3><p>A ToF circle is frosting. It is the light, celebratory, shareable layer that gets people to show up. Nobody opens toastourfriend.com to think about mentorship and epigenetics. They open it because a friend sent a link to say something nice about someone, and that is an easy, pleasant thing to do on a Tuesday. Frosting belongs on birthdays, and so do circles.</p><p>But frosting sits on a cake, and the cake is the slower work: the lab reunion, the plane to talk with undergrads, graduate students, and postdocs, telling someone twenty-five years younger exactly what you got wrong at their age so they do not have to find out the same way. The celebration is the secret sauce that gets people engaged. The substance underneath is what carries forward.</p><p>I turn sixty in a few weeks, with friends in a room that has been designed to share music and conversation (at the same time!). This time with a circle Kirsten started filling before the candles are lit. Sung-Hou got to read his. Kirsten got to read hers. I am reading mine.</p><p>If you are sitting on knowledge somebody younger needs, find the room where you can hand it over. And if there is someone you have been meaning to celebrate, <a href="https://toastourfriend.com/">start your own circle</a>. Feel free to template your own, or make use of the platform I enabled. Regardless, do not wait for an occasion. Celebrate the people in your life today. Take my word for it, being philanthropic with your time is better than any material gift.</p><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/0*-Sd-B72k7EzXk60M.png" /><figcaption><a href="https://toastourfriend.com/">ToastOurFriend</a></figcaption></figure><h3>References</h3><ul><li>Robertson et al., Negativity drives online news consumption, Nature Human Behaviour (2023): <a href="https://www.nature.com/articles/s41562-023-01538-4">nature.com/articles/s41562–023–01538–4</a></li><li>Fitz-James and Cavalli, Molecular mechanisms of transgenerational epigenetic inheritance, Nature Reviews Genetics (2022): <a href="https://www.nature.com/articles/s41576-021-00438-5">nature.com/articles/s41576–021–00438–5</a></li><li>Wealth inequality is worse than ever as K-shaped economy widens, CNBC (January 2026): <a href="https://www.cnbc.com/2026/01/30/wealth-inequality-k-shaped-economy-united-states-consumer-spending-trump.html">cnbc.com</a></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1CCuV5Riad9lm0Wp.jpeg" /></figure><p>Steven Muskal, Ph.D. is the CEO of <a href="https://www.eidogen-sertanty.com/">Eidogen-Sertanty, Inc.</a> — a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at <a href="http://stevenmuskal.com/">stevenmuskal.com</a></p><p>I normally include examples of recent mixes, but this time— I thought I would direct you to scan through any that might resonate with you. It’s a nice collection and growing — <a href="https://www.stevenmuskal.com/music.html">Music Videos</a> — I’m sure I’ll add a few after my upcoming birthday…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5fon5QSKQz7lt3W5.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b0f15981ee0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/age-of-awareness/bring-them-forward-7b0f15981ee0">Bring Them Forward</a> was originally published in <a href="https://medium.com/age-of-awareness">Age of Awareness</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Toast Our Friend]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/illumination/toast-our-friend-274725d89646?source=rss-7c2cfeb3b1e7------2"><img src="https://cdn-images-1.medium.com/max/1400/0*mhP8nUHnMQJyuw-5.png" width="1400"></a></p><p class="medium-feed-snippet">Celebrating the people who made us &#x2014; before it&#x2019;s too late</p><p class="medium-feed-link"><a href="https://medium.com/illumination/toast-our-friend-274725d89646?source=rss-7c2cfeb3b1e7------2">Continue reading on ILLUMINATION »</a></p></div>]]></description>
            <link>https://medium.com/illumination/toast-our-friend-274725d89646?source=rss-7c2cfeb3b1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/274725d89646</guid>
            <category><![CDATA[life-lessons]]></category>
            <category><![CDATA[family]]></category>
            <category><![CDATA[grief]]></category>
            <category><![CDATA[friendship]]></category>
            <dc:creator><![CDATA[Steven Muskal]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 04:22:55 GMT</pubDate>
            <atom:updated>2026-11-06T04:22:55.726Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Energizing the Machine]]></title>
            <link>https://aiadvances.org/energizing-the-machine-139107ca1be7?source=rss-7c2cfeb3b1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/139107ca1be7</guid>
            <category><![CDATA[energy]]></category>
            <category><![CDATA[biology]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Steven Muskal]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 12:26:35 GMT</pubDate>
            <atom:updated>2026-11-04T12:26:35.643Z</atom:updated>
            <content:encoded><![CDATA[<h4>What Mitochondria Can Teach Us About Powering the Next Generation of Robots</h4><h3>Large batteries are the wrong answer. Nature solved this problem 2 billion years ago.</h3><p>By <strong>Steven Muskal, Ph.D.</strong> | <strong>The Rise of the Machine II</strong>| <a href="http://stevenmuskal.com/">stevenmuskal.com</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YXmTd0TZuA_lXdwi.jpeg" /></figure><p><em>The centralized battery paradigm: hundreds of kilograms of electrochemical cells that define and constrain modern electric vehicles.</em></p><h3>The Crushing Weight of Our Best Ideas</h3><p>Here’s a number that should unsettle you: <strong>480 kilograms</strong>.</p><p>That’s the mass of the battery pack in a Tesla Model S Long Range. Nearly half a metric ton of electrochemical cells bolted to the floor of what we’ve been told is the future of transportation. To put it viscerally: you’re hauling around the weight of a baby elephant every time you drive to the grocery store.</p><p>And it’s not just Tesla. The Rivian R1T pickup carries a 800 kg battery. The GMC Hummer EV? A staggering 1,326 kg of battery alone, more than many entire compact cars weigh. These vehicles have become architectural monuments to a single engineering decision: centralized energy storage.</p><p>The consequences ripple outward with punishing physics. Heavier batteries demand reinforced suspension systems. Reinforced suspensions require beefier brakes. Beefier brakes add mass. More mass requires more energy to accelerate and decelerate, which demands… more battery capacity. We’ve engineered ourselves into a feedback loop where our solution to the energy problem <em>is</em> the energy problem. Recent research has confirmed that EV-induced road damage scales dramatically with axle load, following the well-established fourth-power law<a href="#ref-1"><strong>[1]</strong></a>.</p><p>But here’s what keeps me awake at night, and what I’ve been building toward across this <em>Rise of the Machine</em> series: we’re about to make this same mistake with humanoid robots.</p><p>Boston Dynamics’ Atlas. Tesla’s Optimus. Figure AI’s 01. The next generation of AI-embodied systems. Every one of them faces the same existential constraint: <strong>mass is the enemy</strong>. A robot that must carry a 50 kg battery pack to operate for two hours is a robot that has already lost. Every gram allocated to energy storage is a gram stolen from payload capacity, from actuator power, from the sensors and compute that make intelligence possible.</p><p>What if the entire paradigm is wrong? What if the answer isn’t “better batteries” but <em>no batteries at all</em>, at least not in the way we’ve conceived them?</p><p>Nature solved this problem two billion years ago. And she didn’t use lithium.</p><h3>Section 1: The Mass Problem Scales Catastrophically</h3><p>Let’s sit with the physics for a moment, because the numbers are damning.</p><p>In modern electric vehicles, battery mass typically represents 25–30% of total vehicle weight. The Tesla Model 3’s battery accounts for approximately 28% of its curb weight. For the Hummer EV, batteries approach 35%. These aren’t engineering compromises; they’re structural commitments that cascade through every other design decision.</p><p>Consider what this means mechanically. A 2024 study on hidden externalities of electric vehicles documented that EVs impose significantly greater wear on road surfaces and tire systems precisely because of this concentrated mass<a href="#ref-2"><strong>[2]</strong></a>. The fourth-power law of road damage, which states that wear scales with the fourth power of axle load, means even modest weight increases impose dramatically disproportionate consequences. Suspension components in EVs show accelerated wear patterns. Brake systems require oversizing. Tire compounds must be reformulated to handle the additional loading.</p><p>Now scale this to a humanoid robot.</p><p>A robot operating in a warehouse, a hospital corridor, or alongside humans in a factory faces constraints far more severe than any vehicle. Every additional kilogram is:</p><ul><li>More torque demand on every joint actuator</li><li>Greater inertial load during acceleration and deceleration</li><li>Increased ground reaction forces transmitted through the structure</li><li>Larger actuators required, which themselves add mass</li><li>More energy consumed per movement cycle</li></ul><p>This is the <strong>compounding mass penalty</strong>, and it’s why battery weight isn’t just inconvenient for robots; it’s architecturally fatal.</p><p>Current humanoid prototypes allocate 20–40% of their mass budget to batteries. Some early systems were effectively battery packs with legs. The operational consequences are predictable: limited runtime (often under two hours), restricted payload capacity, and compromised dynamics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*JDcGcY_vCYhb-n02.gif" /></figure><h3>Core Insight</h3><p>In mobile systems, energy storage and structural mass are in zero-sum competition. Every kilogram devoted to storing energy is a kilogram that cannot contribute to capability. Biology never makes this trade-off. And that should tell us something profound.</p><h3>The Caloric Reality Check: Human vs. Machine</h3><p>Before we can solve the power problem, we need to quantify it. Let’s put human and robot energy demands side by side.</p><p>An active adult human consumes approximately 2,000–2,500 kilocalories per day. Converting to physics: 2,000 kcal = 8.4 megajoules, spread across 86,400 seconds, yields roughly 97 watts of continuous average power. During 16 waking hours of movement, cognition, and homeostasis, the effective draw rises to approximately 145 watts average. Peak output during sprinting can reach 1,500–2,000W, but only for seconds.</p><p>From this modest ~100W continuous budget, a 70 kg human walks, thinks, maintains body temperature, repairs tissues, and runs a 20-watt brain. The secret: biological muscles are only ~25% mechanically efficient (75% becomes heat), but they excel at passive holding via tendons (zero energy to maintain position) and elastic energy recovery.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*iUU0CyAVJJlhxBHF.png" /></figure><p>The striking finding: despite being 3–4x more efficient electrically, the robot requires 3–4x more absolute power than a human because electric actuators lack muscles’ passive holding ability, and because the robot has no equivalent of tendons that store and release energy for free. The daily energy budgets are remarkably similar (~8.4 MJ for both), but the human achieves continuous operation through constant metabolic intake while the robot must periodically stop and recharge.</p><p>This equivalence is not a coincidence. It tells us that a 57–70 kg mobile system performing human-equivalent work in Earth’s gravity well requires approximately 8–10 megajoules per operational day, regardless of whether the prime mover is muscle or motor. The question isn’t “how much energy?” but “how do you deliver it?”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SToK9oeOhIB_G0dM.png" /></figure><p>Fluorescence microscopy reveals the intricate internal architecture of living cells. The colored organelle networks visible here include mitochondria, the distributed energy systems refined over two billion years of evolution. <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6747937/"><em>Mitochondria: not merely powerhouses, but distributed, adaptive, intelligent energy management systems refined over two billion years of evolution</em></a><em>. See </em><a href="https://www.google.com/books/edition/Mitochondria_and_the_Future_of_Medicine/GMxJDwAAQBAJ?hl=en&amp;gbpv=1&amp;printsec=frontcover"><em>Mitochondria and future of medicine.</em></a></p><h3>Section 2: What Biology Figured Out</h3><p>If I were to ask you what mitochondria do, you’d probably answer: “They’re the powerhouse of the cell. They make ATP.”</p><p>You’d be right. And you’d be missing almost everything interesting.</p><p>Mitochondria are not simply ATP factories. They are distributed, adaptive, intelligent energy management systems that also serve as signaling hubs, quality control monitors, and, remarkably, executioners. Understanding what mitochondria <em>actually</em> do is the key to understanding what our machines are missing.</p><h3>Two Billion Years Ago: Energy Came First</h3><p>Here is a fact that should reframe your entire conception of how complex life arose: mitochondria were not <em>built</em> by cells. They were free-living bacteria that <em>invaded</em> them.</p><p>Approximately two billion years ago, an alpha-proteobacterium, a free-living energy-producing microbe, was engulfed by a primitive archaeal host cell. Rather than being digested, it survived. A symbiosis formed. The bacterium provided a massive surplus of ATP; the host provided raw materials and protection. Over eons, the bacterium lost its independence, shedding most of its genome into the host nucleus, but it retained its own DNA, its own replication machinery, and its own double membrane, vestiges of that ancient merger.</p><p>This is the endosymbiotic theory, now supported by overwhelming molecular evidence: mitochondria retain their own circular DNA (like bacteria), replicate by binary fission (like bacteria), and are bounded by a double membrane (the inner membrane being the original bacterial membrane). They are, in a very real sense, domesticated bacteria living inside every cell of your body.</p><p>Nick Lane’s landmark work (<em>The Vital Question</em>, 2015; “How energy flow shapes cell evolution,” <em>Current Biology</em>, 2020) establishes the critical causality: the energy source preceded and enabled cellular complexity, not the reverse. Without the 100,000-fold energy surplus that internalized mitochondria provided per gene, eukaryotic cells could never have evolved nuclei, organelles, multicellularity, or any of the machinery we associate with complex life. Every complex organism on Earth, every animal, plant, and fungus, carries mitochondria. There are no exceptions. This is not coincidence; it is causation.</p><p>Two further observations deepen the insight:</p><p><strong>Maternal inheritance:</strong> In humans and most animals, mitochondrial DNA is inherited almost exclusively from the mother. The mitochondria in your cells are descendants of those in your mother’s cells, which were inherited from her mother, and so on through a long maternal lineage. This pattern of inheritance allows mitochondrial DNA to be used to trace maternal ancestry across many generations. The mitochondria present in modern organisms ultimately descend from an ancient symbiotic relationship between an ancestral cell and a bacterium that occurred roughly 1.5 to 2 billion years ago, although mitochondrial genomes and functions have evolved substantially since that time.</p><p><strong>The lesson for machines:</strong> The evolution of complex cells was closely linked to the acquisition of mitochondria, which greatly increased the amount of energy available to support cellular processes. Many researchers view this increase in energy capacity as an important factor enabling the evolution of larger genomes, greater cellular complexity, and multicellular organisms. In contrast, robotics and artificial intelligence development often emphasize advances in computation, sensing, and mechanical capability while relying on incremental improvements in energy storage and power systems. This comparison suggests that breakthroughs in energy generation, storage, and distribution may be as important to future robotic capabilities as advances in hardware and software.</p><h3>Beyond ATP: The Hidden Lives of Mitochondria</h3><p><strong>Calcium signaling and buffering:</strong> Mitochondria actively absorb and release calcium ions, modulating the calcium waves that propagate through cells during signaling events. They’re not passive bystanders to cellular communication; they’re active participants, shaping the temporal dynamics of signals that control everything from muscle contraction to neurotransmitter release.</p><p><strong>Programmed cell death (apoptosis):</strong> Here’s a fact that should reframe your entire conception of these organelles: mitochondria decide when cells die. Through the release of cytochrome c and other pro-apoptotic factors, mitochondria trigger the cascade that dismantles a cell from within. They are quality control systems with executive authority.</p><p><strong>Reactive oxygen species (ROS) as signals:</strong> We’ve long thought of ROS as purely damaging, byproducts of metabolism that cause cellular aging. Recent research reveals a more nuanced picture: mitochondria use controlled ROS production as a signaling mechanism, communicating metabolic state to the rest of the cell.</p><p><strong>Fission and fusion dynamics:</strong> This is where things get truly remarkable. Mitochondria are not static structures. They continuously divide (fission) and merge (fusion) in response to cellular energy demands. When energy demand is high and sustained, mitochondria fuse into elongated networks that maximize efficiency. When demand is localized or the organelles are damaged, they fragment for targeted repair or removal. A 2024 review in <em>Nature Reviews Molecular Cell Biology</em> characterized these molecular mechanisms of mitochondrial dynamics in comprehensive detail<a href="#ref-3"><strong>[3]</strong></a>.</p><p><strong>Physical migration:</strong> Mitochondria move. They travel along cytoskeletal tracks to position themselves precisely where energy is needed. In neurons, mitochondria migrate to synapses during periods of high activity. In migrating cells, they relocate to the leading edge where ATP demand is greatest.</p><h3>The Molecular Turbine</h3><p>At the heart of mitochondrial ATP production sits what I consider the most elegant nano-machine evolution has ever produced: <strong>ATP synthase</strong>.</p><p>This protein complex is literally a rotary motor. The F0 component, embedded in the inner mitochondrial membrane, channels protons down their concentration gradient. This proton flow drives the rotation of a central shaft (the gamma subunit) at speeds approaching 100–150 revolutions per second. The rotating shaft induces conformational changes in the F1 component that catalyze the synthesis of ATP from ADP and inorganic phosphate.</p><p>Think about this: inside your cells, right now, are trillions of molecular turbines spinning hundreds of times per second, powered by proton currents, generating the universal energy currency that allows you to read these words, move your eyes, and form thoughts about what you’re reading.</p><p>A 2023 study in <em>Nature Communications</em> elucidated the actual mechanism by which ATP hydrolysis couples to rotation in bacterial ATP synthase, revealing a sophisticated molecular ratchet mechanism<a href="#ref-4"><strong>[4]</strong></a>. This isn’t metaphorical engineering; it’s literal engineering at nanometer scales, refined over billions of years.</p><h3>Video Resources: ATP Synthase in Action</h3><p>These videos provide outstanding visual reference for the molecular machinery discussed above.</p><ol><li><strong>Molecular Machines — ATP Synthase: The Power Plant of the Cell</strong> (<a href="https://youtube.com/watch?v=XI8m6o0gXDY">youtube.com/watch?v=XI8m6o0gXDY</a>)</li><li>Describes ATP synthase as a “high-tech micromolecular power generator” with a barrel-shaped rotator of 10–15 protein subunits.</li><li><strong>ATP Synthesis | HHMI BioInteractive Video</strong> (<a href="https://youtube.com/watch?v=CSrtewCJbpg">youtube.com/watch?v=CSrtewCJbpg</a>)</li><li>HHMI BioInteractive production showing mitochondria as “electrochemical batteries that convert energy from food and oxygen into ATP.”</li><li><strong>ATP in Use | HHMI BioInteractive Video</strong> (<a href="https://youtube.com/watch?v=vUC5njkOq7o">youtube.com/watch?v=vUC5njkOq7o</a>)</li><li>Shows ATP in action powering cellular machinery: calcium ion pumps, muscle contraction, nerve transmission, gene regulation.</li></ol><h3>The Metabolic Cascade</h3><p>Here’s another number that should reshape your thinking: <strong>4 grams</strong>.</p><p>That’s approximately how much glucose is in your bloodstream at any given moment. Four grams, less than a sugar packet, circulating through roughly five liters of blood. From this seemingly minimal reserve, your body powers every thought, every heartbeat, every breath.</p><p>How? Because the body never stores energy proportional to its demand. Instead, it maintains a sophisticated metabolic cascade:</p><ol><li><strong>Ingestion:</strong> Complex carbohydrates, fats, and proteins enter the gut</li><li><strong>Digestion:</strong> Macronutrients are broken down into absorbable components</li><li><strong>Distribution:</strong> Glucose enters the bloodstream, maintained at tight homeostatic levels</li><li><strong>Cellular uptake:</strong> Glucose enters cells and is processed through glycolysis</li><li><strong>Oxidative phosphorylation:</strong> Pyruvate enters mitochondria, feeding the citric acid cycle and electron transport chain</li><li><strong>ATP synthesis:</strong> Those spinning molecular turbines convert the proton gradient into ATP</li><li><strong>Immediate use:</strong> ATP is consumed within seconds to minutes of its production</li></ol><p>Your body produces approximately <strong>40 kilograms</strong> of ATP per day. But you never <em>store</em> 40 kilograms of ATP. The total ATP content of the human body at any moment is roughly 250 grams. The system works by continuous generation and immediate consumption, not bulk storage.</p><p>This is the design principle we’ve ignored: <strong>flow, not reservoir</strong>.</p><p><em>Supercapacitors and micro-electronic energy storage: the technological building blocks of a distributed power architecture.</em></p><h3>Section 3: The Mitocapacitor Vision</h3><p>Let me introduce a concept I’ve been developing that I call <strong>mitocapacitor architecture</strong>: a distributed energy storage and delivery system for robots that mirrors the biological strategy of mitochondria.</p><p>The foundational technology already exists. It’s called the supercapacitor.</p><h3>The Supercapacitor Difference</h3><p>Supercapacitors (also called ultracapacitors) store energy through electrostatic charge separation rather than electrochemical reactions. This distinction has profound implications:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*RPxcFNETtpWvU6Cn.png" /></figure><p>Look at those numbers carefully. Yes, supercapacitors store far less energy per kilogram. But they deliver energy 10–40 times faster. They can charge in milliseconds. They last through hundreds of thousands of cycles without degradation.</p><p>We’ve been asking the wrong question. The engineering community has been obsessed with: “How do we store more energy in one place?” The biological question is different: “How do we distribute the right amount of energy, instantly, exactly where it’s needed?”</p><h3>Structural Integration</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*T4Qyhn49DvLEiRcF.png" /></figure><p><em>Carbon fiber weave: the distinctive crosshatch pattern of woven carbon fiber composites that can simultaneously serve as load-bearing structure and energy storage medium. Research at Chalmers, Cambridge, and MIT is making this dual function a reality.</em></p><p>Here’s where it gets interesting. Recent research from Cambridge, MIT, and Chalmers University has demonstrated <strong>structural supercapacitors</strong>, carbon fiber composite materials that simultaneously serve as load-bearing structures and energy storage devices.</p><p>A 2024 review on carbon fiber electrodes for composite structural supercapacitors documented preparation and modification methods for creating these dual-function materials<a href="#ref-5"><strong>[5]</strong></a>. Chalmers University of Technology has demonstrated that carbon fiber structural batteries could reduce EV weight by 25 percent while maintaining mechanical integrity<a href="#ref-6"><strong>[6]</strong></a>.</p><p>Imagine a robot whose skeletal frame <em>is</em> its distributed energy storage system. Not a battery pack bolted to a chassis, but energy storage woven into the very structure that provides mechanical support. Every strut, every beam, every joint housing becomes a local energy reserve.</p><p>This concept has already been validated in robotics. A landmark 2020 paper in <em>Science Robotics</em> demonstrated “biomorphic structural batteries for robotics,” showing that structural zinc batteries integrated into the robot body could provide up to 72 times more energy than standalone cells by eliminating dedicated battery enclosures<a href="#ref-7"><strong>[7]</strong></a>.</p><p>This is the mitocapacitor vision: thousands of micro and supercapacitors distributed throughout the structural frame of a robot, embedded in the architecture itself, as mitochondria are embedded in muscle fibers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BAguffYwNUJnJtHl.jpeg" /></figure><p><em>AI-generated illustration of ultrafast energy storage and power delivery via electrostatic microcapacitors directly integrated on-chip for </em><a href="https://engineering.berkeley.edu/news/2024/05/researchers-achieve-giant-energy-storage-power-density-on-a-microchip/"><em>next-generation microelectronics</em></a><em>.</em></p><p>There’s a further advantage biology discovered: d<strong>istributed energy conversion distributes heat</strong>. A centralized battery pack concentrates thermal load, demanding liquid cooling systems that add mass and failure modes. Thermal runaway, where one overheating cell ignites its neighbors, remains lithium-ion’s existential risk. A mitocapacitor architecture eliminates this by design. Thousands of small energy nodes, each generating minimal heat, turn the robot’s structural frame into a passive radiator. The same principle keeps your body at 37 degrees: trillions of mitochondria each producing tiny amounts of heat, collectively maintaining homeostasis without any centralized cooling system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/756/0*JK2AL4iwV5yv54u0.jpeg" /></figure><p>a 3D reconstruction of the brain’s dense vascular network showing how distributed blood flow handles thermal regulation</p><p><strong>Consider the brain: </strong><a href="https://hypertextbook.com/facts/2001/JacquelineLing.shtml"><strong>20 watts of continuous power</strong></a><strong> dissipated within 1.4 kilograms of tissue that is exquisitely sensitive to temperature fluctuation</strong>s. Neurons malfunction with even small thermal deviations. Yet the brain never overheats because biology distributes both energy conversion and cooling. Dense vascular networks carry heat away from every cubic millimeter of neural tissue simultaneously. There is no centralized radiator, no heat pipe, and no fan. The architecture <em>is</em> the cooling system. A mitocapacitor robot inherits this principle: by distributing energy conversion across its entire frame, it eliminates thermal hot spots and the heavy cooling infrastructure they demand.</p><h3>Micro-Supercapacitors: The Building Blocks</h3><p>A 2020 review in <em>npj Flexible Electronics</em> on “Recent developments of advanced micro-supercapacitors” documented extraordinary progress in fabricating micro-scale energy storage devices<a href="#ref-8"><strong>[8]</strong></a>. These devices can be printed onto flexible substrates, embedded in composites, and integrated into remarkably compact systems.</p><p>Meanwhile, 2D MXene-based supercapacitor electrodes (two-dimensional transition metal carbides) are pushing energy densities higher while maintaining the instantaneous charge/discharge characteristics that make distributed architectures viable<a href="#ref-9"><strong>[9]</strong></a>.</p><h3>The ATP Pool Analog</h3><p>The microcapacitor array isn’t a battery in the traditional sense. It’s an always-ready local energy reserve, continuously topped off by a slower baseline source. This mirrors the ATP pool in biological cells:</p><ul><li>Mitochondria maintain a local ATP concentration</li><li>ATP is consumed within milliseconds by nearby enzymes and motors</li><li>Mitochondria continuously regenerate ATP to maintain the pool</li><li>The pool never depletes because generation matches consumption</li></ul><p>The key insight is that peak demand and average demand are radically different numbers. A robot joint might require 500W for 100 milliseconds during a rapid movement, but only 5W averaged over a minute. Batteries are sized for peak demand. Mitocapacitor arrays are sized for average demand, with distributed capacitors handling the instantaneous peaks.</p><h3>Section 4: The Bioreactor Bridge</h3><p>Every architecture needs a complete energy story. For biology, that story begins with food and ends with ATP. For the mitocapacitor architecture, we need to identify what plays the role of the digestive system, the slow, continuous energy source that keeps the distributed array charged.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yC3o6olh18X_uZy7.png" /></figure><p><a href="https://www.mdpi.com/2071-1050/18/12/5898">Schematic diagram represents a dual chamber Microbial Fuel Cell (MFC)</a> showing bacterial biofilm formed on the anode and electricity generation from wastewater through microbial oxidation of organic matter.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZAyyS_vyYivDSLyx.png" /></figure><p><a href="https://onlinelibrary.wiley.com/doi/10.1155/2021/4570388">Systematic presentation of microbial fuel cell (MFC</a>) A bioelectrochemical approach based on metabolic activities inoculated bacteria, and the MFC acts as biocatalysts for the natural conversion to energy of organic substrates</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/389/0*exzh7FN7NfUDxwOh.jpeg" /></figure><p><a href="https://www.sciencedirect.com/science/article/pii/S2667141721000525">Schematic drawing of an </a><a href="https://www.sciencedirect.com/topics/engineering/enzymatic-biofuel-cell">Enzymatic biofuel cells</a> (EBFCs), consisting of a GOx-based bioanode (PDB: <a href="http://www.rcsb.org/pdb/explore.do?structureId=3QVP">3QVP</a> — <strong>crystal structure of glucose oxidase</strong>) and a BOx-based biocathode (PDB: <a href="http://www.rcsb.org/pdb/explore.do?structureId=2XLL">2XLL</a> — <strong>crystal structure of bilirubin oxidase from Myrothecium verrucaria</strong>).</p><h3>The Charging Options</h3><p><strong>Inductive charging plates:</strong> The simplest approach treats electricity from the grid as “food,” delivered through wireless induction pads embedded in floors at workstations, assembly lines, or anywhere a robot stands for sustained periods. Figure AI’s Figure 03 already demonstrates this: charging coils in the robot’s feet allow it to step onto a wireless pad and charge at 2 kW while continuing to work. No plug, no docking station, no interruption. Anywhere a humanoid stands for more than a few minutes becomes a charging surface. This is the metabolic equivalent of grazing: continuous low-level energy intake during normal activity, not a discrete “refueling” event.</p><p><strong>Microbial fuel cells (MFCs):</strong> These bioelectrochemical systems use bacterial metabolism to generate electricity from organic matter. A 2025 comprehensive review on “Microbial fuel cells powering robots and beyond” documents the complete historical arc from early experiments to fully autonomous robots powered by waste<a href="#ref-10"><strong>[10]</strong></a>. The theoretical framework of Gastrobotics (Wilkinson, 2000) envisioned robots that locate, harvest, and digest food to power themselves<a href="#ref-11"><strong>[11]</strong></a>.</p><p><strong>The EcoBot lineage:</strong> Perhaps the most compelling proof-of-concept comes from the Bristol Robotics Laboratory, where EcoBot-II demonstrated an artificial agent with a “natural metabolism,” using MFCs fed with flies and waste to power fully autonomous behavior<a href="#ref-12"><strong>[12]</strong></a>. EcoBot-III extended this to complete energy autonomy from wastewater. These weren’t toy demonstrations; they were engineering proofs that biological energy pathways can sustain robotic systems.</p><p><strong>Enzymatic biofuel cells:</strong> These systems use isolated enzymes rather than whole microbes to catalyze electrochemical reactions. A 2024 review on glucose-based biofuel cells documented how enzymatic systems operating on glucose can achieve higher power densities than MFCs<a href="#ref-13"><strong>[13]</strong></a>. A 2019 study in <em>Scientific Reports</em> demonstrated practical glucose-reactive enzymatic biofuel cell performance for biomedical applications<a href="#ref-14"><strong>[14]</strong></a>.</p><p><strong>Hybrid bio-electrochemical systems:</strong> The 2021 review “Towards Bio-Hybrid Energy Harvesting in the Real-World” pushes the boundaries further, examining how bio-electrochemical and bio-mechanical processes can be combined for practical energy generation<a href="#ref-15"><strong>[15]</strong></a>. Stanford’s microbial battery technology demonstrates how an anode coated with exoelectrogenic microbes can ingest organic waste to produce sustained electricity<a href="#ref-16"><strong>[16]</strong></a>.</p><h3>The Biofuel Provocation</h3><p>What if the charging station of the future isn’t a wall outlet but a bioreactor? A robot that can convert waste biomass, wastewater, or even ambient organic matter into electricity has fundamentally different operational constraints than one tethered to the grid. The EcoBot lineage proved this is physically possible. The question is whether we can scale the power densities to match the demands of dynamic locomotion. Current MFC power densities remain in the milliwatts-to-watts range per unit, but enzymatic cells are pushing higher, and the combinatorial approach (MFC + supercapacitor buffer) could bridge the gap between biological energy rates and mechanical power demands.</p><h3>Running the Numbers: Could a Robot Have a Gut?</h3><p>The provocation demands quantification. Let’s work through the engineering arithmetic to determine whether a bio-digestive architecture could realistically power a humanoid robot.</p><p><strong>The power target:</strong> A lightweight (57 kg) humanoid optimized for efficiency needs approximately 200–300W continuous for mixed factory tasks. With distributed supercapacitors handling peak demands, the bioreactor only needs to provide the average trickle charge, not the instantaneous peaks. Target: ~130–200W continuous from the bioreactor.</p><p><strong>MFC power densities (current state of art):</strong></p><ul><li>Early MFCs (2000s): 0.001 W/m2 electrode area</li><li>Modern MFCs (2024): 1–7 W/m2 (best: MXene-enhanced anodes reaching ~7 W/m2)</li><li>At 3 W/m2 (practical average): need 44–67 m2 of electrode area for 130–200W</li><li>Verdict: MFCs alone are volumetrically impractical for a humanoid form factor</li></ul><p><strong>Enzymatic biofuel cells (glucose-fed):</strong></p><ul><li>Demonstrated power density: 1–10 mW/cm2 (= 10–100 W/m2)</li><li>At 5 mW/cm2 (achievable mid-range): need 2.6–4.0 m2 for 130–200W</li><li>A coiled tube, 3 cm diameter, 14 meters long = 1.32 m2 bare surface</li><li>With engineered micro-villi (30x surface enhancement, mimicking intestinal architecture): effective area = 39.6 m2</li><li>At 5 mW/cm2 on villi-enhanced surface: theoretical output = ~198W</li></ul><p><strong>The gut-tube architecture:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xBdf8uC7VHRlrIjn.png" /></figure><p><strong>Fuel requirements:</strong> At 200W continuous output with ~40% glucose-to-electricity conversion efficiency, the system would consume approximately 430 grams of glucose per 16-hour operating day, dissolved in roughly 2–4 liters of solution. That’s less than a pound of sugar per shift, a remarkably achievable logistics burden compared to the infrastructure required for battery charging.</p><p><strong>The hybrid architecture that emerges:</strong></p><ol><li>Distributed supercapacitor array (structural, 500–800 Wh total): handles instantaneous peaks</li><li>Enzymatic gut-tube bioreactor (torso-mounted, 2–4L): provides 100–200W continuous trickle charge</li><li>Robot “eats” glucose solution periodically (like refueling a car, but with sugar water)</li><li>The bioreactor never needs to match peak demand because the capacitors buffer everything</li><li>During “sleep” (low-activity charging periods), the bioreactor tops off the full capacitor array</li></ol><p><strong>Critical caveat:</strong> These are theoretical upper bounds. Real-world enzyme degradation, mass transport limitations, biofouling, and thermal management will reduce practical output by 30–60%. A realistic near-term system might achieve 50–80W, insufficient alone but potentially viable as a range extender that doubles operational time between conventional charges. The path to full energy autonomy requires continued advances in enzyme stability (currently 30–90 day lifetimes before replacement) and micro-villi fabrication at scale.</p><p>The conclusion is provocative but grounded: a robotic digestive system is not thermodynamically impossible. The surface areas, power densities, and volumetric constraints are all within an order of magnitude of feasibility. The engineering challenges are fabrication, durability, and integration, not physics.</p><h3>The Complete Biological Mapping</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kl-_jKWCwSzDgW9e.png" /></figure><h3>The Central Insight</h3><p>The body never stores energy proportional to its mass in a centralized reservoir. Your fat stores are distributed. Your glycogen reserves are distributed (primarily in liver and muscle). Your ATP pools are distributed across trillions of mitochondria. Evolution ruthlessly eliminated centralized energy storage in mobile organisms because the mass penalty was too severe.</p><p>We’ve been fighting evolution’s verdict with brute-force battery engineering. It’s time to stop.</p><p><em>The next generation of humanoid robots will need power architectures as sophisticated as the biological systems they aim to emulate.</em></p><h3>Section 5: What Research Exists Today</h3><h3>Already Deployed</h3><p><strong>Supercapacitors in transportation:</strong> Hybrid buses in China use supercapacitor banks for regenerative braking, charging in seconds at stops and discharging during acceleration. Some urban tram systems operate entirely on supercapacitors, recharging at each station. The technology is proven at scale for high-power, frequent-cycle applications.</p><p><strong>Grid-scale supercapacitor systems:</strong> Power quality applications use supercapacitor arrays to smooth frequency fluctuations and provide instantaneous backup power.</p><h3>Active Research</h3><p><strong>Bioinspired Distributed Energy in Robotics:</strong> A 2023 paper in <em>Advanced Intelligent Systems</em> directly addressed “Bioinspired Distributed Energy in Robotics and Enabling Technologies,” covering flexible piezoelectric and triboelectric nanogenerators, 3D-printed energy storage, and distributed power architectures for mobile robots<a href="#ref-17"><strong>[17]</strong></a>.</p><p><strong>Structural supercapacitors:</strong> Research groups at Cambridge, MIT, and Chalmers are developing carbon fiber composites with energy storage capability. Current energy densities remain modest (typically single-digit Wh/kg for the structural component), but the concept is proven and improving rapidly.</p><p><strong>Biomorphic batteries:</strong> The University of Michigan’s biomorphic battery work demonstrates that when you stop treating the battery as a separate component and instead integrate it into the structural body plan, the energy available to a robot can increase by orders of magnitude<a href="#ref-7"><strong>[7]</strong></a>.</p><p><strong>MXene-based energy materials:</strong> Doped MXene supercapacitors are being explored as potential “game-changers” for next-generation energy landscapes, with performance characteristics that bridge the gap between batteries and traditional capacitors<a href="#ref-18"><strong>[18]</strong></a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ECin2dE-hZyfM4LH.jpeg" /></figure><p><a href="https://www.powerelectronictips.com/how-is-power-limiting-the-adoption-of-physical-artificial-intelligence-in-humanoid-robotics/"><strong>The energy storage, power conversion, and motion systems can amount to 30% of the cost of a humanoid robot</strong></a><strong>. (Image: </strong><a href="https://www.macquarie.com/lu/en/insights/are-we-turning-a-corner-on-the-humanoid-robot-age.html"><strong>Macquaire</strong></a><strong>) </strong><em>The next frontier: humanoid robots that integrate distributed energy storage into their structural frame, eliminating the centralized battery paradigm entirely.</em></p><h3>The Integration Gap</h3><p>Here’s what’s missing: nobody has assembled these pieces into a complete mitocapacitor architecture for a mobile robot.</p><p>We have structural supercapacitors. We have micro-supercapacitors. We have AI-driven energy management. We have bio-electrochemical power generation. But these remain separate research streams. No team has yet built a robot with:</p><ul><li>A carbon fiber composite frame serving as distributed energy storage</li><li>Thousands of embedded micro-supercapacitors positioned at points of actuation</li><li>A slow-rate charging system (whether bioreactor-based or conventional)</li><li>An AI-driven system that predicts motion sequences and pre-charges relevant capacitor clusters</li><li>Fault-tolerant architecture allowing graceful degradation when components fail</li></ul><p>This integration challenge is the opportunity.</p><h3>The Path Forward</h3><p>In the first two volumes of <em>The Rise of the Machine</em>, I argued that truly capable AI-physical systems would require not just better algorithms but fundamentally rethought embodiment. Better sensors. Better actuators. Better integration between perception and action.</p><p>Now I’m arguing for something equally fundamental: <strong>better power architecture</strong>.</p><p>The machines that will move and think like biology will also need to <em>power</em> like biology.</p><p>This isn’t just about efficiency, though the efficiency gains would be substantial. It’s about creating robots that can operate in the real world with the grace, adaptability, and endurance of biological systems. A humanoid robot with mitocapacitor architecture could:</p><ul><li>Sustain hours of operation with a fraction of the battery mass</li><li>Handle peak power demands through instantaneous capacitor discharge</li><li>Tolerate component failures through distributed redundancy</li><li>Scale power delivery to match local demand without centralized bottlenecks</li><li>Potentially harvest energy from waste streams, extending operational range indefinitely</li></ul><p>The next generation of AI-physical systems, robots that work alongside humans, that move gracefully, that run for hours without returning to charging stations, will be powered by distributed intelligence as much as distributed energy. The control system that predicts which joints will fire in the next 100 milliseconds and pre-charges those specific capacitors is as important as the capacitors themselves.</p><h3>A Challenge</h3><p><strong>To the materials scientists:</strong> prove that structural supercapacitor composites can achieve energy densities sufficient for practical robotics applications.</p><p><strong>To the power electronics engineers:</strong> design the DC-DC conversion and distribution systems that can manage thousands of independent energy storage nodes.</p><p><strong>To the AI researchers:</strong> build the predictive systems that can anticipate power demand milliseconds to seconds in advance.</p><p><strong>To the roboticists:</strong> stop accepting battery packs as a given and start designing for distributed energy from first principles.</p><p><strong>To the bio-electrochemists:</strong> push microbial and enzymatic fuel cells toward the power densities that mobile systems demand.</p><p>The mitocapacitor is not just a metaphor. It is an engineering specification waiting to be built. Evolution spent two billion years optimizing distributed energy systems for mobile organisms. The designs are proven. The principles are established. The component technologies are emerging.</p><p>All that’s missing is the will to assemble them, and the recognition that the battery, as we’ve known it, was always the wrong answer.</p><p><strong>The machines are rising. Let’s make sure they can actually keep moving.</strong></p><h3>References</h3><p><strong>[1]</strong> “Evaluation of potential electric vehicles load-induced damage on flexible pavements.” <em>Transportation Research Part D</em>, 2024. <a href="https://www.sciencedirect.com/science/article/abs/pii/S1361920924004322">sciencedirect.com</a></p><p><strong>[2]</strong> “Hidden effects and externalities of electric vehicles.” <em>Energy Policy</em>, 2024. <a href="https://www.sciencedirect.com/science/article/pii/S0301421524003550">sciencedirect.com</a></p><p><strong>[3]</strong> Giacomello, M. et al. “Molecular mechanisms of mitochondrial dynamics.” <em>Nature Reviews Molecular Cell Biology</em>, 2024. <a href="https://www.nature.com/articles/s41580-024-00785-1">nature.com</a></p><p><strong>[4]</strong> “Mechanism of ATP hydrolysis dependent rotation of bacterial ATP synthase.” <em>Nature Communications</em>, 14, 2023. <a href="https://www.nature.com/articles/s41467-023-39742-5">nature.com</a></p><p><strong>[5]</strong> Wang, K. et al. “Carbon fiber electrodes for composite structural supercapacitor: Preparation and modification methods.” <em>Journal of Energy Storage</em>, 2024. <a href="https://www.sciencedirect.com/science/article/abs/pii/S2352152X24027154">sciencedirect.com</a></p><p><strong>[6]</strong> “Carbon fiber structural battery paves way for light, energy-efficient vehicles.” <em>Chalmers University of Technology</em>, 2024. <a href="https://techxplore.com/news/2024-09-carbon-fiber-battery-paves-energy.html">techxplore.com</a></p><p><strong>[7]</strong> Wang, M. et al. “Biomorphic structural batteries for robotics.” <em>Science Robotics</em>, 5(45), eaba1912, 2020. <a href="https://www.science.org/doi/10.1126/scirobotics.aba1912">science.org</a></p><p><strong>[8]</strong> “Recent developments of advanced micro-supercapacitors: design, fabrication and applications.” <em>npj Flexible Electronics</em>, 4, 2020. <a href="https://www.nature.com/articles/s41528-020-00093-6">nature.com</a></p><p><strong>[9]</strong> “Advancements in 2D MXene-based supercapacitor electrodes.” <em>Journal of Materials Chemistry A</em>, RSC, 2024. <a href="https://pubs.rsc.org/en/content/articlelanding/2024/ta/d4ta00328d">rsc.org</a></p><p><strong>[10]</strong> “Microbial fuel cells powering robots and beyond; historical perspective and future directions.” <em>Biosensors and Bioelectronics</em>, 2025. <a href="https://www.sciencedirect.com/science/article/pii/S1567539425001136">sciencedirect.com</a></p><p><strong>[11]</strong> Wilkinson, S. “‘Gastrobots’: Benefits and Challenges of Microbial Fuel Cells in Food-Powered Robot Applications.” <em>Autonomous Robots</em>, 9, 99–111, 2000. <a href="https://link.springer.com/article/10.1023/A%3A1008984516499">springer.com</a></p><p><strong>[12]</strong> Ieropoulos, I. et al. “EcoBot-II: An Artificial Agent with a Natural Metabolism.” <em>International Journal of Advanced Robotic Systems</em>, 2005. <a href="https://journals.sagepub.com/doi/10.5772/5777">sagepub.com</a></p><p><strong>[13]</strong> Chakraborty, I. et al. “Glucose-based biofuel cells and their applications in medical implants: A review.” <em>Heliyon</em>, 2024. <a href="https://www.sciencedirect.com/science/article/pii/S2405844024096464">sciencedirect.com</a></p><p><strong>[14]</strong> “Performance of a glucose-reactive enzyme-based biofuel cell system for biomedical applications.” <em>Scientific Reports</em>, 9, 2019. <a href="https://www.nature.com/articles/s41598-019-47392-1">nature.com</a></p><p><strong>[15]</strong> “Towards Bio-Hybrid Energy Harvesting in the Real-World.” <em>Applied Sciences</em>, 11(5), 2220, 2021. <a href="https://www.mdpi.com/2076-3417/11/5/2220">mdpi.com</a></p><p><strong>[16]</strong> “Microbial Batteries Produce Energy from Wastewater and Other Organic Matter.” <em>Stanford Technology Finder</em>. <a href="https://techfinder.stanford.edu/technology/microbial-batteries-produce-energy">techfinder.stanford.edu</a></p><p><strong>[17]</strong> Mukherjee, S. et al. “Bioinspired Distributed Energy in Robotics and Enabling Technologies.” <em>Advanced Intelligent Systems</em>, 2023. <a href="https://advanced.onlinelibrary.wiley.com/doi/10.1002/aisy.202100036">wiley.com</a></p><p><strong>[18]</strong> “Can Doped-MXene-Based Supercapacitors Be the Game-Changer for Future Energy Landscape?” <em>Langmuir</em>, ACS, 2024. <a href="https://pubs.acs.org/doi/abs/10.1021/acs.langmuir.4c01242">acs.org</a></p><p><strong>[19]</strong> Asp, L.E. et al. “A Structural Battery and its Multifunctional Performance.” <em>Advanced Energy and Sustainability Research</em>, 2(3), 2100093, 2021. <a href="https://onlinelibrary.wiley.com/doi/full/10.1002/aesr.202100093">wiley.com</a></p><p><strong>[20]</strong> Gonzalez, C. et al. “Structural composites for multifunctional applications.” <em>Progress in Materials Science</em>, 89, 194–251, 2017. <a href="https://www.sciencedirect.com/science/article/pii/S0079642517300476">sciencedirect.com</a></p><p><strong>[21]</strong> Simon, P. &amp; Gogotsi, Y. “Perspectives for electrochemical capacitors and related devices.” <em>Nature Materials</em>, 19, 1151–1163, 2020. <a href="https://www.nature.com/articles/s41563-020-0747-z">nature.com</a></p><p><strong>[22]</strong> Greenman, J. et al. “Microbial fuel cells and their electrified biofilms.” <em>Biofilm</em>, 3, 100057, 2021. <a href="https://www.sciencedirect.com/science/article/pii/S2590207521000186">sciencedirect.com</a></p><p><strong>[23]</strong> Pomerantseva, E. et al. “Energy storage: The future enabled by nanomaterials.” <em>Science</em>, 366(6468), eaan8285, 2019. <a href="https://www.science.org/doi/10.1126/science.aan8285">science.org</a></p><p><strong>[24]</strong> Zhu, M. et al. “A Highly Durable All-Polymer Micro-Supercapacitor with Plug-and-Play Function.” <em>Advanced Materials</em>, 33(18), 2006091, 2021. <a href="https://onlinelibrary.wiley.com/doi/abs/10.1002/adma.202006091">wiley.com</a></p><p><strong>[25]</strong> Picard, M. &amp; Shirihai, O.S. “Mitochondrial signal transduction.” <em>Cell Metabolism</em>, 36(6), 1248–1253, 2024. <a href="https://www.cell.com/cell-metabolism/fulltext/S1550-4131(24)00127-3">cell.com</a></p><p><em>Steven Muskal writes at the intersection of molecular biology, materials science, and AI systems. This is apart of “The Rise of the Machine” series. Previous installments explored embodied intelligence and sensorimotor integration.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*247rPVydz0snlgNS.png" /></figure><p>Steven Muskal, Ph.D. is the CEO of <a href="https://www.eidogen-sertanty.com/">Eidogen-Sertanty, Inc.</a> — a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at <a href="http://stevenmuskal.com/">stevenmuskal.com</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=139107ca1be7" width="1" height="1" alt=""><hr><p><a href="https://aiadvances.org/energizing-the-machine-139107ca1be7">Energizing the Machine</a> was originally published in <a href="https://aiadvances.org/">AI Advances</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Don’t Buy into the AI IPOs]]></title>
            <link>https://aiadvances.org/dont-buy-into-the-ai-ipos-1f58938e8d10?source=rss-7c2cfeb3b1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/1f58938e8d10</guid>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[investing]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Steven Muskal]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 00:48:11 GMT</pubDate>
            <atom:updated>2026-11-04T01:08:52.163Z</atom:updated>
            <content:encoded><![CDATA[<h4>History Doesn’t Repeat, But Compute Rhymes</h4><p><em>Anthropic and OpenAI are selling you the mainframe at the top of the swing</em></p><p>Anthropic filed to go public last week. Analysts estimate it could list at <strong>$1.5 trillion to $2 trillion</strong>, potentially one of the largest IPOs in history. OpenAI has filed confidentially and is reportedly raising private money at a valuation between $1.25 trillion and $1.4 trillion. Before you buy, consider the question some have heard me sound off about: could AI disappear? Every so often a technology shows up a little ahead of its time and gets buried. Sometimes it stays buried for a decade or more.</p><p>Don’t get me wrong, I believe AI itself is here to stay. The centralized, hyperscale version of it, the one being financed with trillions of dollars of data center commitments, could fade, at least for a while. It may well come back later, because that is how these cycles run. In between, I expect a large correction. I have been working in AI for over 40 years, and I have seen this movie before. Several times.</p><h3>The car they crushed</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oHCXPyMm9FWdRlnNoaMijQ.jpeg" /><figcaption>One of the few EV1s that escaped the crusher, now at the National Motor Museum, Birdwood, South Australia</figcaption></figure><p><strong>In 1996 </strong>General Motors put the EV1 on the road. It was the first modern electric car from a major American automaker, and the people who drove it loved it. You could not buy one. GM only leased them, in California and Arizona, and built roughly 1,100.</p><p>When the leases ran out, drivers asked to buy their cars. GM said no, took them back, and sent most of them to the crusher. A handful went to museums and universities with their drivetrains disabled.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*aXdpVtdkUGqqWcKPsg_YiA.jpeg" /><figcaption>Crushed EV1s. Photo: Plug In America</figcaption></figure><p>Drivers held a mock funeral for the car at Hollywood Forever Cemetery in 2003, bagpiper and all. The 2006 documentary <em>Who Killed the Electric Car?</em> laid out the suspects: oil companies, regulators, and GM itself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1iQfD783kOJ3PEnwukXecg.jpeg" /><figcaption>The EV1 “funeral,” July 2003</figcaption></figure><p>Whatever you believe about the conspiracy, the lesson I take from it is simple. The technology worked. The incumbent that built it had no reason to want it to succeed, and once California softened its zero emission mandate, the business case went away. Lease only meant GM kept ownership, and the crusher closed out the liability.</p><p>The idea did not die. Tesla was founded in 2003, shipped the Roadster in 2008, and within a decade every major automaker was back in the electric car business. It came back from outside the incumbent, about fifteen years later.</p><p>And the pendulum keeps swinging. Since late 2025, Stellantis, Ford and GM have written off more than $50 billion on electric vehicle plans, Ford killed the electric F-150 Lightning in favor of a hybrid, and Honda shelved three U.S. electric models, all after building for demand that arrived slower than they forecast.</p><h3>Big Blue, big iron</h3><p>Computing has been swinging back and forth between centralized and distributed for seventy years. I have watched most of it firsthand.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*wp7OhTSsKj63osBH0hbcWw.jpeg" /><figcaption>Twin IBM 7090s in NASA’s computer room, around 1962</figcaption></figure><p>First came the mainframe. IBM, Big Blue, filled whole rooms with big iron, and the System/360 family announced in 1964 became the standard for business computing. All the intelligence lived in the machine room. Terminals were dumb. You submitted a job and waited your turn.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hrcuLDG-EvFT-hY8owDy6g.jpeg" /><figcaption>IBM System/360 Model 40, circuit gates open.</figcaption></figure><h3>Cray, the Connection Machine, and Silicon Graphics</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-ypdDkogLno78d_Dm6OSPQ.jpeg" /><figcaption>The Cray-1, 1976, with its famous padded bench hiding the power supplies</figcaption></figure><p>Mainframes ran the business. Supercomputers ran the science. In 1976 Seymour Cray delivered the Cray-1 to Los Alamos for roughly $8 million. It peaked at about 160 million floating point operations per second, and it was the fastest machine in the world.</p><p>Then came parallelism. The Cray X-MP in 1982 put more than one processor on the same problem. The Cray-2 in 1985 packed so much into so little space that its circuit boards sat submerged in liquid coolant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yB0R7B-BB495mpXwQd9oGw.png" /><figcaption>Left: the Cray X-MP, Cray’s first multiprocessor. Right: the Cray-2, 1985, cooled by immersion in Fluorinert.</figcaption></figure><p>Thinking Machines took a different road. Its Connection Machine, also from 1985, used 65,536 very simple processors working at once. That idea, many small processors attacking one problem together, is what pulled me into parallel computing and neural networks early on.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FI33a1IQKLfqpWXgsvvB5w.jpeg" /><figcaption>Thinking Machines Connection Machine CM-1, 1985, at the Computer History Museum.</figcaption></figure><p>At the same time, Silicon Graphics was building workstations with dedicated graphics hardware, machines like the Indigo and the Onyx. They rendered 3D in real time, which was unheard of on a desktop, and they powered the movie effects of the early 1990s. That specialized graphics silicon was the ancestor of the GPU. In 1999 Nvidia shipped the GeForce 256 and called it the first GPU. Twenty five years later, chips descended from it train and run every large AI model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*95Ol5yV9kjX4tQIuofwsQA.png" /><figcaption>Silicon Graphics Indigo (1991) and Onyx (1993)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xutl4OtM5hG5wgvKjLf8Yg.jpeg" /><figcaption>Nvidia GeForce 256, 1999, marketed as the first GPU</figcaption></figure><p>None of those companies survived as they were. Silicon Graphics went bankrupt and its name was sold. Cray changed hands several times and is now part of Hewlett Packard Enterprise. The power those machines held in one room ended up on everyone’s desk, and then in everyone’s pocket. The phone you are holding is thousands of times faster than a Cray-1.</p><h3>When it landed on the desk</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/962/1*vdZCwxNNwRAMHM1QZJMa_Q.jpeg" /><figcaption>The IBM PC 5150, introduced August 1981</figcaption></figure><p>In 1981 IBM shipped the PC, and the compute came to your desk. No more begging the data processing department for time. Your software ran on your machine and your data stayed on your disk. The pendulum had swung hard toward the individual.</p><h3>Client server</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*kJNVdv6N8qIEXpC5nxCUeA.jpeg" /><figcaption>A Compaq ProLiant, the kind of box that sat in thousands of back rooms in the 1990s</figcaption></figure><p>Through the late 1980s and 1990s we met in the middle. The desktops stayed smart, but servers came back for the shared pieces: databases, email, file storage. Every company of any size had a server closet. Small businesses ran their own machines and owned their own data.</p><h3>Then Bezos wanted to be the server room for everyone</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MeQm-kikNFw5KPJdQimTgQ.jpeg" /><figcaption>BusinessWeek, November 13, 2006: “Amazon’s Risky Bet.”</figcaption></figure><p>Amazon Web Services launched S3 and EC2 in 2006. That November, BusinessWeek put Jeff Bezos on the cover under the headline “Amazon’s Risky Bet.” Wall Street hated it. Why would a bookseller want to rent out its computers?</p><p>Bezos keeps that cover framed. In 2022 he posted a photo of it and pointed out that the risky bet had brought in more than $62 billion the year before. He was right. Startups stopped buying servers. Enterprises closed their server closets. The pendulum swung all the way back to centralization, this time with an API and a credit card.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZgZQHxUfRRr-LrBmPyvVww.png" /><figcaption>Left: AWS buildings in Ashburn, Virginia. Right: Google’s Council Bluffs, Iowa campus.</figcaption></figure><p>AI rode that wave. Training a frontier model takes tens of thousands of GPUs, and only the hyperscalers could pay for them. Serving the models to everyone also ran best in their data centers, and they were happy to meter it. Every question you ask flows back to them.</p><p>I think we are near the top of that swing.</p><h3>Why I think the build out is overdone</h3><p>McKinsey puts global data center spending at nearly $7 trillion by 2030. Bain estimates the industry needs $2 trillion in new annual revenue by then just to pay for the scaling it has planned. A lot of that is being financed with debt, long term power contracts, and leases.</p><p>Here is what bothers me. A trained model is a large collection of neural network weights, numbers that encode what the network learned. Even the most powerful models today fit on a thumb drive. That has two consequences.</p><p>First, they can be distilled. A smaller model can be trained to imitate a larger one at a fraction of the cost.</p><p>Second, they can walk out the door. One employee with access and a USB stick can leave the building carrying an entire IP position in their back pocket. That is no longer hypothetical. In January 2026 a federal jury convicted a former Google engineer of economic espionage for taking more than 2,000 pages of Google’s AI trade secrets while lining up work with Chinese companies.</p><h3>Open weights changed the math</h3><p>The bigger story is that much of the best AI is now free to download. Open weight models publish the numbers themselves, so anyone can run them on their own hardware, inspect them, and fine tune them on their own data.</p><p>China has been the most aggressive here. DeepSeek released its R1 reasoning model in January 2025 under an MIT license. It came close to the best American closed models, reportedly at a small fraction of the training cost, and on the Monday after it hit the top of the app charts Nvidia lost roughly $589 billion in market value in a single day. Alibaba’s Qwen family followed with models from tiny to enormous under the Apache 2.0 license, and Qwen has become one of the most popular starting points for derivative models on Hugging Face. Moonshot’s Kimi K2, Zhipu’s GLM, and MiniMax have all released large open weight models of their own.</p><p>The West answered. Meta’s Llama, Mistral in France, and Google’s Gemma are all open, and in August 2025 even OpenAI released open weight models, its first since GPT-2.</p><p>The result is that capability which cost billions to create is now a download. These open models write code, read documents, reason through problems, and hold a conversation better than most of what we will ever need. Using a trained network, running inference through it, does not require retraining it. It only requires a place where the weights fit.</p><p>So I do not see a need to keep pushing ever bigger models, and I do not see a need for the data center build out at the scale being promised. If a developer can download a model, run it locally, and get most of the way to the frontier without sending a byte to a hyperscaler, why pay rent?</p><h3>Who is listening to your questions?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/336/0*4IBmIa3UM0sWM8Yd.png" /></figure><p>In the late 1990s IBM put up a free patent server on the web. You could search U.S. patents and see what was out there, and for its day it was a great tool. I remember companies telling their people not to use it. The worry was simple: every query told IBM what you were working on. Type in a few searches around a new idea and you have just described your next product to one of the largest patent holders in the world. Whether IBM ever looked at those logs hardly mattered. Nobody wanted to find out. IBM spun the service out as Delphion in 2000.</p><p>Today’s AI chat window is that patent server times a thousand. People do more than search. They paste in the draft patent, the source code, the board deck, the customer list, the clinical data, and ask the model to improve it. In drug discovery, a medicinal chemist asking a cloud model about a new scaffold has just revealed the program. Each prompt is a disclosure to a third party, and you have little real control over where it is stored, who can see it, how long it is kept, or whether it ends up in a future training run.</p><p>This has already happened. In 2023 Samsung engineers pasted confidential source code into ChatGPT, and Samsung banned generative AI tools on company devices. Apple, JPMorgan, Amazon, Verizon and others restricted employee use for the same reason. In 2025 a federal judge in the New York Times copyright case ordered OpenAI to preserve users’ ChatGPT conversations, including ones they had deleted. Your “deleted” questions became potential evidence in someone else’s lawsuit.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8JkwwQgknu77k0WB1YNY1Q.png" /><figcaption>Left: the Navier-Stokes equations. Right: the Fields Medal.</figcaption></figure><p>Mathematicians just lived through a version of this. <strong>In September 2026 OpenAI announced a solution to the Navier-Stokes problem</strong>, one of the million dollar Millennium Prize problems. Hours earlier, NYU mathematician Tristan Buckmaster had said that work he and Levent Alpöge had been doing, and had run through OpenAI’s own Codex to clean up and rewrite, had leaked into OpenAI’s effort. OpenAI denies it, and says none of his inputs after July 3 could have influenced the result. A month later OpenAI posted 722 math manuscripts in a single drop, and then claimed a second Millennium problem, Yang-Mills, ahead of two mathematicians who had been posting their progress in public. One Fields Medalist said OpenAI had “annihilated my field of research.”</p><p>I cannot tell you whose ideas went where, and that is exactly the point. When the company refining your work is also racing to publish in your field, you have no way to know. Every question, draft, and proof you hand a frontier lab tells them what you are working on and how far along you are. The IBM patent server only saw what you searched for. These systems see the work itself.</p><p>Then there is the cyber threat. A hyperscaler holding the prompts of millions of companies is the richest espionage target ever assembled. <strong>Researchers found a DeepSeek database sitting open on the internet in January 2025 with chat histories and secret keys inside. </strong>In November 2025 Anthropic reported that a Chinese state sponsored group had used its own Claude model to automate much of an espionage campaign against roughly thirty organizations. The same tools that answer your questions are being pointed at your network.</p><p>Serious businesses will respond the way they always have with their crown jewels. They will lock them down. For the most sensitive work that means nearly air gapped systems: models running on machines the company owns, inside its own walls, with no path out to the internet. Defense, pharma, finance, and law are already moving this way. Trade secret law protects only information a company takes reasonable steps to keep secret, and pasting it into someone else’s chatbot is a hard step to defend.</p><p><strong>I believe this, more than anything else, will undo the hyperscale investment thesis</strong>. Corporations protecting their intellectual property and individuals protecting their privacy want the same thing: AI that stays home. When the biggest customers decide their most valuable questions can never leave the building, the revenue those trillion dollar data centers are counting on goes with them.</p><p>The capital commitments assume the pendulum stays where it is. I doubt the companies talking about hundreds of billions in committed spend will be able to pay back their investors, and later the creditors who lent them the money, if it does not.</p><h3>Read the S-1 before you buy</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s10RjWHzTqM6cLXfmwOxQw.png" /><figcaption>Nasdaq MarketSite, Times Square. Anthropic plans to list here as ANT.</figcaption></figure><p>Anthropic filed its S-1 with the SEC on October 26 and plans to list on Nasdaq as ANT. Its last private round, in May, valued it at $965 billion, and analysts now talk about $1.5 trillion to $2 trillion at listing. OpenAI, valued at $852 billion as recently as March, has filed confidentially and is reportedly raising private money at up to $1.4 trillion. <strong>These will be two of the largest IPOs in history, and they are arriving at exactly the moment I think the pendulum is turning.</strong></p><p>The filing is worth reading , or at least let one of the AI models summarize it for you. Anthropic’s revenue grew fast, to $16.3 billion in the first half of 2026. It also lost $80.7 billion in those six months. About $72 billion of that was an accounting charge tied to its convertible notes, but even without it the loss was roughly $8 billion. Meanwhile its noncancelable purchase commitments climbed from $54.6 billion at the end of 2025 to $265.6 billion by June 30, including $110 billion to Amazon, $114.5 billion for Google’s TPUs, and at least $44.9 billion to Microsoft. That is the build out I described above, written into a contract.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ysKGRB7k9LOh-6GOHn7znw.png" /><figcaption>Left: Anthropic CEO Dario Amodei (right) at TechCrunch Disrupt 2023. Right: OpenAI CEO Sam Altman at TechCrunch Disrupt 2019</figcaption></figure><p>The risk factors say much of the rest. <strong>Anthropic lists open weight models among its competitors.</strong> It is tough to sell against free, especially if models can be run in the privacy of your business or home, i.e. on your servers, laptops, or mobile devices. It notes that Amazon, Google, and Microsoft, its landlords, all build competing models of their own. Seven co-founders keep 50.1% of the vote, so public shareholders are buying the economics without the control.</p><p>I use these models every day and I admire what both companies have built. <strong>That is a separate question from whether their stock is a good buy at a trillion dollars, sitting on hundreds of billions in compute commitments</strong>, while open weights get better every quarter and corporations work out how to keep their questions inside their own walls. I would not buy either IPO. If you want to bet on AI, buy a good laptop and download a model.</p><p><em>This is my opinion, not investment advice.</em></p><h3>Back to the desk, and the phone</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WUQaQIJbW2apdAvlzj188w.jpeg" /><figcaption>Apple’s M1 on a 2020 Mac mini logic board. Chips like this run capable models locally</figcaption></figure><p>The chips in our laptops and phones are already strong enough. Apple’s M series, Qualcomm’s Snapdragon parts, and ordinary desktop GPUs run billion parameter models at useful speed. A 7 billion parameter model fits in a few gigabytes of memory, and quantization squeezes it further with little loss. What needed a server rack two years ago runs on a laptop today. A single workstation can now run the open weight models I described above.</p><p>Run it locally and your questions never leave your device. Nobody logs them. Nobody builds an ad profile from them. Nobody changes the terms of service and starts training on your inputs. For an individual that is privacy. For a business it is compliance and protection of its IP. For anyone handling medical, legal, or classified material it is a requirement.</p><p>Yes, content is king, and these models have an insatiable appetite for it. But they have already absorbed most of what humans have written over our entire history. Add new content locally as it arrives and you get most of the benefit without shipping your data back to a frontier lab.</p><p>That opens the door for local small businesses again, the same way client server did. Customer service, coding help, document review, all running on hardware you already own.</p><h3>From macro to micro</h3><p>AI is transformative. I have no doubt about that. The question is who controls it and where it runs.</p><p>The EV1 was buried by the company that built it, and the idea came back from outside the incumbent fifteen years later. Now the same industry is writing off tens of billions because it built ahead of demand. Electric cars did not disappear. The overbuilding got corrected.</p><p>Compute has followed the same rhythm. Big iron gave way to the desktop. The Cray and the Connection Machine gave way to workstations, then to commodity chips. Client server gave way to the cloud. Each time, the power that sat in one room eventually landed in everyone’s hands.</p><p>I believe AI is next. The weights are open and portable, the chips are already in our pockets, and the models are good enough today. Companies guarding their intellectual property and people guarding their privacy will pull the work back inside their own walls. Power is moving back toward the individual, the small company, and the large corporation that does not want to depend on a handful of highly concentrated entities sitting in the S&amp;P.</p><p>So I expect a large correction in the hyperscale build out, much like the one the automakers are working through now, and I expect the AI IPOs to arrive right at the top of it. Centralized AI may well come back later, when the cycle turns again. It always does. For the next stretch, though, I would bet on the small, the local, and the private, and I would keep my money out of the mainframe.</p><h3><strong>References</strong></h3><h4>Electric cars</h4><ul><li><a href="https://en.wikipedia.org/wiki/General_Motors_EV1">General Motors EV1</a>, Wikipedia</li><li><a href="https://en.wikipedia.org/wiki/Who_Killed_the_Electric_Car%3F">Who Killed the Electric Car?</a>, Wikipedia</li><li><a href="https://www.npr.org/2024/12/06/nx-s1-5116270/ev1-general-motors-electric-vehicle">This little electric car made history. 25 years ago, GM stopped making it</a>, NPR, December 6, 2024</li><li><a href="https://amp.claimsjournal.com/news/national/2026/02/09/335557.htm">Global Carmakers Book $55 Billion Hit From EV Rollback</a>, Bloomberg via Claims Journal, February 9, 2026</li><li><a href="https://www.businessday.co.za/world/international-companies/2026-02-06-stellantis-plunges-on-r427-billion-bill-for-ev-pullback">Stellantis plunges on R427 billion bill for EV pullback</a>, BusinessDay, February 6, 2026</li><li><a href="https://jalopnik.com/2122668/ev-restructuring-cost-automakers-70-billion">EV Restructuring Has Cost Automakers Over $70 Billion So Far</a>, Jalopnik</li><li><a href="https://www.itiger.com/news/2618588031">Honda to Write Off $15.7 Billion as EV Winter Arrives</a>, Barron’s</li></ul><h4>Mainframes, supercomputers and the GPU</h4><ul><li><a href="https://en.wikipedia.org/wiki/IBM_System/360">IBM System/360</a>, Wikipedia</li><li><a href="https://en.wikipedia.org/wiki/Cray-1">Cray-1</a>, Wikipedia</li><li><a href="https://en.wikipedia.org/wiki/Cray_X-MP">Cray X-MP</a>, Wikipedia</li><li><a href="https://en.wikipedia.org/wiki/Cray-2">Cray-2</a>, Wikipedia</li><li><a href="https://en.wikipedia.org/wiki/Connection_Machine">Connection Machine</a>, Wikipedia</li><li><a href="https://en.wikipedia.org/wiki/Silicon_Graphics">Silicon Graphics</a>, Wikipedia</li><li><a href="https://en.wikipedia.org/wiki/GeForce_256">GeForce 256</a>, Wikipedia</li></ul><h4>PC and cloud</h4><ul><li><a href="https://www.ibm.com/history/personal-computer">The IBM PC</a>, IBM History</li><li><a href="https://en.wikipedia.org/wiki/Amazon_Web_Services">Amazon Web Services</a>, Wikipedia</li><li><a href="https://www.bloomberg.com/news/articles/2006-11-12/jeff-bezos-risky-bet">Jeff Bezos’ Risky Bet</a>, BusinessWeek, November 2006</li><li><a href="https://www.cnbc.com/2022/05/20/why-jeff-bezos-keeps-a-reminder-that-aws-was-once-just-a-risky-bet.html">Why Jeff Bezos keeps a reminder that AWS was once just a risky bet</a>, CNBC, May 20, 2022</li><li><a href="https://twitter.com/JeffBezos/status/1527003895393812480">Jeff Bezos on X</a>, May 2022</li></ul><h4>AI build out and open weights</h4><ul><li><a href="https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers">The cost of compute: a $7 trillion race to scale data centers</a>, McKinsey</li><li><a href="https://www.bain.com/about/media-center/press-releases/20252/$2-trillion-in-new-revenue-needed-to-fund-ais-scaling-trend---bain--companys-6th-annual-global-technology-report/">$2 trillion in new revenue needed to fund AI’s scaling trend</a>, Bain</li><li><a href="https://www.iiss.org/publications/strategic-comments/2025/04/deepseeks-release-of-an-open-weight-frontier-ai-model/">DeepSeek’s release of an open-weight frontier AI model</a>, IISS</li><li><a href="https://hai.stanford.edu/policy/beyond-deepseek-chinas-diverse-open-weight-ai-ecosystem-and-its-policy-implications">Beyond DeepSeek: China’s diverse open-weight AI ecosystem</a>, Stanford HAI</li><li><a href="https://en.wikipedia.org/wiki/DeepSeek">DeepSeek</a>, Wikipedia</li><li><a href="https://en.wikipedia.org/wiki/Qwen">Qwen</a>, Wikipedia</li><li><a href="https://openai.com/index/introducing-gpt-oss/">Introducing gpt-oss</a>, OpenAI, August 2025</li></ul><h4>Security, espionage and privacy</h4><ul><li><a href="https://www.infotoday.com/searcher/may01/lambert.htm">Delphion and the former IBM patent server</a>, Searcher, May 2001</li><li><a href="https://www.tprinternational.com/ipi-award-2014-recognizes-stephen-boyer-for-remarkable-advances-in-access-to-patents">IPI Award 2014 recognizes Stephen Boyer (founder of the IBM Intellectual Property Network)</a>, TPR International</li><li><a href="https://thestar.com.my/tech/tech-news/2023/05/02/samsung-bans-staffs-ai-use-after-spotting-chatgpt-data-leak">Samsung bans staff’s AI use after spotting ChatGPT data leak</a>, Bloomberg via The Star, May 2, 2023</li><li><a href="https://extratv.com/2023/05/25/more-companies-are-banning-chatgpt-due-to-privacy-risks">More companies are banning ChatGPT due to privacy risks</a>, May 25, 2023</li><li><a href="https://www.cbsnews.com/sanfrancisco/news/former-google-engineer-linwei-ding-found-guilty-stealing-ai-secrets-chinese-firms/">Former Google engineer found guilty of stealing AI secrets for Chinese firms</a>, CBS News, January 2026</li><li><a href="https://news.bloomberglaw.com/esg/ex-google-engineer-guilty-of-stealing-ai-tech-for-chinese-firm">Ex-Google engineer guilty of stealing AI tech for Chinese firm</a>, Bloomberg Law, January 2026</li><li><a href="https://www.wiz.io/blog/wiz-research-uncovers-exposed-deepseek-database-leak">Wiz Research uncovers exposed DeepSeek database leaking sensitive information</a>, Wiz, January 2025</li><li><a href="https://www.anthropic.com/news/disrupting-AI-espionage">Disrupting the first reported AI-orchestrated cyber espionage campaign</a>, Anthropic, November 2025</li><li><a href="https://openai.com/index/response-to-nyt-data-demands/">How we’re responding to The New York Times’ data demands</a>, OpenAI, June 2025</li></ul><h4>AI IPOs</h4><ul><li><a href="https://www.cnbc.com/2026/10/26/anthropic-ipo-live-updates-ai-s-1-filing.html">Anthropic IPO live updates: S-1 filing</a>, CNBC, October 26, 2026</li><li><a href="https://www.investing.com/news/stock-market-news/anthropic-revenue-hits-163-billion-as-preliminary-s1-outlines-ipo-risks--wsj-4960928">Anthropic revenue hits $16.3 billion as preliminary S-1 outlines IPO risks</a>, WSJ via Investing.com</li><li><a href="https://siliconangle.com/2026/06/08/openai-confidentially-files-ipo-paperwork-inches-closer-stock-market-debut/">OpenAI confidentially files IPO paperwork</a>, SiliconANGLE, June 8, 2026</li></ul><h4>AI and mathematics</h4><ul><li><a href="https://www.scientificamerican.com/article/the-inside-story-behind-openais-million-dollar-math-proof/">The inside story behind OpenAI’s million-dollar math proof</a>, Scientific American, October 22, 2026</li><li><a href="https://www.scientificamerican.com/article/openai-just-scooped-human-researchers-on-a-million-dollar-quantum-math-problem/">OpenAI just scooped human researchers on a million-dollar quantum math problem</a>, Scientific American, October 27, 2026</li><li><a href="https://en.wikipedia.org/wiki/OpenAI_bulk_release_of_mathematical_papers">OpenAI bulk release of mathematical papers</a>, Wikipedia</li><li><a href="https://www.nature.com/articles/d41586-026-03196-8">OpenAI posts 700 maths preprints online: mathematicians are up in arms</a>, Nature</li></ul><h4>IBM Patent Server</h4><ul><li><a href="https://web.archive.org/web/19980213074039/http://patent.womplex.ibm.com:80/">IBM Patent Server home page, February 1998</a>, Internet Archive</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7pCsfPCz_Qs_CMOh.png" /></figure><p><strong>About the Author:</strong> <em>Dr. Steven Muskal is CEO and founder of Eidogen Sertanty, with more than four decades of experience in AI driven drug discovery. He holds a Ph.D. in Chemistry from UC Berkeley (1991), where his thesis focused on neural networks for protein structure prediction.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1f58938e8d10" width="1" height="1" alt=""><hr><p><a href="https://aiadvances.org/dont-buy-into-the-ai-ipos-1f58938e8d10">Don’t Buy into the AI IPOs</a> was originally published in <a href="https://aiadvances.org/">AI Advances</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Moat Is a Puddle]]></title>
            <link>https://medium.com/the-generator/the-moat-is-a-puddle-1e17f362d903?source=rss-7c2cfeb3b1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e17f362d903</guid>
            <category><![CDATA[finance]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Steven Muskal]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 21:44:57 GMT</pubDate>
            <atom:updated>2026-10-25T20:57:20.889Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Why the market is way over its skis.</em></p><p>Seven companies now account for roughly a third of the value of the S&amp;P 500. One sponsored market note puts the Magnificent Seven at a 35% peak in June 2026, against a previous high near 26% set twice before, in the 1980 energy boom and at the dot-com top in March 2000. A third of the largest equity market in the world is levered to one buildout thesis, and most of that exposure sits in index funds that retirement savers never chose stock by stock.</p><p>The comparison being drawn in many podcasts right now is 1873, Jay Cooke, the Northern Pacific, and a founder with a cult following and a bond desk. The parallel is real. We need to stop treating bubbles as one species. Every one of them runs on two engines, a technological engine and a financial engine, and the engine that dominates tells you what breaks, who survives, and what the wreckage is worth afterward.</p><h3>Every Bubble Has Two Engines</h3><p>Carlota Perez made this argument: <strong>the installation phase of a technological revolution and the speculative frenzy of financial capital are coupled</strong>. A study of 51 major innovations introduced between 1825 and 2000 found equity bubbles in 73% of them, with bubble magnitude rising with how radical the innovation was. The technology is usually not the error. The financing of the technology is where the error lives.</p><p>Dominant engine by episode. Both engines run in every case; the question is which one sets the pace.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*J72_05XkTbY4C51UwOkRuA.png" /></figure><p>Two entries in that table have both engines at full throttle. <strong>One of them is 1873. The other is the one we are living in.</strong> That is the reason the 1873 comparison keeps surfacing, and it is a better reason than the resemblance between two charismatic men.</p><p>When the financial engine dominates, as in 2008, the assets left behind are houses and loan books, and they reprice. <strong>When the technology engine dominates, as in 2000, the assets left behind are physical capacity that outlives the companies that built it</strong>, and someone else inherits it cheap. When both engines run hot, you get both outcomes at once: a credit event and a decade of stranded capacity. In 1873 that combination produced a depression that ran, by some measures, six years.</p><h3>The Cell Runs Two Currencies Too</h3><p>Biology made this same split. A<strong>TP is the energy currency of the cell. GTP is the information currency</strong>. ATP<strong> </strong>is the molecule hydrolyzed to do work, to contract a muscle fiber, to pump an ion uphill, to drive a reaction that would not otherwise run. GTP is used by the cell for switching and timing: the heterotrimeric G proteins behind hormone and light signaling, the Ras family of small GTPases, the elongation factors that gate each step of translation.</p><p>The chemical difference is almost nothing. Guanine and adenine are both purines, C₅H₅N₅O against C₅H₅N₅. One oxygen. Counting nitrogen and oxygen sites the way a medicinal chemist does, guanine offers 6 against adenine’s 5, along with one additional hydrogen bond donor. An oxygen takes the place of the amino group, the amino group moves one position over, and a hydrogen settles on the nitrogen between them. Guanine presents a different face, with one more place to make contact.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/950/1*6Q-kC8kN6hjhYzCt9GDwWg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PDtwxTbPYNJs3GvqfbOX5g.png" /><figcaption>The whole difference sits in the base at the top, where guanine carries one additional oxygen and one more hydrogen bonding site than adenine.</figcaption></figure><p>Hydrogen bonds are how molecules speak to each other. They are directional, they are specific, and a protein reads a nucleotide by counting them. The nucleotide that carries information therefore has one more channel of communication than the nucleotide that carries energy, and every extra channel is also an extra place to be perturbed, misread, or hijacked. Signal transduction is built on the richer molecule and it is correspondingly easier to break.</p><p>By analogy, the same asymmetry holds for the two engines of a bubble. The financial engine is energy; it moves capital to where capital is called. The technological engine is information; it carries the thesis, the expectations, and the story that justifies the capital. The information side has more surfaces at which a shock lands. That argues for watching the technology engine more closely than the financing, while conceding that both of them can fail, and that in this cycle both of them are running hot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UG1GeHrLjh7t-KFvdMJcpA.png" /><figcaption>September 20, 1873: the doors of the New York Stock Exchange closed on its own members. The exchange stayed shut for ten days. Wood engraving, 1873</figcaption></figure><h3>Jay Cooke Built Railroads Nobody Needed Yet</h3><p>Jay Cooke &amp; Company underwrote the Northern Pacific Railway on the promise that a northern transcontinental line would unlock settlers, freight, and commerce. The vision was correct. The northern route did get built, it did matter, and the territory it opened is still shipping grain. What was wrong was the pace and the structure. Cooke borrowed heavily and sold bonds faster than track demand justified, on the assumption that revenue would arrive before the obligations did.</p><p>Cooke was not a fraud. He was a believer who financed his belief with other people’s money on a schedule the freight could not meet. His firm failed in September 1873, the New York Stock Exchange closed for ten days, and the Northern Pacific was finally driven to completion in 1883 by other people, with Cooke’s investors long since wiped out.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QWb3M25C_oV4YvAcRNlaSQ.png" /><figcaption>Jay Cooke, photographed during the Civil War, when he had made his reputation distributing Union war bonds to ordinary households. He applied the same distribution machinery to railroad debt.</figcaption></figure><p><strong>Cooke’s role has a name in the cell. Nucleoside diphosphate kinase</strong> takes ATP and hands back GTP, converting the energy currency into the signaling currency at a fixed rate and creating nothing in the exchange. Cooke did the same with capital and belief, through distribution machinery he had already built selling war bonds to households. He did not invent the locomotive and he did not generate the freight. He converted.</p><p>Swap the Northern Pacific for gigawatt data centers, swap the bond syndicate for off-balance-sheet financing vehicles and circular chip-and-cloud supply agreements, and the structural bet is identical. Build the capacity now, on debt, for demand the believers are certain about and nobody has yet metered.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ap6Q9zdIlenaLJtVLjcbPw.png" /><figcaption>1883, ten years after the panic: the Northern Pacific is completed and celebrated, under new ownership. The track was worth building. The financing schedule was not survivable.</figcaption></figure><p>While I agree that this technology is as consequential as rail, I separate two claims that the market has fused into one. That AI will matter enormously is a claim about the technology. That every dollar spent accelerating its physical buildout today will be repaid on the current schedule is a claim about financing. Cooke conflated them too.</p><h3>Remember the Dot-Com Wasteland</h3><p>The 1920s had radio. Sets sold into American homes by the millions, broadcasting was genuinely new, and the equities attached to it were bid to levels that assumed the household penetration curve would never flatten. The technology delivered. The valuations did not survive 1929.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3Zy42xTWRIiBR_z6ZNKSFw.png" /><figcaption>An early 1920s regenerative receiver with horn speaker. Radio was the real thing, and radio equities were still among the casualties of 1929, Charles William Taussig, before 1922.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MJz1PgtK8JNu88rDcTedrg.png" /><figcaption>Wall Street, October 24, 1929. The financial engine set the pace that year; the technologies of the decade, radio, autos, aviation, electrification, were real and were repriced anyway. Associated Press, 1929</figcaption></figure><p><strong>The episode closer to hand is 2000, and the part of it that gets left out is Y2K</strong>. <em>Governments and corporations spent extraordinary sums in 1998 and 1999 on remediation</em>, and much of that money moved through the same vendors, switches, servers, and fiber routes that the internet buildout rode in on. When the remediation cliff arrived in January 2000 and the follow-on revenue did not, the capacity stayed in the ground. Dark fiber sat unlit for years. We had built the track ahead of the freight again.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L3NeyMfFCH1ISoMHeb6ttQ.png" /><figcaption>A buried optical fiber cable in cross-section. The white cores are empty. Capacity installed ahead of demand is the signature artifact of a technology-led bubble, and it is why the survivors of one get cheap infrastructure.</figcaption></figure><p>The absurdities of that period are worth recalling only because they were accompanied, at the time, by universal insistence that they were rational. <strong>Pets.com bought a Super Bowl ad for a business that lost money on every bag of dog food</strong>. Webvan spent close to a billion dollars on warehouses for a model nobody had proven. <strong>Graduate students at name-brand universities advertised as egg and sperm donors</strong> to cover rent while waiting on options that never vested.</p><p>Amazon fell about 95% peak to trough and lived, because underneath the story was a logistics and retail business that generated cash on every order. Google was not even public when the bubble burst; it listed in 2004, after the carnage, on a search advertising model that earned on every query. The survivors had unit economics. The casualties had a story and a burn rate.</p><h3>Weights, Not Wonders</h3><p>I have spent forty years building computational and machine learning models, starting with how small molecules bind to proteins and ending up in the tooling that grew out of that discipline. <strong>The mechanics drive the investment case.</strong></p><p><strong>A large language model is a very large array of numbers</strong>. Billions of floating-point weights, in some cases more than a trillion, arranged in matrices and fitted by gradient descent to predict the next token, until the statistics of that prediction task encode something that behaves like reasoning. <strong>Nothing in that description is structurally uncopyable</strong>. There is no patent-grade secret inside the matrix. Getting there the first time costs an enormous amount of compute and engineering, which is real and deserves to be paid for, and which is not the same thing as a durable structural barrier to entry.</p><p>Chinese labs, DeepSeek most visibly, have used distillation, training a smaller model to imitate a larger one’s outputs, to reach competitive performance at a small fraction of frontier training cost, and have published the weights. Those models run today on a workstation. Four-bit quantization brings a 70 billion parameter model down to roughly 40 GB, which fits in one machine under a desk, with no H100 cluster anywhere in the picture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RIqHpOQC_Y73gxDSszjE8Q.png" /></figure><h3>Dumping, With Weights Instead of Steel</h3><p>We have seen this film before.<strong> China built steel capacity near a billion metric tons a year</strong>, <strong>well past what its own construction demand could absorb, and exported the surplus</strong>. Crude steel output was 960.81 million metric tons in 2025, roughly 52% of the world total of 1.85 billion, while exports hit a record just above 119 million and Chinese hot-rolled coil offered at about $465 a metric ton against $480 from India, $495 from Japan, and $535 from Turkey. <strong>Western producers lost on price set by someone with capacity to clear and a reason to clear it</strong>.</p><p><strong>Open weights are the same move with a crucial difference: the marginal cost of the next copy is zero</strong>. Alibaba’s Qwen family already accounted for more than 30% of all Hugging Face model downloads in 2024, and by August 2025 Qwen derivatives were over 40% of new language-model variants on that hub. A government that wants to compress the margin of American model vendors does not need an export terminal. It needs one upload.</p><h3>Good Enough Is a Business Strategy</h3><p>We are no longer constrained by model capability. We are constrained by deployment (and content).</p><p>The industry’s own behavior says so. The last real jump in usefulness came from spending more compute at inference time, the reasoning models that work step by step before answering, rather than from a larger base model. The pretraining scaling law has hit diminishing returns. Meanwhile the toolkit for extracting value from what already exists has improved faster than the frontier has: quantization that shrinks the memory footprint at minor accuracy cost, speculative decoding, distillation into small specialists, and retrieval that hands a mid-sized model the context of a much larger one. <strong>A well-engineered application on a mid-sized open model now beats what a frontier model could do eighteen months ago</strong> on the tasks a business actually cares about, at a fraction of the cost per call.</p><p>Serving those models has become trivial, and trivial is the operative word. The weights are published, the runtimes are commodity, the hardware is available, and the deployment is a weekend of work for a competent engineer. The scarce thing is integration, product design, data plumbing, and dull reliability engineering. <strong>Spending another quarter-trillion dollars on the next training cluster, while most of the economy has not finished exploiting the model it already has, is Cooke laying a fourth parallel line past a town that has not filled its first three cars</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6tiumNqB3u5jlfieFxy1xg.png" /><figcaption>Racks and cabling in a commercial data center. Training capacity is the asset being financed; inference is where the revenue has to come from, and inference runs on far less than this.</figcaption></figure><h3>We Are Running Out of Things to Read</h3><p>There is a harder ceiling than compute, and it gets far less attention. Epoch AI put the effective stock of quality, repetition-adjusted public human text at about 300 trillion tokens and projected that training runs would consume it between 2026 and 2032, sooner under aggressive overtraining.<strong> Nature reported their estimate that typical training set size reaches the total stock of public online text around 2028</strong>.</p><p>Set that against the models. Parameter counts are now quoted in the hundreds of billions, and <strong>open-weight releases at a trillion parameters have already shipped</strong>. The double-digit trillions of tokens that the web, the libraries, and the archives of human writing represent have largely been absorbed into weight matrices measured in trillions of parameters. The corpus has been read. Access to it is tightening at the same time, as rights holders stop treating crawling as weather.</p><p><strong>Compute can be bought and chips can be fabbed. New writing cannot be manufactured by the people who need it</strong>.</p><h3>The Real Moat Is Proprietary Content</h3><p>Follow that to its conclusion and the defensibility everyone is paying for sits in the wrong place. The weights are copyable. The serving stack is commodity. What is neither is a corpus that nobody else holds and that keeps growing: proprietary data, curated knowledgebases, instrument records, clinical and operational logs, and the working output of people who produce original material for a living.</p><p><strong>Writers, researchers, educators, analysts, and curators should read this as leverage rather than threat</strong>. They are a dependency, not a casualty. Reported licensing terms are already telling: News Corp’s agreement with OpenAI was reported by the Wall Street Journal at more than $250 million over five years, and Reddit’s IPO filing disclosed data licensing contracts with an aggregate value of $203 million over two to three years. Those are the opening prices in a market where the buyers are running out of alternatives.</p><p>The pricing implication is the part creators keep underselling. A corpus rented through low-cost subscriptions earns a few dollars a month per reader. The same corpus licensed for training is an asset sale, and the move from subscription to perpetual license is reasonably a 10 to 100 fold step up in realized value for the contributor.</p><p>Better still, the asset regenerates. Models are retrained frequently, and each cycle needs material that postdates the last one. A<strong> perpetual license on the 2026 corpus does not foreclose a perpetual license on the 2029 corpus</strong>. <em>For a prolific curator, that is an annuity with a new principal payment every cycle</em>, and the way to earn it is to keep producing original, high-quality material that a model cannot source anywhere else.</p><h3>The Pyramid Balanced on One Point</h3><p>Layer the market structure back on top. A handful of hyperscalers and the chipmaker that supplies them carry the index. Beneath them sits a dependent tier of specialty cloud providers, utilities signing long-dated contracts for gigawatt loads, and financing arrangements in which the chip supplier, the cloud provider, and the model company invest in each other and contract with each other, which makes end demand difficult to separate from the financing itself. The dot-com telecom sector ran a milder version of this with capacity swaps between carriers that inflated reported revenue without a single new end customer appearing.</p><p>This is not a forecast about any one company. It is an observation that when a third of an index depends on a buildout thesis rather than on realized cash flow from that buildout, the overconcentration pattern that preceded every prior technology-driven correction is back, with larger numbers and better marketing.</p><h3>The Thermodynamics Are Settled. The Kinetics Are Not.</h3><p>I trained as a chemist before I wrote software, and the instinct that leaves you with is about gradients. <strong>Capital is flowing into AI infrastructure at a rate decoupled by a wide and widening margin from the cash flow that infrastructure currently produces</strong>.<em> A gap like that is a pressure differential, and pressure differentials equalize.</em> A supersaturated solution crystallizes eventually, and the identity of the speck of dust that starts it is almost beside the point.</p><p>What is genuinely open is the rate. Maybe a financing vehicle fails to refinance. Maybe one quarter arrives in which a hyperscaler concedes that capex is not converting to revenue on schedule. Maybe open weights land at a capability that makes a frontier subscription hard to defend on price. I do not know the trigger and I do not know the quarter. I am confident about the direction.</p><h3>Keep the Powder Dry, and Keep Writing</h3><p>Two lessons follow, and they point in the same direction.</p><p>The first is dry powder. <strong>Hold more cash than is comfortable</strong>, more than a model portfolio prescribes, and more than the fear of missing the next 10% wants you to hold. The investors who did best out of 2000 were not the ones who squeezed the last few points out of the rally. They were the ones with money available when Amazon traded in single digits and the survivors were priced at wartime rather than peacetime. The correction, when it comes, will not end AI. It will be the moment when companies with real unit economics buy their way into dominance, the way Google and Amazon did, while the firms that financed capacity ahead of the freight get reorganized quietly.</p><p>The second is content. <em>If you write, research, teach, record, or curate, your material is the input that is actually running short, and it is the only part of this stack that cannot be distilled out of someone else’s model.</em> Produce it, keep the rights to it, catalog it, and price it as the strategic asset it has become. The systems will be back for more in a few years, and they will need what you made after the last time they asked.</p><p>You can stay fully invested and hope the kinetics stay slow. Or you can hold your powder, keep producing the one thing the models cannot generate for themselves, and be solvent and useful on the other side. I know which one I am doing.</p><h3>References</h3><ul><li>Carlota Perez, “The Advance of Technology and Major Bubble Collapses”: <a href="https://carlotaperez.org/wp-content/downloads/media/articles-and-blogs/PEREZTechnologyandbubblesforEngelsbergseminar.pdf">carlotaperez.org</a></li><li>“Two Centuries of Innovations and Stock Market Bubbles”: <a href="https://www.researchgate.net/publication/326424022_Two_Centuries_of_Innovations_and_Stock_Market_Bubbles">ResearchGate</a></li><li>Panic of 1873: <a href="https://en.wikipedia.org/wiki/Panic_of_1873">Wikipedia</a>; Jay Cooke: <a href="https://en.wikipedia.org/wiki/Jay_Cooke">Wikipedia</a></li><li>Dot-com bubble: <a href="https://en.wikipedia.org/wiki/Dot-com_bubble">Wikipedia</a></li><li>S&amp;P 500 concentration, Magnificent Seven peak: <a href="https://www.institutionalinvestor.com/article/sponsored-content/sp-500s-concentrated-rally-starting-diversify">Institutional Investor (sponsored content, CME Group)</a></li><li>Knowledge distillation: <a href="https://en.wikipedia.org/wiki/Knowledge_distillation">Wikipedia</a>; DeepSeek-R1 (arXiv:2501.12948): <a href="https://arxiv.org/abs/2501.12948">arXiv</a></li><li>Epoch AI, “Will we run out of data? Limits of LLM scaling based on human-generated data”: <a href="https://epoch.ai/publications/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data">epoch.ai</a></li><li>China steel output, exports, and price offers, 2025: <a href="https://www.sunsirs.com/uk/detail_news-30925.html">SunSirs</a>, <a href="https://news.metal.com/en/newscontent/103748560-smm-steel-import-and-export-special-topic-outward-leap-under-domestic-demand-restructuring-a-panoramic-analysis-of-the-d">SMM</a></li><li>News Corp and OpenAI licensing terms as reported: <a href="https://malaysia.news.yahoo.com/news-corp-inks-openai-licensing-211417483.html">Yahoo News, citing the Wall Street Journal</a></li><li>AI content licensing deal tracker: <a href="https://llmpulse.ai/blog/ai-content-licensing-deals/">LLMPulse</a></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rzS8-oxC88F0onab.png" /></figure><p><strong>About the Author:</strong> <em>Dr. Steven Muskal is CEO and founder of Eidogen Sertanty, with more than four decades of experience in AI driven drug discovery. He holds a Ph.D. in Chemistry from UC Berkeley (1991), where his thesis focused on neural networks for protein structure prediction.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e17f362d903" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-generator/the-moat-is-a-puddle-1e17f362d903">The Moat Is a Puddle</a> was originally published in <a href="https://medium.com/the-generator">The Generator</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Dexcom Is Live. Apple Health and Health Connect Are Three Hours Behind.]]></title>
            <link>https://levelup.gitconnected.com/your-dexcom-is-live-apple-health-and-health-connect-are-three-hours-behind-37fc83fd8646?source=rss-7c2cfeb3b1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/37fc83fd8646</guid>
            <category><![CDATA[diabetes]]></category>
            <category><![CDATA[apple-health]]></category>
            <category><![CDATA[healthtech]]></category>
            <dc:creator><![CDATA[Steven Muskal]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 21:13:08 GMT</pubDate>
            <atom:updated>2026-11-04T15:27:00.710Z</atom:updated>
            <content:encoded><![CDATA[<h4>Closing the three-hour glucose gap on iPhone and Android, and what I’m learning from meals, voice notes, and face check-ins.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*T5IdYkwVYHv-rJzL.png" /><figcaption><a href="https://foodhealthscan.com/">Food Health TxD</a> home on iPhone (left) and Android (right): today’s glucose and carbs, a mini trend chart, and the four actions, Scan Food, Voice Note, Scan Menu, Log Glucose. Screenshots as shown on the App Store and Play Store.</figcaption></figure><p>If you wear a Dexcom Stelo, G7, or ONE+ and you have ever pulled up your glucose trend in a third-party app only to notice the timestamp is hours behind the clock on the wall, you are not imagining it, and you are not doing anything wrong. <strong>The sensor is fine. The handoff between apps is the slow part.</strong></p><h3>The gap is in the handoff</h3><p>Dexcom’s sensor samples interstitial glucose about every five minutes and sends each value to the Dexcom app over Bluetooth. That part works exactly as advertised. The delay shows up one layer downstream, when the Dexcom app exports those values to Apple Health on iPhone or Health Connect on Android.</p><p><strong>That export is held back on purpose.</strong> Dexcom’s own support pages say so, in one sentence each: for the G7, “Apple Health receives your Dexcom G7 information after a three-hour delay, not in real time”; for Stelo, “Stelo sends your glucose data to health apps with a 3-hour delay”; for Health Connect on Android, “Data is shared with a 3-hour delay, not in real time.” None of the three pages says why. <strong>My reading is that it is a regulatory line, not a technical one.</strong> The Dexcom app is the cleared medical display, and its alerts are the ones people act on. Apple Health and Health Connect are general stores that any app can read, and a three-hour hold makes sure no app downstream of them can be mistaken for a real-time glucose monitor. That is a defensible choice for Dexcom. It is also why every app that reads from Apple Health or Health Connect shows you a number that is three hours old.</p><p>For a once-a-day glance at the trend, that lag barely registers. For anyone trying to connect a meal to the glucose response it caused, <strong>three hours is the difference between a chart that teaches you something and a chart that is noise.</strong> You eat lunch at noon. The readings that follow reach Apple Health about three hours later. That wait makes it harder to connect what you just ate with the response you are seeing.</p><p>I have spent 35-plus years building software at the intersection of biology and computation, and this is a timestamp problem, which makes it a solvable one.</p><h3>How the direct read works, and whose shoulders it stands on</h3><p>Food Health TxD is the diabetes companion to Food Health AI: Scan &amp; Score, built by Eidogen-Sertanty for people managing Type 1, Type 2, or prediabetes. It scans a meal photo for carbs and macros and lines the meal up against your glucose on one chart. The feature relevant here is a beta called Direct Sensor Reading, and I want to be specific about where it comes from, because it is built on two open-source projects the diabetes community has maintained for years.</p><p><strong>On iPhone,</strong> the direct read is built on the engine of xDrip4iOS ( <a href="https://github.com/JohanDegraeve/xdripswift">https://github.com/JohanDegraeve/xdripswift</a>), the community CGM app. xDrip4iOS can connect to a Dexcom G7, ONE+, or Stelo over native Bluetooth in what it calls Coexistence mode: the Dexcom app keeps owning the sensor, and the listener joins the same Bluetooth session. No pairing code, nothing changes in the Dexcom app. When you turn on Direct Sensor Reading in Settings, TxD does one job with that engine: it listens for each reading and writes it to Apple Health the moment it arrives, with backfill. The Bluetooth transmitter code is the community’s work, and the license is GPL-3.0, same as upstream.</p><p><strong>On Android,</strong> the direct read is built on xDrip+ ( <a href="https://github.com/NightscoutFoundation/xDrip">https://github.com/NightscoutFoundation/xDrip</a>), the Nightscout Foundation’s CGM engine. Android only lets one app hold a sensor’s Bluetooth connection, so the recommended path uses xDrip+’s notification-based collector: the Dexcom app keeps the sensor, and TxD reads each reading off the Dexcom app’s notification and writes it to Health Connect. For people who want the listener to own the sensor outright, the direct Bluetooth path is there too, through the optional real-time bridge in Settings.</p><p><strong>Either way, sensor data you have already paid for and are already wearing reaches your chart in minutes instead of hours.</strong> A meal and the rise it caused end up on the same timeline.</p><p>Things to be direct about, because this is a health app and the caveats matter as much as the feature:</p><ul><li>Direct reading is a beta and it is off by default. You turn it on deliberately in Settings.</li><li>TxD never pairs with, writes to, or takes over your sensor. The Dexcom app stays in charge of your sensor and your alerts. Keep those alerts on.</li><li>TxD’s own notices (readings stopped or resumed, sensor status, lows, highs, fast rises or drops against your own target range) are convenience notices and do not replace medical alarms.</li><li><strong>Nothing in TxD is a glucose estimate or a prediction.</strong> Values come from your sensor, from Apple Health or Health Connect, or from your own manual entry. The app is for education and tracking, and none of it replaces your healthcare provider.</li><li>Every other CGM that writes to Apple Health or Health Connect keeps working in TxD as before. Abbott Lingo and Libre 3 use encrypted Bluetooth and stay on the delayed Health path.</li></ul><p>Dexcom, Stelo, G7 and ONE+ are trademarks of Dexcom, Inc. Food Health TxD is not affiliated with or endorsed by Dexcom, and it is not a medical device.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Cqs278DCU4n5Aw7e.png" /><figcaption>Left: Food Health TxD on iPhone, Settings, Direct Sensor Reading on, a Dexcom Stelo read by the app itself, 127 mg/dL 3 min ago. Right: the same section on Android, Settings, Real-time CGM bridge, with the two paths, read from the Dexcom app’s notification or listen to the sensor directly. iPhone screenshot as shown on the App Store; Android captured from the 2.2.0 build.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0aORKjGhs2G98KAp.png" /><figcaption>The day chart on iPhone (left) and Android (right): glucose on the left axis, carbs as bars at meal times, tap or drag to read any moment. This is the chart the three-hour delay smears.</figcaption></figure><h3>Why minutes matter: the voice and face check-ins</h3><p>Once readings arrive in near real time, a second thing becomes possible, and it is the reason the Voice Note button sits on the TxD home screen. I wrote about the case for talking to a computer instead of typing at it in a separate piece (Why I Started Talking to My Computer Instead of Typing at It, linked below). The short version: <strong>the human voice carries information about the body that a keyboard never captured</strong>, and the research on voice and glucose is real, recent, and unfinished.</p><p>What the literature says, in brief:</p><ul><li>Kaufman and colleagues at Klick Labs (Mayo Clinic Proceedings: Digital Health, 2023) recorded 267 people saying a fixed six-second phrase up to six times a day, 18,465 recordings in all, and classified Type 2 diabetes from 14 acoustic features at 0.89 accuracy in women and 0.86 in men, with age and BMI in the model.</li><li>The same group (Scientific Reports, 2024) put 505 people on a CGM for two weeks with the same fixed phrase, and found that within a person, fundamental frequency rises about 0.02 Hz per mg/dL of glucose. Small, real, and the authors concluded any predictive model would have to be personalized.</li><li>Lehmann and colleagues (Diabetes Care, 2026) collected 540 smartphone recordings from 22 adults with Type 1 diabetes under clamped normal and low glucose and detected hypoglycemia from sustained vowels with machine learning.</li><li>Hepkarsi and colleagues (Journal of Voice, 2026) measured vocal shimmer in 101 pregnant women and found it elevated in gestational diabetes.</li><li>Hamdan and colleagues’ 2024 systematic review and meta-analysis found no significant group difference in the classic single measures (jitter, shimmer, harmonics-to-noise ratio, F0) between Type 2 diabetes and controls. The machine-learning papers find signal where single-feature group comparisons do not, which says the signal is distributed and small.</li><li>Shukla and colleagues (Journal of Medical Internet Research, 2026) audited the Bridge2AI voice dataset and showed that age imbalance alone can produce the high accuracies reported for voice screening. Demographic confounding is a live risk in this field.</li></ul><p>Two things stand out from that list. First, the within-person, longitudinal question (does my voice track my glucose over time) is the one that matters for a person wearing a CGM, and it is the one with the thinnest evidence. Second, we checked every public repository we could find (Zenodo, PhysioNet, Synapse, figshare, Kaggle, Hugging Face and the rest) and <strong>there is still no public dataset that pairs raw voice or face captures with glucose values.The only labeled voice-plus-CGM data anyone will have is first-party data, collected by the person whose glucose it is.</strong></p><p>We ran two of our own experiments before building anything into the app. On the public VOCADIAB voice-embedding set (607 adults, Type 2 yes/no), a random forest on voice alone reached an AUC of 0.64 to 0.71 depending on sex, and age plus BMI alone did as well or better. The published cross-sectional accuracies are demographics plus voice. Then, with no audio at all, we lined up 1,359 of my own dictation transcripts against Stelo readings within 15 minutes: speaking rate correlated with glucose, most of it was time of day, and a small residual survived after removing day and hour. Real, tiny, per-person. Exactly the shape of the Klick F0 result, and enough to justify collecting audio.</p><h3>What the app does with a check-in</h3><p><strong>So TxD collects the data the field is missing, on the phone, and it does the first modeling there too.</strong></p><p>A face check-in is a front-camera photo, once a day or when you choose. A voice check-in is a clip of up to 30 seconds, either a free note about a meal or your day, or a fixed read-aloud passage so clips are comparable from one day to the next. After a face capture the app offers that passage, so voice and face are captured at the same moment. When a direct reader is active and your glucose is rising or falling fast, the app can also ask for a photo or a note right then, so the outer bands of your range get sampled too. All of this is off by default and lives under Settings, Voice &amp; Face.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Fh_wKmONFrFQSIQY.png" /><figcaption>Left: the Voice Note screen with its three prompts (describe a meal, how’s your day, fixed phrase). Center: Captures, with the Research status card counting labeled voice and face captures toward the 30 recommended, and the Patterns &amp; model button. Right: on Android, a meal entry whose voice note is labeled with the nearest reading, 99 mg/dL.</figcaption></figure><p>Each capture is labeled with the nearest glucose reading within 15 minutes, with the slope and the number of readings, and the label stays provisional until 17 minutes after the capture so the readings after it count as well. That labeling is the part the three-hour delay would have broken. With the direct read, a label is available within minutes of the capture.</p><p>On the phone, every voice clip is reduced to nine acoustic features: F0 mean and spread, jitter, shimmer, intensity, voiced and pause fractions, spectral centroid, and words per second from the transcript. The pitch tracker is autocorrelation on 25 ms windows and was checked against Praat on real clips. Every face photo is reduced to ten features from Apple’s Vision landmarks and CIELAB color patches: eye openness, brow-to-eye gap, mouth aspect, head roll, under-eye and cheek and forehead lightness, periorbital darkness, and cheek redness and yellowness. No geometric warping of the face, and the photo itself goes nowhere.</p><p>Once a person has 10 labeled captures over 3 days (30 recommended), the Patterns page runs three analyses on the device:</p><ol><li>A Spearman correlation of every feature against mg/dL, with a day-block permutation test, so many clips on one day cannot masquerade as many independent days.</li><li>A leave-one-day-out ridge regression on standardized features, compared against predicting the training days’ mean. If ridge does not beat that, the verdict says so.</li><li>A three-band random forest (low, baseline, high), with the bands cut at the sensor’s own 15th and 85th percentiles for that person, evaluated leave-one-day-out with accuracy, Matthews correlation, per-band precision and recall, and a permutation p. The same model runs on the phone and desktop.</li></ol><p><strong>Transcription uses Apple’s on-device recognizer with on-device recognition required, so audio never leaves the phone.</strong> Voice and face captures, their features, their labels, and the model run all stay on the device, and a TxD export carries them if you want to analyze them yourself. Nothing on the Patterns page is a glucose estimate. It is a per-person research readout, and it says in plain words when there is no signal yet.</p><p>That last point is where my own data sits today. The most recent run on my phone covers 72 voice captures over nine days, labeled between 88 and 142 mg/dL. Pitch had the largest correlation with glucose, a Spearman rho of 0.29, and nothing passed a permutation p of 0.05. Leave-one-day-out ridge predicted held-out days with a mean absolute error of 8.87 mg/dL against 8.92 for the training-days mean, and the three-band forest scored 0.25 on held-out days.</p><p>Rho measures how consistently two quantities rise or fall together, on a scale from -1 to 1. My pitch correlation of 0.29 is a loose positive association, as the scatter below shows. The permutation p asks a harder question: could day-to-day drift alone produce a rho this large? The app shuffles the day-mean glucose labels across days 400 times and counts how often the shuffled data correlates at least as well. For pitch that happened 1 time in 10, so p is 0.10. The reason the test works in day blocks is visible in the figure. The clips from 10/09/26 sit at a mean pitch of 117 Hz against 92 to 96 Hz on every other day, and they also hold the highest glucose. Leave that one day out and rho drops to 0.14. Within single days the sign is not stable either, from 0.73 on 10/13 to -0.77 on 10/18. A second pass over the same clips on the Mac points at the likely culprit: the relationship appears in the iPhone recordings (0.38) and not in the Android ones (-0.10), which says microphone and processing, and the apparent slope of about 0.4 Hz per mg/dL is roughly 19 times what Kaufman’s 505-person study measured. <strong>An effect that much larger than a large study found is a warning sign.So the honest reading is a small positive tendency that cannot yet be separated from whatever else changed between days</strong>: hydration, time of day, a cold, which phone was held. The literature says the within-person effect is on the order of 2 Hz of pitch per 100 mg/dL against day-to-day pitch noise of 10 to 20 Hz, which is why the instrument needs many more labeled clips, spread across many more days and across the full glucose range, before it can see anything. That is what the app is collecting, one check-in at a time, and the number on the screen will change only when the data does.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lgJYNMJeI2PmcO2_.png" /><figcaption>The on-device Patterns run from my phone, 10/19/26: 72 voice captures over 9 days. Left, Spearman rho of each acoustic feature with the labeled glucose, dark bars where the day-block permutation p is below 0.10. Right, mean pitch against glucose, one point per capture, colored by day. Drawn from the app’s own model_run.json.</figcaption></figure><h3>What two weeks of my own data say so far</h3><p>Voice and face are the newest capture, but the app has been pairing meal photos with the sensor for longer, so I ran all three channels through the same two-week window on the Mac. The food scans carried the clearest signal.</p><p><strong>Meals.</strong> 125 scanned meals, with scans less than 20 minutes apart counted as one meal. 95 had sensor coverage from before eating to three hours after, and 47 had no other carb-containing food logged in the following two hours, so each rise in that set comes from one meal. The typical response was a median rise of 23 mg/dL, peaking about 65 minutes after eating. Carbs set a threshold:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*G1aLsMZvt57ttm03.png" /></figure><p><strong>Every meal under 30 g of carbohydrate kept my glucose rise under 30 mg/dL, while about 40% of larger meals went past it.</strong> A model using the scan’s nutrients plus starting glucose and time of day predicted the size of each rise 28% better than guessing my average rise, with predicted and actual rises correlating at 0.73; carbs alone gave 15% and 0.55. Each day was predicted using only the other days. <strong>The food that spiked me most was a snack:</strong> Trader Joe’s dark chocolate nut bars, about 45 g of carbs, produced my two largest rises (+82 and +41 mg/dL), more than meals with twice the carbs, and across the 7 meals that included them the average rise was 41 mg/dL, the highest of any food I eat often. This echoes the Weizmann Institute’s finding that responses to the same food differ markedly between people (Zeevi et al., Cell 2015). A food-by-food spike list is not yet possible: most foods appear in fewer than 10 clean meals. The app could flag a likely spiker once a food has been eaten about 5 to 10 times, and the next feature on the list is a post-scan line that reads “likely spike: +30 to 40 mg/dL, peak in about an hour”.</p><p><strong>Cheek redness.</strong> The best lead in the face channel. Redness in my cheeks, measured from the daily selfie, rose with glucose at a correlation of 0.51 across 24 photos on 10 days, with a day-to-day uncertainty range of 0.18 to 0.75. It holds within a single day (0.45), so it is not good days against bad days, and it holds after removing photo brightness and time of day (0.49). It may be ordinary flushing after eating, since food raises both skin blood flow and glucose. Blanes-Vidal and colleagues (Journal of Diabetes Research, 2019) found that people with diabetes showed more pulse-to-pulse variation in facial redness measured from video, while average redness did not differ from controls. TxD measures average redness from a single photo, so that paper supports the idea without confirming this finding.</p><p><strong>Voice pitch.</strong> A correlation of 0.26 across 71 recordings, with a plausible range that includes zero (-0.02 to 0.56), and the phone split described above. Kaufman’s group concluded that voice-based glucose models will need to be personalized to each person, which is the case for collecting more of my own recordings rather than fewer.</p><p>Every figure in this section comes from one person over about two weeks, in a narrow glucose range. <strong>They are leads to follow, and the point of the app is that the same analysis runs for each user on their own data.</strong></p><h3>Available on both platforms, today</h3><p>TxD runs on iPhone (iOS 17.0 or later) and on Android, with a free trial and a subscription after that. The direct Dexcom read and the voice and face check-ins are live on both. If you have been quietly annoyed at your glucose chart for months without knowing why, the fix is a Settings toggle away.</p><p><strong>Reader note: Free access codes are available, one per person.</strong> Pick iPhone or Android, enter your email, and the code comes to you. We ask code recipients to share a few meals with the Food Health community.</p><p><a href="https://foodhealthscan.com/free"><strong>Get a free access code at foodhealthscan.com/free →</strong></a></p><p><a href="https://foodhealthscan.com/txd/#dexcom"><strong>See Direct Dexcom Reading on Food Health TxD →</strong></a></p><h3>The take-home</h3><p><strong>Your sensor was never the slow part.</strong> The handoff between apps was, and two open-source communities had already solved it. Once readings arrive in minutes, a glucose label can be attached to something as ordinary as a voice note, and the dataset nobody has can start to exist on your own phone.</p><h3>References</h3><ul><li>Food Health TxD, Direct Dexcom Reading: <a href="https://foodhealthscan.com/txd/#dexcom">https://foodhealthscan.com/txd/#dexcom</a></li><li>Dexcom, “How do I share my G7 glucose data with the Apple Health app?”: <a href="https://www.dexcom.com/en-us/m/faqs/how-do-i-share-my-glucose-data-with-the-apple-health-app">https://www.dexcom.com/en-us/m/faqs/how-do-i-share-my-glucose-data-with-the-apple-health-app</a></li><li>Dexcom, Stelo, “How do I share my Stelo glucose data with health apps?”: <a href="https://www.stelo.com/en-us/faqs/using-stelo/how-do-i-share-my-glucose-data-with-health-apps">https://www.stelo.com/en-us/faqs/using-stelo/how-do-i-share-my-glucose-data-with-health-apps</a></li><li>Dexcom, “Can I share my glucose data with Health Connect?”: <a href="https://www.dexcom.com/en-gb/faqs/how-do-i-share-my-glucose-data-s-health-app">https://www.dexcom.com/en-gb/faqs/how-do-i-share-my-glucose-data-s-health-app</a></li><li>xDrip4iOS (xdripswift), the engine behind the direct read on iPhone: <a href="https://github.com/JohanDegraeve/xdripswift">https://github.com/JohanDegraeve/xdripswift</a></li><li>xDrip+, the Nightscout Foundation engine behind the direct read on Android: <a href="https://github.com/NightscoutFoundation/xDrip">https://github.com/NightscoutFoundation/xDrip</a></li><li>Muskal, S. “Why I Started Talking to My Computer Instead of Typing at It.” October 10, 2026. Substack: <a href="https://www.drstevenmuskal.com/p/why-i-started-talking-to-my-computer">https://www.drstevenmuskal.com/p/why-i-started-talking-to-my-computer</a></li><li>Muskal, S. “The Metabolic Crisis Is Not a News Story.” May 6, 2026. Medium: <a href="https://medium.com/@smuskal/the-metabolic-crisis-is-not-a-news-story-3ea7ffada493">https://medium.com/@smuskal/the-metabolic-crisis-is-not-a-news-story-3ea7ffada493</a> · Substack: <a href="https://www.drstevenmuskal.com/p/the-metabolic-crisis-is-not-a-news">https://www.drstevenmuskal.com/p/the-metabolic-crisis-is-not-a-news</a></li><li>Kaufman JM, et al. Acoustic Analysis and Prediction of Type 2 Diabetes Mellitus Using Smartphone-Recorded Voice Segments. Mayo Clinic Proceedings: Digital Health, 2023. doi:10.1016/j.mcpdig.2023.08.005</li><li>Kaufman JM, Jeon J, Oreskovic J, Fossat Y. Linear effects of glucose levels on voice fundamental frequency in type 2 diabetes and individuals with normoglycemia. Scientific Reports 14, 2024. doi:10.1038/s41598–024–69620-z</li><li>Lehmann V, Hilpert M, Mostaani Z, et al. Listening to Hypoglycemia: Voice as a Biomarker for Detection of a Medical Emergency Using Machine Learning. Diabetes Care, 2026. doi:10.2337/dc25–1680</li><li>Hepkarsi S, Bildaci TB, Yavas Z. Vocal Shimmer: A Potential Clinical Biomarker for Gestational Diabetes. Journal of Voice, 2026. doi:10.1016/j.jvoice.2026.04.040</li><li>Hamdan AL, Abi Zeid Daou C, et al. Effect of Diabetes Mellitus Type 2 on Voice: A Systematic Review and Meta-Analysis. Folia Phoniatrica et Logopaedica, 2024. doi:10.1159/000534271</li><li>Elbéji A, et al. A voice-based algorithm can predict type 2 diabetes status in USA adults: Findings from the Colive Voice study. PLOS Digital Health, 2024. doi:10.1371/journal.pdig.0000679</li><li>Shukla S, Naliyatthaliyazchayil P, Gichoya JW, et al. Demographic Confounding in Voice-Based Parkinson Disease Screening: Methodological Analysis of the Bridge2AI Voice Dataset. Journal of Medical Internet Research, 2026. doi:10.2196/95609</li><li>Zeevi D, Korem T, Zmora N, et al. Personalized Nutrition by Prediction of Glycemic Responses. Cell, 2015. <a href="https://www.wisdom.weizmann.ac.il/~eran/zeevi_cell_2015.pdf">https://www.wisdom.weizmann.ac.il/~eran/zeevi_cell_2015.pdf</a></li><li>Blanes-Vidal V, et al. Invisible color variations of facial erythema: a novel early marker for diabetic complications? Journal of Diabetes Research, 2019. <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC6745171/">https://pmc.ncbi.nlm.nih.gov/articles/PMC6745171/</a></li><li>Guermazi B, et al. V2D: Detection of type 2 diabetes using only smartphone voice recordings via spectro-temporal transformer embeddings. IEEE Journal of Biomedical and Health Informatics, 2026. doi:10.1109/JBHI.2026.3730729</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JRErxVwr4EvjQv8_.png" /></figure><p>Steven Muskal, Ph.D. is the CEO of <a href="https://www.eidogen-sertanty.com/">Eidogen-Sertanty, Inc.</a> — a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at <a href="http://stevenmuskal.com/">stevenmuskal.com</a></p><p><em>Originally published at </em><a href="https://www.drstevenmuskal.com/p/your-dexcom-is-live-apple-health"><em>https://www.drstevenmuskal.com</em></a><em> on October 19, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=37fc83fd8646" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/your-dexcom-is-live-apple-health-and-health-connect-are-three-hours-behind-37fc83fd8646">Your Dexcom Is Live. Apple Health and Health Connect Are Three Hours Behind.</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Revisiting Old Work]]></title>
            <link>https://pub.towardsai.net/revisiting-old-work-8241cf3ba308?source=rss-7c2cfeb3b1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/8241cf3ba308</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[drug-discovery]]></category>
            <category><![CDATA[biotechnology]]></category>
            <dc:creator><![CDATA[Steven Muskal]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 03:26:22 GMT</pubDate>
            <atom:updated>2026-10-19T06:08:54.293Z</atom:updated>
            <content:encoded><![CDATA[<h4>Twenty five years of my own claims, retested, and the three things they all turned out to share</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WFhIHkh8ef-aGuIP.png" /></figure><p>I have been on a tear. For the last few months I have been going back through work I did over the last twenty five years, some of it from 1998, some from 2003, and rebuilding it. Not for nostalgia. To find out whether it still holds.</p><p>Science is supposed to be hypothesis testing and validation. We seem to have quietly stopped applying that to our own back catalogs. You publish, the paper goes into the record, you move to the next problem, and the claim sits there untested against anything larger than the dataset you had at the time. Most of those claims were built on what was available then, which usually means a few dozen to a few thousand measurements. The data available today is often an order of magnitude larger. That is the part people underrate. Content is still king. New methods are interesting, but the reason a twenty five year old hypothesis can be meaningfully retested in 2026 is that there is now enough public data to put real weight on it.</p><p>So I went back and checked. Five preprints later, here is what I found, and what the process itself taught me. They also turned out to share three things I did not plan on, and I will come back to those.</p><h3>The loop runs the other way now</h3><p>The sequence I grew up with went like this. Do the work. Write it up. Submit. Wait months for review. Publish. Then field reprint requests one email at a time, for years. Eventually someone asks for the code or the data, and by then the grant has closed, the group has moved on, and the technology stack it ran on no longer exists. The science was sound and the access was terrible.</p><p>Here is the loop I am running now:</p><ol><li>Rework the code. Agentic coding makes it realistic to resurrect a project that would otherwise require weeks of archaeology through old scripts and dead dependencies.</li><li>Revalidate on a larger dataset. The original hypothesis gets tested against data that did not exist, or was not public, the first time around.</li><li>Stand up a website. Open access, free to use, downloadable, on GitHub where it fits, before any reviewer has seen it.</li><li>Draft the preprint and post it to bioRxiv. Then submit to a journal at some point.</li></ol><p>Access comes first and the formal record catches up. Peer review still matters to me, and every one of these is in a journal pipeline. It is simply no longer the gate that stands between the work and the people who might use it.</p><h3>1998: six descriptors for human intestinal absorption</h3><p>In 1998, Matt Wessel, Peter Jurs, John Tolan and I published a model that predicted percent human intestinal absorption from molecular structure. This paper has been cited by 500 other papers. It was good work. I am fairly certain it helped get Matt into Pfizer in his early career. Eighty six compounds. Six descriptors chosen by a genetic algorithm with a small neural network as the fitness function.</p><p>The 1998 result rested on a single division of the compounds into training, cross validation and external prediction sets. Repeat that division 250 times and only one of the 250 draws matches the external prediction error we published. We had a fortunate split. A single split was the convention in 1998, but with today’s compute we can do so much more.</p><p>Then the real question. Do those six descriptors still carry the signal on a dataset nine times larger? Booyah — they do. On 780 compounds from a public compilation, the six 1998 descriptors under a modern random forest give an extraction error of 24.72 %HIA units at R 0.63. A full modern descriptor search, choosing 25 consensus descriptors over the same 250 extractions, gives 20.47 at R 0.76. The improvement is real and it comes almost entirely from having a larger descriptor pool to choose from, not from the original six being wrong. In fact, we had to keep the number of descriptors down because we had so few examples and we did not want to overdetermine the system. <em>This is something to keep in mind with today’s LLMs which are approaching double-digit trillions of parameters while the number of examples they’re trained on is also in the double-digit trillions. </em><strong><em>Indeed the frontier AI companies are running out of content</em></strong><em>, but that’s a whole different discussion for another post </em>. The 1998 intuition about which molecular properties govern absorption survived a nine fold increase in data and twenty eight years of method development.</p><p>The deployed model (<a href="https://predicthia.ai/">predicthia.ai</a>) places a compound on the correct side of the 30.00 percent absorption threshold in 92.64 percent of predictions, which is the number that matters if you are using this to triage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lQ77RtqU2xQAXG_T.png" /><figcaption><em>Distribution of extraction error over 250 stratified extractions using the published six descriptors and the 1998 network. Left, the original 86 compounds. Right, the larger set of 780. The dashed line marks the external prediction error published in 1998. Figure 1 from Muskal and Wessel, Human Intestinal Absorption Revisited, bioRxiv, doi 10.64898/2026.10.03.756481, CC BY 4.0</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/671/0*9nP4cmrrCG5Jit13.png" /><figcaption><em>Predicted against measured absorption for the deployed model and the model it replaces. The marked points are the original 1998 training, cross validation and external prediction compounds. Figure 4 from the same preprint.</em></figcaption></figure><p>The method also extended to a problem the 1998 paper never touched. The same procedure, unchanged, now fits permeability across Caco-2 monolayers for 3,771 small molecules and 1,281 cyclic peptides, and across artificial membranes for 7,298 cyclic peptides. One procedure, four endpoints, including a molecular class that barely existed as a drug modality when we started. The model, the descriptors and an interactive predictor are at <a href="https://predicthia.ai/">predicthia.ai</a>.</p><p>Matt Wessel is a co-author again, twenty eight years later. He is at Thermo Fisher now. That reconnection was one of the real pleasures of this year.</p><h3>2003: let a compound’s neighbors vote</h3><p>In 2003 we showed that a compound’s toxicity can be gauged from the toxicities of the compounds it resembles. The structural neighbors form a consortium and vote. It worked, and it ran on a proprietary reference set that is no longer public, which is exactly the failure mode I described earlier. The method outlived its data.</p><p>Twenty three years later the public data is there. Rebuilt on 11,989 curated rat oral LD50 values, the method answers for 67.0 percent of held out compounds and 92.1 percent of those answers land within one log unit of the measured value. In 2003 the figures were 94 percent within one log unit on 57.3 percent of compounds. The method answers for more compounds now at essentially the same accuracy, on data it has never seen and that we did not own.</p><p>Then we pushed it past where the original went. The same machinery, with nothing retuned except the similarity a neighbor must meet before it may vote, now runs over 28 further toxicity endpoints: 149,224 measurements on 34,162 compounds, from cytochrome inhibition through mutagenicity to clinical toxicity. On the compounds it has close enough neighbors to answer for, it correctly identifies a median 72.9 percent of those that showed the toxicity and 96.5 percent of those that did not.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YDjGwlEXnP-MLMka.png" /><figcaption><em>The 2003 machinery applied to 28 further endpoints from TOXRIC. No setting is retuned except the required similarity, which every endpoint sets by the same rule. Figure 2 from Muskal, Jha, Kishore and Tyagi, Toxicity by Consortium Revisited, bioRxiv, doi 10.64898/2026.10.04.756566, CC BY 4.0</em></figcaption></figure><p>The second signal in this revision is the one I did not expect. We asked whether a reverse screen carries information about toxicity beyond molecular size and lipophilicity, which is the usual explanation for why some compounds misbehave everywhere. Screen a compound against 27,910 ligands solved inside proteins and count how many distinct protein families those matches reach. A compound whose matches reach no protein family inhibits CYP3A4 2.5 percent of the time. A compound whose matches span sixteen or more families inhibits it 51.2 percent of the time. Added to heavy atom count, molecular weight and cLogP, the screen raises the mean area under the ROC curve from 0.72 to 0.76 across all 28 endpoints. So yes, the more families a molecule may interact with, the more likely it will have toxicity. <strong>Indeed, promiscuity often kills (a project).</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cfh8j_-KclHuql3X.jpeg" /><figcaption><em>Every compound is screened against 27,910 ligands solved in protein structures, and families reached counts the distinct protein families those matches belong to. Eight endpoints where the active rate rises with reach. From toxpred.ai, where the same scoring runs interactively</em></figcaption></figure><p>Then came the part that makes it a real test rather than a retrospective one. The OpenADMET consortium has since measured CYP3A4 inhibition on 2,335 compounds. The pattern held. Compounds whose matches reach no protein family inhibit CYP3A4 7.3 percent of the time, and those reaching sixteen or more families inhibit it 24.6 percent of the time.</p><p>That is strong enough to act on before anything is made. Score a library on how far its compounds reach across protein families together with their size and lipophilicity, set aside the riskiest fifth, and you remove two thirds of the androgen receptor binders and about half of the CYP3A4 inhibitors. Scoring needs a two dimensional structure and nothing else. No docking, no conformer generation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*986mIbbICGTULR43.png" /><figcaption><em>Running a library through the score, measured on compounds it was never fitted on. Figure 4 from the same preprint.</em></figcaption></figure><p>Interactive scoring, the data and the compound pool are at <a href="https://toxpred.ai/">toxpred.ai</a>.</p><p>Sanjiv Kumar Jha, Phani Kishore and Prashant Tyagi are on this one. All three were at my first startup company, and Prashant and I had worked together earlier at Affymax. They are at AM Green, DigitalOcean and Palo Alto Networks now.</p><h3>Pharmacophore fingerprints without the conformers</h3><p>A three dimensional pharmacophore fingerprint records the binding features a molecule can present. I have described it as a hand in search of a glove. Because it is defined by presented features rather than two dimensional structure, it finds similarity between compounds with unrelated scaffolds, which is what scaffold hopping and non obvious me too design actually require.</p><p>It stayed a niche tool for one reason: cost, and the cost is conformer generation. In the reference pipeline, generating 100 conformers takes 2.82 seconds of the 2.86 seconds needed to fingerprint a single screening compound. The bit calculation itself takes 0.039 seconds. Ninety nine percent of the expense is spent producing conformers you then throw away.</p><p>So we removed the conformational stage entirely. PharmCast is a feedforward neural network that predicts the 10,549 bit fingerprint directly from two dimensional structure. Same descriptor, same meaning, no conformers. Once again, a neural network learned very complex chemistry and physics, all embedded in a set of numerical descriptors or weights. Booyah. This theme has been repeated and exploited many times throughout my career.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fKGnrTSVF0hPH3Ap.png" /><figcaption><em>The two routes to a fingerprint, shown for saquinavir and indinavir, HIV-1 protease inhibitors of unrelated scaffold. The conventional route generates a conformer ensemble and applies the reference calculation. Figure 5 from Muskal and McGregor, PharmCast, bioRxiv, doi 10.64898/2026.09.02.748999, CC BY 4.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KMP4iHQjhxK_hNuD.png" /><figcaption><em>Predicted against real pairwise similarity for PharmCast. Peptides are the reserved holdout, screening collection compounds are test cases the model never saw. Figure 6 from the same preprint.</em></figcaption></figure><p>PharmCast runs as a service, with the model and the benchmark, at <a href="https://pharmcast.ai/">pharmcast.ai</a>.</p><p>Malcolm McGregor, now at SRI International, worked on the original pharmacophore fingerprinting with me when we were at Affymax. He is a co-author here too.</p><h3>Turning the screen around</h3><p>Once fingerprints are cheap, something that was impractical becomes routine. The PharmCast work has opened up many new possibilities. For example — traditional forward screening asks which compounds hit a target. Reverse screening asks the opposite: given this molecule, which proteins might it touch. The latter can be very useful to identify off target liability or for repurposing. Done exhaustively by docking, one molecule against the 28,579 ligand bound proteins assembled here takes 40.5 hours on twenty CPU cores. That is why nobody does it casually.</p><p><a href="https://reversescreen.ai/">ReverseScreen.ai</a> puts a retrieval step in front of the docking. Every co-crystallized ligand in the Protein Data Bank is indexed by the three dimensional pharmacophore fingerprint it presents, predicted from two dimensional structure by PharmCast, together with the proteins it was solved against. A query retrieves the proteins belonging to its nearest neighbors, and only those get docked.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Kqwf6KLN-7jJs6sw.png" /><figcaption><em>The whole pipeline on one query. Orforglipron enters as a SMILES string, PharmCast predicts its 10,549 bit pharmacophore fingerprint, the match to the indexed ligand UK1 at Tanimoto 0.90 ranks first of 27,797, and the index carries that ligand back to the GLP-1 receptor. The second arm runs Morgan similarity on the same molecule, where the match is 0.63. Figure 1 from Muskal and Nicola, ReverseScreen.ai application note, 2026.</em></figcaption></figure><p>There are now two arms, and that turned out to be the interesting part — because the editor of the journal we submitted the preprint to suggested we consider this prior to going to peer review. Lo and behold, this added significant value and was very inexpensive to add, only an extra couple of days of work. The first arm indexes by 3D pharmacophores (i.e. PharmCast’s fingerprint), the second by Morgan 2D fingerprint, and a protein is admitted when either similarity clears its own cutoff. The arms find different things. Each holds recorded targets the other misses. Compounds that look unrelated in two dimensions while presenting a similar pharmacophore (i.e. “non-obvious me-toos”) are nearly four times more likely than chance to share a target with the query, and a Morgan cutoff alone never puts them on the list. Across the full sweep of cutoffs, the setting that recovers the most targets for a given list length uses both arms rather than one in 49 of 60 cases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mpiIiCwRqK_o_8vu.png" /><figcaption><em>A candidate drawn at random shares a recorded co-complex target with the query 0.59 percent of the time. The low two dimensional, high pharmacophore quadrant runs 3.9 times above that, and those are precisely the candidates a Morgan cutoff discards. Figure 2 from the same application note.</em></figcaption></figure><p>The evaluation is deliberately hard — something the editor also “suggested.” Before a query runs, its own entry and every indexed ligand sharing its generic scaffold are removed, a mean of 97 of them, so a target can be reached only through a different molecule crystallized against the same protein. For 4.5 percent of queries no such molecule exists at all, which caps recall at 0.96 before the method does anything.</p><p>Within that, the balanced setting leaves a typical query with 35 proteins to dock instead of 28,579, and the recorded target is among them for 71 percent of queries. A looser setting catches 86 percent and hands docking a longer list. Fingerprinting and searching both arms takes 45.7 milliseconds per query, roughly 79,000 molecules per hour on a single core.</p><p>Two 2025 clinical molecules were run as cross validation. <strong>Orforglipron (a once-daily oral, non-peptide, small-molecule GLP-1 receptor agonist developed by Eli Lilly and Company and approved by the FDA in April 2026 for chronic weight management</strong>) returns the glucagon-like peptide-1 receptor first, through a non identical analog, at a pharmacophore similarity of 0.90 against a Morgan similarity of 0.63. <strong>Daraxonrasib (a first-in-class, once-daily oral targeted medication developed by Revolution Medicines to treat advanced cancers driven by RAS gene mutations</strong>) returns the KRAS and cyclophilin A tri-complex fourth, at 0.84 against 0.64. In both cases the pharmacophore arm is what put the right answer near the top.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FkmNDb6MWGCcayYC.png" /><figcaption><em>Entries released by the RCSB Protein Data Bank over time. The addressable target space grows every week without anyone doing anything. Figure 1 from the preprint, doi 10.64898/2026.09.09.750461, CC BY 4.0.</em></figcaption></figure><h3>Two comparators may be all we need</h3><p>This is the one I expect pushback on, and I mean that as a compliment to the idea. I almost always involve multiple co-authors on my papers. I single-authored this one only because I did not want to inflict likely heat on anyone else. That said, <strong>this one I think could reset how the field thinks.</strong></p><p>A compound inside a cell meets a spectrum of proteins from many families at once. Screening interrogates one target at a time. But the questions that actually drive decisions in a rank ordering workflow are comparative: which of two targets does this compound prefer, and which of two compounds does this target prefer. So I built one model for each question. Two chemical structures and a sequence, or two sequences and a structure. No conformational analysis, no protein structure, no binding site, no docked pose.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FviJ3g3nBlOMzWK_.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gbsReQUuOIaCOcgf.png" /><figcaption><em>The two comparisons, each with a real case. Each comparison is one feature row, and the order is the question. Figure 1 from Muskal, Two Comparators May Be All We Need, bioRxiv, doi 10.64898/2026.09.15.751854, CC BY 4.0.</em></figcaption></figure><p>The compound comparator reached 0.71 on 65,725 held-out comparisons, increasing to 0.96 when filtering on prediction strength. The target comparator reached 0.75 accuracy on 8,689 held out comparisons across protein families, and 0.94 when the comparisons are filtered by the model’s own prediction strength. Crossing a family boundary costs about three points against staying inside one. The useful property is that the model tells you when to trust it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qAUhbcoacNReq2yB.png" /><figcaption><em>Accuracy against prediction strength. Prediction strength is the larger of the two returned probabilities. Figure 2 from the same preprint.</em></figcaption></figure><p>The roster is 2,279 human proteins across 34 families for the target comparison and 2,079 across 32 for the compound comparison. Both comparators are live at <a href="https://familyfoundationmodel.com/">familyfoundationmodel.com</a>, alongside the family level models they came from. The same comparative framing is what sits underneath <a href="https://kinasefoundationmodel.com/">kinasefoundationmodel.com</a> and <a href="https://gpcrfoundationmodel.com/">gpcrfoundationmodel.com</a>.</p><h3>Three things they all turned out to share</h3><p>That is the work. I did not plan it as a set, and it came back as one. It took until the fourth of them before I saw what they had in common, and now that they are all on the table, here is what runs through every one.</p><p><strong>Ask which, not how much.</strong> Every one of these projects got better the moment it stopped trying to compute a number and started making a comparison. The 1998 absorption model is judged on whether it puts a compound on the correct side of a threshold, not on how close its percentage lands. The 2003 toxicity method never calculates a toxicity at all. It asks which compounds this one resembles and lets them vote. The reverse screen asks which proteins have already been seen holding something like this. And the last paper presents the <strong>aha realization that much of preclinical discovery is about prioritization and can thus boil down to two pairwise questions asked over and over: which target does a compound prefer, and which compound does a target prefer.</strong> This field has spent decades trying to predict absolute quantities. Almost every decision anyone actually makes with them is a comparison.</p><p><strong>The expensive step is usually optional.</strong> Generating a pharmacophore fingerprint takes 2.86 seconds, and 2.82 of those are spent building conformers that get thrown away, so we removed that stage and predicted the fingerprint directly. Docking one molecule against every crystallized protein site takes forty hours on twenty cores, so we put a retrieval step in front and dock 35 sites instead of 28,579. The revisits themselves needed no new experiment, because the data was already public and one to two orders of magnitude larger than what we had the first time. And the comparator models need no protein structure (just its sequence), no pocket and no pose. In every case the expensive step was buying a precision the decision never used.</p><p><strong>The question outlives the tools.</strong> This is the one I did not expect, and it is the least scientific of the three. None of these old hypotheses turned out to be wrong. They turned out to be underpowered. The six descriptors were right, and there were 86 compounds. The idea that a compound’s neighbors can speak for it was right, and the public data to prove it would not exist for another twenty years. What expired was never the idea. It was the tooling and the sample size, and both of those have now moved by factors that make the original questions answerable almost casually. The same turned out to be true of the people. Every one of these revisits put me back in touch with a co-author I had not worked with in years, and in every case the question we left open was still interesting to both of us. <strong>A career is not really a list of results. It is a list of questions, most of them still open, and most of the people who cared about them are one email away.</strong></p><p>None of those three were in the plan. They are what the work turned out to be about.</p><h3>The part I did not plan for</h3><p>Every one of these revisits put me back in touch with someone I had not worked with in years. Matt Wessel on absorption. Malcolm McGregor on pharmacophores. Sanjiv, Phani and Prashant on toxicity. They have all moved on to other companies and other fields. Thermo Fisher, SRI, AM Green, DigitalOcean, Palo Alto Networks. Everyone said yes.</p><p>There is something specific about reopening a working relationship around a question you both cared about once and never fully settled. It is a different conversation than a cold introduction. The shared context is already there, and so is the unfinished business. If you have old work worth retesting, you probably also have old collaborators worth calling, and the retest is a better reason to call than catching up.</p><h3>A quieter thread: sound, the microbiome, and the voice</h3><p>Alongside the drug discovery work I have been pulling on something further out. Expect more on this as I turn this momentum toward other past projects.</p><p>To us, music is perceived in the brain. But sound is a vibration, a pressure wave, and it has measurable effects on organisms with no ears and no nervous system to hear it with. Music has been shown to shift microbial growth and the production of certain metabolites. Bacteria appear to both emit and respond to sound waves, with at least one study suggesting sound can act as a growth regulatory signal between cells. The mechanisms are not settled. What interests me is whether carefully designed sound could nudge the gut microbiome in a deliberate, therapeutic direction. We carry a pharmacy inside us. Activating it with something as cheap and non invasive as sound would be a genuinely new route to personalized medicine. Just before the pandemic, I reached out to a now collaborator — Prof. Vijay Kothari, Institute of Science, Nirma University. Vijay’s lab and I explored the effect of various sound frequencies on microbiota and published in 2019 “<a href="https://www.eurekaselect.com/article/94241">Sonic Stimulation and Low Power Microwave Radiation Can Modulate Bacterial Virulence Towards Caenorhabditis elegans</a>.” Amazingly, sound corresponding to 400 Hz, and the divine sound ‘Om’ conferred a protective effect on C. elegans in face of bacterial infection. I look forward to picking back up on this trajectory as I think there are many, many new possibilities, and frankly it is one of the main reasons I continue to record music in my studio. It’s all about the content baby!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ryXlk5V4jd6tOKRw.png" /><figcaption><strong>Effect of audible sound in form of music on microbial growth and production of certain important metabolites — </strong><a href="https://link.springer.com/article/10.1134/S0026261715020125"><strong>https://link.springer.com/article/10.1134/S0026261715020125</strong></a></figcaption></figure><h3>Vitamin D, and a question I wrote down ten years ago</h3><p>Ten years ago I wrote down a question and had no practical way to answer it.</p><p>Here is the question. Most Americans are short on vitamin D. And the vitamin D you get, whether from the sun or from a pill, does not work as it arrives. Your body has to switch it on in two steps, first in the liver and then in the kidneys.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DUxrEzf9dp13hbEG.png" /><figcaption>Vitamin D, as either D3 or D2, does not have significant biological activity. It must be metabolized within the body to the hormonally active form known as 1,25-dihydroxycholecalciferol. This transformation occurs in two steps. Within the liver, cholecalciferol is hydroxylated to 25-hydroxycholecalciferol by the enzyme 25-hydroxylase. Within the kidney, 25-hydroxycholecalciferol serves as a substrate for 1-alpha-hydroxylase, yielding 1,25-dihydroxycholecalciferol, the biologically active form. Each of the forms of vitamin D is hydrophobic and is transported in blood bound to carrier proteins. The major carrier is called, appropriately, vitamin D binding protein. Sources: <a href="http://www.vivo.colostate.edu/hbooks/pathphys/endocrine/otherendo/vitamind.html">Colorado State pathophysiology of the endocrine system</a> and <a href="http://www.ncbi.nlm.nih.gov/books/NBK56061/">NCBI Bookshelf NBK56061</a>.</figcaption></figure><p>There are reports that taking probiotics raises the amount of usable vitamin D in the blood. That made me wonder whether some of the bacteria we live with are doing part of that switching on for us, and whether we could encourage them to do more of it.</p><p>Ten years ago that was a thought on a page. Chasing it down would have meant months in the literature, and I never did.</p><p>Last night I checked it in about a third of a second.</p><p>I typed vitamin D into <a href="https://reversescreen.ai/try.html?screen=2026-10-14T234213Z-UVIaGb4g">reversescreen.ai</a>. The tool does one thing. You give it a molecule, and it gives back a list of proteins that molecule might stick to, drawn from the tens of thousands of molecule-inside-protein snapshots that scientists have deposited in a public archive over the past fifty years. Think of it as a reverse image search for drugs. Instead of asking which molecules hit this protein, you ask what this molecule touches.</p><p>Hold the diagram above in your head, because the search reproduces most of it without being told any of it.</p><p>Two answers came back as near certainties, because vitamin D has already been photographed sitting inside them. Both of those proteins belong to bacteria, and one of them is an enzyme whose job is to modify vitamin D.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*27hxOe354H2P3xfo.png" /><figcaption>Vitamin D, drawn in the middle, sitting inside a bacterial enzyme. The green labels are the parts of the protein close enough to touch it. The single blue dotted line is the one firm connection between them.</figcaption></figure><p>Then came the suggestions. Seventy one more proteins, reached not by a direct photograph but by chemical family resemblance. The liver enzyme from step one of the diagram is in there, the human one. So is the carrier protein that moves vitamin D around in the blood. So is the receptor the active form finally acts on, four times over, in humans, rats, zebrafish and lamprey. The search was given a molecule and nothing else, and it handed back most of a textbook page.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R-pRo7oXggwJJQR0.png" /><figcaption>The human liver enzyme from step one, with a close relative of vitamin D inside it. The search finds this one by resemblance rather than by a direct match.</figcaption></figure><p>One thing in the diagram does not come back, and the reason is worth more than the hits are. The second step, the one in the kidney, is carried out by an enzyme called CYP27B1. Nobody has ever determined its structure. There is not a single picture of it in the archive, anywhere, so there is nothing for a search like this to match against. What turns up in its place is a bacterial enzyme that performs both steps at once. A search of this kind can only show you what somebody has already photographed, and the most important human enzyme on that slide has never been photographed at all.</p><p>Which is a fair place to say what the rest of the list is worth.</p><p>A photograph of a molecule inside a protein tells you the molecule was there, under laboratory conditions. It does not tell you the molecule does anything useful once it is there. And a molecule that merely resembles one that was photographed tells you even less than that. Resemblance is a lead. It is not a result.</p><p>Look at the bacterial enzyme image again with that in mind. Vitamin D sits in it making exactly one firm connection, and otherwise just resting against it. Yet when you look up how strongly that same molecule has been measured against various human proteins, the numbers spread over a factor of several thousand. The picture predicts none of that.</p><p>Which is the honest shape of all of this. A search like that takes you from a vague idea to a short list of real candidates in under a second, and it tells you plainly where the archive runs out. Finding out whether any of it is true is still an experiment, and still takes a year.</p><p>Sound and vitamin D look like two separate interests. They are the same one. Both are cases where something cheap and already present, a pressure wave, a public archive, a change in what you eat, might move a system we normally treat as fixed.</p><h3>Voice, and why both of my food apps listen</h3><p>And vitamin D is where that thread runs into its plainest version, because the most reliable lever most people have over it is not a molecule at all. <strong>It is sunlight, and it is food</strong>. Which is a large part of how I ended up building food apps.</p><p>The same instinct, that there is signal in channels we routinely ignore, is what pulls me toward voice as a biomarker. There are real associations between vocal characteristics and metabolic conditions, Type 2 diabetes among them. That is one reason both of my food apps now listen. See “ <a href="https://www.drstevenmuskal.com/p/why-i-started-talking-to-my-computer"><strong>Why I Started Talking to My Computer Instead of Typing at It.</strong></a>”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JZKTajNVvJYwiZFs.jpeg" /><figcaption><em>Two apps, one question. Food Health logs meals from a photograph and tracks nutrition trends. Food Health TxD sits next to continuous glucose data and answers a sharper question: which meals spike you. Both take a spoken note as readily as a picture. </em><a href="https://foodhealthscan.com/"><em>foodhealthscan.com</em></a></figcaption></figure><p>Food Health is the general one. Photograph a meal, get calories, macronutrients and a score, and watch the trend over weeks. Food Health TxD is the metabolic sibling. It pulls glucose from Apple Health, sets it against what you logged, and surfaces the meals that moved you. Scan a menu before you order. Log a reading. Leave a voice note instead of typing.</p><p>The voice capture is there for convenience, and it is also there because the recording itself may carry signal. If vocal features track metabolic state, then every spoken note is a measurement nobody had to ask the user to take. That is a hypothesis, and it is the kind I now know how to test.</p><h3>Why this matters more than any one result</h3><p>None of this is really about a single preprint or application.</p><p>Five pieces of work, four of them resurrections, each rebuilt, revalidated on data that is one to two orders of magnitude larger than the original, published openly with the code and a working website, inside one year. In the era I trained in that would have been a department. Multiple people on data acquisition and curation alone, someone on infrastructure, someone on the writing, and a year of calendar time before anyone outside the building could use any of it.</p><p>That is the shift, and it is not that the science got easier. The tooling around doing science got radically faster. Agentic coding collapses the archaeology. Public data collapses the acquisition. Open preprints and a static website collapse the distribution. What is left is the part that was always the real work: knowing which question is worth asking, and being honest about what the answer says.</p><p>And the three things that tie them together are not really about chemistry. Ask which rather than how much, because comparisons are what decisions are made of. Skip the expensive step, because precision you do not use is not rigor, it is cost. And treat an old question as underpowered rather than closed, because the thing that expired was the tooling, not the idea.</p><p>I have written before that the AI we have now feels like fire to me. <a href="https://www.drstevenmuskal.com/p/the-fire-this-time">The Fire This Time</a>. The more of this revisiting I do, the more I suspect fire undersells it. Fire changed what one person could do with their hands. This changes what one person can do with a career’s worth of unfinished questions, and with the unfinished questions of everyone they ever worked with.</p><p>We are at an inflection point. I think people looking back at this stretch of years will see it clearly, the way inflection points are always obvious in retrospect and easy to miss from inside. I would rather name it while it is happening.</p><p>If you want to see where this goes next, what I am building lives at <a href="https://www.stevenmuskal.com/projects.html">stevenmuskal.com/projects</a>.</p><p><em>Steven Muskal, Ph.D. is Chief Executive Officer of Eidogen-Sertanty, Inc. All five preprints are open access under CC BY 4.0, and the figures above are taken from them.</em></p><p><strong>Where the work lives</strong></p><ul><li><a href="https://pharmcast.ai/">pharmcast.ai</a> pharmacophore fingerprints from two dimensional structure</li><li><a href="https://reversescreen.ai/">reversescreen.ai</a> reverse screening, service and downloadable index</li><li><a href="https://toxpred.ai/">toxpred.ai</a> interactive toxicity scoring, data and compound pool</li><li><a href="https://predicthia.ai/">predicthia.ai</a> human intestinal absorption and permeability</li><li><a href="https://familyfoundationmodel.com/">familyfoundationmodel.com</a> the two comparators and the family models</li><li><a href="https://kinasefoundationmodel.com/">kinasefoundationmodel.com</a> and <a href="https://gpcrfoundationmodel.com/">gpcrfoundationmodel.com</a> the family specific models</li><li><a href="https://foodhealthscan.com/">foodhealthscan.com</a> Food Health and Food Health TxD</li><li><a href="https://eidogen-sertanty.com/">eidogen-sertanty.com</a> the company</li><li><a href="https://www.stevenmuskal.com/projects.html">stevenmuskal.com/projects</a> everything in one place</li></ul><p><strong>The papers</strong></p><ul><li>PharmCast: rapid generation of three dimensional pharmacophore fingerprints from two dimensional structure without conformer generation. Muskal and McGregor. doi <a href="https://doi.org/10.64898/2026.09.02.748999">10.64898/2026.09.02.748999</a></li><li>Toxicity by Consortium Revisited: Structural Neighbors and Cross-Family Reach. Muskal, Jha, Kishore and Tyagi. doi <a href="https://doi.org/10.64898/2026.10.04.756566">10.64898/2026.10.04.756566</a></li><li>Human Intestinal Absorption Revisited: The 1998 Descriptors Survive, and the Method Extends to Permeability. Muskal and Wessel. doi <a href="https://doi.org/10.64898/2026.10.03.756481">10.64898/2026.10.03.756481</a></li><li>ReverseScreen.ai: Pharmacophore-Guided Reverse Screening Across the Growing Co-Complex Proteome. Muskal and Nicola. doi <a href="https://doi.org/10.64898/2026.09.09.750461">10.64898/2026.09.09.750461</a>. Now also an application note submitted to the Journal of Chemical Information and Modeling</li><li>Two Comparators May Be All We Need. Muskal. doi <a href="https://doi.org/10.64898/2026.09.15.751854">10.64898/2026.09.15.751854</a></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8kj95fbjuvm4m2NP.png" /></figure><p>Steven Muskal, Ph.D. is the CEO of <a href="https://www.eidogen-sertanty.com/">Eidogen-Sertanty, Inc.</a> — a drug discovery informatics company. He has spent four decades working at the intersection of computational biology, AI, and drug discovery. He writes about AI, health, and the intersection of biology and technology at <a href="http://stevenmuskal.com/">stevenmuskal.com</a></p><p>For a couple mix videos, the first was from a recent ridiculously productive session with Maya, David, Tim, Dom, and Don. Such a good example of Bass and vocals laying down the track for the rest of us to follow. The second is from a party mix session earlier this year with Rick, Grant, Dom, and Alan. My first opportunity after all these years to hit real drums (v. air drums) on this Phil Collins classic. Booyah!</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FkizQxC5gZBA%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DkizQxC5gZBA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FkizQxC5gZBA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ccd150b49f19989895fe33e476f5f267/href">https://medium.com/media/ccd150b49f19989895fe33e476f5f267/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F-wHc3O0UHXc%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D-wHc3O0UHXc&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F-wHc3O0UHXc%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/c2e3808a28ee32f8cc6fd3fe015cb746/href">https://medium.com/media/c2e3808a28ee32f8cc6fd3fe015cb746/href</a></iframe><p><em>Originally published at </em><a href="https://www.drstevenmuskal.com/p/revisiting-old-work"><em>https://www.drstevenmuskal.com</em></a><em> on October 15, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8241cf3ba308" width="1" height="1" alt=""><hr><p><a href="https://pub.towardsai.net/revisiting-old-work-8241cf3ba308">Revisiting Old Work</a> was originally published in <a href="https://pub.towardsai.net/">Towards AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>