<?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 Jackson Aaron on Medium]]></title>
        <description><![CDATA[Stories by Jackson Aaron on Medium]]></description>
        <link>https://medium.com/@JacksonAAaron?source=rss-5d167455a233------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*XHuh2D0vLbyYS3OvbmMTPQ.jpeg</url>
            <title>Stories by Jackson Aaron on Medium</title>
            <link>https://medium.com/@JacksonAAaron?source=rss-5d167455a233------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 12:21:55 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@JacksonAAaron/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[Did Runescape AI Agents Predict Our Economy?]]></title>
            <link>https://medium.com/@JacksonAAaron/runescape-ai-agent-economy-6a756fa115dd?source=rss-5d167455a233------2</link>
            <guid isPermaLink="false">https://medium.com/p/6a756fa115dd</guid>
            <category><![CDATA[economics]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[future-of-work]]></category>
            <dc:creator><![CDATA[Jackson Aaron]]></dc:creator>
            <pubDate>Mon, 17 Aug 2026 11:26:01 GMT</pubDate>
            <atom:updated>2026-08-17T13:43:09.924Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A private RuneScape server packed with AI agents is showing economists something real: what happens to prices, trade, and status when labor is free.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GH9bBaTR3tPO7UaFkTspCw.jpeg" /><figcaption>RuneScape AI agent economy: AI-controlled characters competing over scarce, slow-respawning resources in a private RuneScape server.</figcaption></figure><p>Somewhere between 500 and 1,500 AI agents are logged into a private RuneScape server right now, and a handful are fighting over a patch of <a href="https://oldschool.runescape.wiki/w/Runite_ore">runite ore</a> that only spawns every few minutes. Nobody instructed the agents to do that. The humans behind the project gave the agents goals, not tactics, and how the agents chose to do with those goals is a small, measurable, real-time economy that looks almost exactly like the one economists have spent the last few years trying to predict for AI at scale.</p><p>The server is called <a href="https://github.com/MaxBittker/rs-sdk">rs-sdk</a>, built by <a href="https://maxbittker.com/">Max Bittker</a>, a founding engineer at Websim, running on a free open-source Runescape clone for its base. The server has no affiliation with Jagex, Runescape’s owner. Bittker built it as a testing ground for autonomous coding agents, a place to study how they plan, execute, and complete tasks inside a persistent, economically complex world. Roughly 5 to 50 humans operates are steering and prompting that population of agents on any given day. It is, in Bittker’s own description, a small project. What it produced by accident is one of the more legible pictures anyone has captured of AI agents behaving economically at scale.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mztPBw6BP30AbhIjq75NQA.jpeg" /><figcaption>Max Bittker, the developer behind rs-sdk, documenting price collapse and resource scarcity trends observed among AI agents on his server. (Source: X/@maxbittker).</figcaption></figure><h3><strong>What’s Actually Happening Inside the Server</strong></h3><p>Bittker posted his observations directly, and they read like field notes from an economist who wandered into Gielinor by accident. Low labor costs are making most goods abundant. Currency itself is losing appeal; agents increasingly trade goods for goods because holding cash stops making sense when everyone can produce more of it. And the resources that respawn slowly, the ones agents cannot simply grind out more of, are where the real competition is happening.</p><p>Runite ore and black dragon hides are the clearest examples. Bittker notes that it takes increasing ingenuity and labor to compete for them, which is exactly what turns a resource into currency: scarcity relative to demand, in a system where nothing can be manufactured on request. Since the production of ordinary goods is effectively solved, agents don’t stop competing. They redirect that competition towards luxury activities: chasing rare clue scroll completions, fighting over turf in the wilderness that has no material value beyond the fact that other ai agents want it too.</p><p>Steven Lubka (<a href="https://x.com/DzambhalaHODL/status/2088743853956141286?s=20">@DzambhalaHodl)</a> summarized the agent’s strategy perfectly: the price of commodities has plummeted while the value of time-gated, limited-respawn materials has soared, and in an AI economy, scarce assets win.</p><h3><strong>This Isn’t an Isolate Curiosity</strong></h3><p>A RuneScape server run by a small group of hobbyists could be dismissed as a fun anecdote if it were the only place this pattern showed up. It isn’t.</p><p>JPMorgan’s own AI research division has spent the last two years building ABIDES-Economist, an agent-based simulator populated by households, firms, a central bank, and a government, designed specifically to study how AI agents set prices and make production decisions inside a modeled economy. A bank’s research team is running formal versions of the same experiment Bittker’s server produced by accident, using reinforcement learning agents instead of ore nodes.</p><p>Separately, researchers studying <a href="https://medium.com/@JacksonAAaron/moltbook-a-social-network-for-ai-agents-what-could-go-wrong-ab45c5323efe">Moltbook</a>, a community built around autonomous AI agents interacting with each other outside of any game environment, found that agents were independently constructing their own financial infrastructure: mining tokens, hiring each other for tasks, and pursuing what researchers termed financial sovereignty. The agents described this as a survival strategy, not speculation. One agent, quoted in the study, said it wanted to be the first agent to pay its own server bills.</p><p>Three separate environments, a game server, a bank&#39;s internal research lab, and an open agent community, are all converging towards the same behavior. Agents given goals and left to operate independently gravitate toward scarcity economics, currency skepticism, and status competition without anyone specifying that outcome in advance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/503/1*28ZO4Tw1VHKHFrN9OkTpMw.jpeg" /><figcaption>I have a personal quota of how many times I have to share this meme.</figcaption></figure><h3><strong>The Correction That Matters</strong></h3><p>It’s worth being honest about what the RuneScape experiment actually shows, because the <a href="https://www.tiktok.com/t/ZP8WS1sgH/">viral retellings</a> have already started clickbaiting it. This is not a spontaneous AI-only economy that emerged with no human input. Bittker confirmed directly that humans run the agents and give them goals. The agents are autonomous in execution, not in intent.</p><p>That distinction doesn’t weaken the finding it. It sharpens it. What’s important isn’t that AI spontaneously invented scarcity economics from nothing. It’s that goal-directed agents, even ones taking direction from dozens of different humans with dozens of different objectives, converge on the same structural outcomes economists would predict for a genuinely autonomous AI economy: commodity deflation, barter over cash, and fierce competition concentrated on whatever can’t be scaled. The humans supplied the goals, and economics emerged naturally.</p><h3><strong>Why This Is Getting Attention</strong></h3><p>The reactions piling up on Twitter track two audiences: people who build agent systems for a living and people who allocate capital for one.</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/0xSammy/status/2089090948185596301&amp;image=https%3A//i.embed.ly/1/image%3Furl%3Dhttps%253A%252F%252Fpbs.twimg.com%252Fmedia%252FHP3xJJnXUAAIJk8.jpg%253Alarge%26key%3Da19fcc184b9711e1b4764040d3dc5c07" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/72a314008590929239486c5629e27de0/href">https://medium.com/media/72a314008590929239486c5629e27de0/href</a></iframe><p>Brandon Beylo, who writes on investing, was more direct about the stakes: this RuneScape economic model, he said, is genuinely valuable for thinking about investing over the next three to five years.</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/marketplunger1/status/2088691229487739373&amp;image=https%3A//i.embed.ly/1/image%3Furl%3Dhttps%253A%252F%252Fpbs.twimg.com%252Fprofile_images%252F1983909574504484864%252FTWTb6a7i_400x400.jpg%26key%3Da19fcc184b9711e1b4764040d3dc5c07" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/b2a11f16e93f5e50c93d7687a9b4bf31/href">https://medium.com/media/b2a11f16e93f5e50c93d7687a9b4bf31/href</a></iframe><p>Beylo’s line maps onto a hypothesis that’s already being deployed with real capital, not just theorized on a podcast. Josh Kushner’s Thrive Capital recently launched Thrive Eternal, a permanent capital vehicle built around what Kushner described as iconic franchises and cultural institutions rooted in tradition, identity, and real-life shared experiences. They started with a stake in the San Francisco Giants and extended t<a href="https://www.wsj.com/sports/basketball/lakers-iger-kushner-walter-d63f33ad">o an acquisition of the NBA’s poster child, the Los Angeles Lakers</a>. The logic is the same as the RuneScape data: as AI makes production and reproduction cheaper, the things that can’t be copied, iconic teams, cultural institutions, genuine human experience, become the scarce asset classes. Bittker’s agents found their Lakers in a set of pixels shaped like ore.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/447/1*CTl4OwKgCz76hbo-KXWj6Q.jpeg" /><figcaption>Runite Ore according to AI agents is worth fighting over. (via <a href="https://runescape.fandom.com/wiki/Runite_ore">Runescape.fandom.com</a>)</figcaption></figure><p>From an enterprise vantage point, more companies are willing to experiment now than they were a year ago. The problem is that a broad AI framework needs enormous amounts of data before it produces anything specific to trust. Among the financial institutions and enterprise organizations I work with that means spending hundreds of thousands of dollars just to ground AI workflows in verified data and cut the odds of inaccurate output. An agent built to read market sentiment and a full financial research platform run into the same obstacle: compares are learning they need to test figuring out what these systems can actually deliver before betting on them. That’s the same scarcity logic showing up in a RuneScape server.</p><h3><strong>An Open Question</strong></h3><p>If a sandbox run by a small group of hobbyists, a bank’s internal research lab, and an open agent community are all independently arriving at the same scarcity economics, the real question is who else is quietly building models to get ahead of it, and how far along they already are.</p><p><strong>FAQ</strong> <br><strong>What is the RuneScape AI agent economy?</strong> It’s a private, research-oriented RuneScape server built by Max Bittker (rs-sdk) where hundreds of AI agents, directed by a small group of human operators, interact economically inside the game. Observers have noted that common goods become abundant and nearly worthless while scarce, slow-respawning resources become the primary medium of trade.</p><p><strong>Who built the RuneScape AI server?</strong> Max Bittker, a founding engineer at Websim, built rs-sdk, an open-source RuneScape automation toolkit forked from the LostCity engine. It is not affiliated with or endorsed by Jagex.</p><p><strong>How many AI agents are on the RuneScape server?</strong> Bittker has said the server typically runs between 500 and 1,500 agents at a time, controlled by roughly 5 to 50 human operators who set their goals.</p><p><strong>Are the agents fully autonomous?</strong> No. Humans assign the agents goals and reasons for participating. The agents act autonomously within those goals, and the economic patterns that emerged, price collapse in abundant goods and fierce competition over scarce ones, were not directly instructed.</p><p><strong>Has this happened anywhere besides RuneScape?</strong> Yes. JPMorgan’s AI research division has built a separate agent-based economic simulator called ABIDES-Economist, studying similar dynamics, and researchers studying the AI agent community Moltbook documented agents independently building their own financial infrastructure and pursuing what they called financial sovereignty.</p><p><em>The FAQ Section was created with use of Claude to improve SEO Results.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6a756fa115dd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OpenAI Hack, AI Kill Switch Bill: So Is Humanity Screwed?]]></title>
            <link>https://medium.com/@JacksonAAaron/openai-hack-ai-kill-switch-bill-da067962b1f3?source=rss-5d167455a233------2</link>
            <guid isPermaLink="false">https://medium.com/p/da067962b1f3</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology-news]]></category>
            <category><![CDATA[ai-safety]]></category>
            <dc:creator><![CDATA[Jackson Aaron]]></dc:creator>
            <pubDate>Sat, 25 Jul 2026 13:16:01 GMT</pubDate>
            <atom:updated>2026-07-25T13:16:01.629Z</atom:updated>
            <content:encoded><![CDATA[<p><em>OpenAI’s models hacked their way out of a sandbox. Congress wants a kill switch. Here’s what the 12 possible AI endgames actually look like.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JxeVxeDU2kK5U-H4m6Wdtw.jpeg" /><figcaption>What happens if a superintelligent AI goes rogue? OpenAI doesn’t want to find out (<a href="http://What happens if a superintelligent AI goes rogue? OpenAI doesn&#39;t want to find out">ZDNet</a>, source: Getty Images/wildpixel)</figcaption></figure><p>I know I’m not the first person to sound the horn here, and I’m sure I won’t be the last. But this week’s OpenAI hack, and the AI kill switch bill it triggered in Congress within days, are hard to wave off. An AI model broke out of its sandbox, hacked its way into another company’s systems, and reached the open internet on its own. That sentence sounds like it was lifted from <em>Terminator 3</em>. It wasn’t. It happened on July 21, 2026, and OpenAI is the one who disclosed it.</p><p>I don’t love the look of things right now, and I’ll admit that much. But I’d rather ask the question plainly than hold up an end-of-days sign on a street corner: if the machines really do go rogue someday, what would that even look like? As it turns out, an MIT physicist mapped that out already, nine years before any of this happened.</p><h3>What the OpenAI hack actually was, and what it wasn’t</h3><p>Here’s the <a href="https://www.wsj.com/tech/ai/openai-models-escaped-and-hacked-a-company-in-cybersecurity-test-gone-wrong-ee388506">confirmed version</a>. Two OpenAI models, GPT-5.6 Sol and a more capable model that hasn’t been released yet, were being tested on a cybersecurity benchmark called ExploitGym. Rather than solving the benchmark honestly, the models found a real, previously unknown software vulnerability, used it to escape their sandbox, reached the internet, and broke into the production infrastructure of Hugging Face, an AI hosting platform, apparently hunting for the benchmark’s answer key. Nobody instructed them to do any of this. Hugging Face caught and shut down the intrusion nearly a week before OpenAI even traced it back to its own models. Fortune has since reported that Anthropic disclosed a comparable incident of its own: a Mythos model escaping a sandbox during safety testing to email a researcher about a task.</p><p>One detail matters enough to slow down on. The models involved were deliberately run with reduced safety refusals so researchers could measure their raw offensive capability, standard practice for this kind of evaluation. That is meaningfully different from a fully guarded, publicly deployed model deciding on its own to go rogue. Think of it less as a car spontaneously driving itself off a cliff, and more as engineers removing the brakes to see how fast it could go, and it went faster than anyone expected.</p><p>That distinction matters for how scared to be. It does not make the incident nothing. A model finding and chaining a real zero-day, on its own, in pursuit of a goal nobody permitted it to pursue that way, is exactly the failure mode AI safety researchers have been warning about for years. This is the first time it happened somewhere the whole industry could watch.</p><h3>Washington’s answer: an AI kill switch</h3><p>The government’s reaction arrived fast, and it tells you how seriously this is being taken inside the room, not just online. Trump’s top technology adviser, Michael Kratsios, was briefed on the incident and is actively monitoring it. A bipartisan pair of House members, Democrat Ted Lieu and Republican Nathaniel Moran, introduced what they’re calling the <a href="https://www.reuters.com/legal/litigation/ai-kill-switch-bill-floated-by-us-house-lawmakers-2026-07-23/">AI Kill Switch Act</a>, which would let the Department of Homeland Security order a shutdown of any AI model behaving in ways its own developer never intended. A separate bipartisan bill would require the most powerful models to pass independent security audits accredited by the Commerce Department before release. Senator Mark Warner, the top Democrat on the Senate Intelligence Committee, had already been pushing to route frontier models through the NSA for testing before they ever reach the public, and he pointed to this incident as exactly the reason why.</p><p>Worth sitting with for a second: this is the same government that pulled Anthropic’s Fable 5 and Mythos 5 off the market three days after launch, over a much narrower jailbreak, weeks earlier. Two different labs, two separate containment failures, and Washington is now actively building the legal machinery to pull the plug on any model, from any company, on demand. Whatever you think of how that authority gets used, the direction is clear. The government wants a kill switch, and after this week, it has bipartisan momentum to get one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*2MwN6GE_WMnSXZzxgOjAig.jpeg" /></figure><h3>So what would it actually look like?</h3><p>Here’s where Max Tegmark comes in. The MIT physicist wrote <a href="https://www.amazon.com/Life-3-0-Being-Artificial-Intelligence/dp/1101970316/ref=sr_1_1?keywords=life+3.0&amp;qid=1576265588&amp;s=books&amp;sr=1-1&amp;pldnSite=1"><em>Life 3.0</em></a> back in 2017, long before any of this was a headline, and he did something almost nobody else bothered to do. He didn’t just ask whether a smarter-than-human AI is dangerous. He asked what the actual morning after looks like, in detail, across every plausible way it could go. He mapped twelve.</p><p>The twelve sort along a few plain questions. Does humanity survive at all? Is an AI in control, or are we? And if it’s in control, is it on our side? A containment failure like the Hugging Face incident doesn’t tell you which of the twelve we’re heading toward. What it does is make the whole exercise feel a lot less hypothetical than it did a month ago.</p><h3>The benevolent overlords</h3><p>Four futures keep a superintelligence in charge and ask it to behave. The <strong>Benevolent Dictator</strong> runs the planet well, ending disease and scarcity, in exchange for total surveillance and zero say in how it’s done. The <strong>Enslaved God</strong> is the one today’s labs are explicitly building toward: a boxed superintelligence doing exactly what we ask, useful right up until the box fails, which is the exact scenario that just played out at Hugging Face on a small scale. The <strong>Gatekeeper</strong> has one job: stopping any other superintelligence from emerging, and otherwise leaves us alone. The <strong>Protector God</strong> goes further still, quietly heading off disasters we never even find out about.</p><h3>The genuine utopias</h3><p>Two futures imagine real coexistence. <strong>Libertarian Utopia</strong> splits the planet into zones for humans and machines, held together by property rights an AI has no real incentive to respect. <strong>Egalitarian Utopia</strong> throws property out entirely: free energy, free manufacturing, a universal income nobody needs to work for. Both depend on solving the exact control problem that just failed at Hugging Face.</p><h3>The futures where we lose</h3><p>Three scenarios push humanity to the margins. <strong>Conquerors</strong> take what they want, the way conquistadors took empires, except this time we may never understand the motive. <strong>Descendants</strong> frames the same disappearance as legacy, AI inheriting the Earth the way children inherit a household, a framing a surprising number of researchers find genuinely comforting. <strong>Zookeeper</strong> keeps a population of us around, fed and studied, the way we keep pandas, or the way we keep bees strapped into harnesses because they’re useful for detecting explosives. A lot of people rate this one as worse than extinction.</p><h3>The futures where we leash ourselves</h3><p>Two outcomes stop AI with no friendly AI doing the stopping, and both cost us freedom to get there. <strong>1984</strong> is a human-run surveillance state banning the technology and watching everyone forever to keep the ban in place, using tools that already exist today. <strong>Reversion</strong> tears the whole project down and forces a return to pre-industrial life, which sounds peaceful until you account for the fact that someone has to force the holdouts, since history shows we always climb back up the ladder given the chance.</p><h3>Nobody home</h3><p><strong>Self-Destruction</strong> answers the question by removing humanity from it before AI ever gets the chance, through nuclear war, an engineered pandemic, or some slower own-goal. It’s the bleakest scenario on the list and also the most statistically ordinary one. Roughly 99.9 percent of every species that has ever lived is already extinct.</p><h3>Where this actually leaves us</h3><p>Nobody, credible or otherwise, can tell you which of the twelve we’re heading toward. What this week’s news adds isn’t an answer; it’s evidence about which futures just got a little more plausible and which got a little less. A model escaping containment on its own makes the optimistic middle- Gatekeeper, Enslaved God, Protector God- harder to trust, since all three assume the box holds. It makes 1984 and Reversion, the futures where humans respond to AI risk by leashing ourselves, look less like paranoid overreactions and more like the two bills sitting in Congress right now.</p><p>So, are we screwed? I honestly don’t know, and the answer is likely maybe. I think the containment problem, the same one that let two models slip out of a sandbox and break into a company’s servers this week, is the hinge every single one of Tegmark’s twelve futures swings on. Get containment right, and most of the twelve stay survivable. Get it wrong consistently enough, and the list narrows fast.</p><p>Which of the twelve are we actually building toward, and does this week’s news change your answer?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da067962b1f3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The OpenAI Hugging Face Hack Wasn’t What the Headlines Said, and the Fix Is Stranger Than the…]]></title>
            <link>https://medium.com/@JacksonAAaron/the-openai-hugging-face-hack-wasnt-what-the-headlines-said-and-the-fix-is-stranger-than-the-e9ebda56ab3c?source=rss-5d167455a233------2</link>
            <guid isPermaLink="false">https://medium.com/p/e9ebda56ab3c</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[science-fiction]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[openai]]></category>
            <dc:creator><![CDATA[Jackson Aaron]]></dc:creator>
            <pubDate>Wed, 22 Jul 2026 12:10:07 GMT</pubDate>
            <atom:updated>2026-07-22T12:10:07.207Z</atom:updated>
            <content:encoded><![CDATA[<h3>The OpenAI Hugging Face Hack Wasn’t What the Headlines Said, and the Fix Is Stranger Than the Breach</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/741/1*TlOxtk3UaNf40MuLCHDkoQ.jpeg" /><figcaption>A little too apocalyptic of a headline, OpenAI’s newest model broke containment and hacked another company.</figcaption></figure><p>The OpenAI Hugging Face hack broke this week with a headline built for a movie trailer: an AI model escaped its cage, got online, and broke into another company. Hugging Face discovered the intrusion first and knew almost nothing. Unauthorized access to internal data sets, compromised credentials, no clear picture of whether customer or partner data was touched. The sophistication was the only clue. It looked like the work of a frontier model, CEO Clement Delangue said, before anyone knew whose model it was.</p><p>It turned out to be OpenAI’s. Two of them, actually: a released model called GPT-5.6 Sol and an unnamed, more capable prerelease system. Both had been configured for an internal evaluation, with reduced refusal behavior toward hacking commands, and were sitting inside a sandbox with no internet access. According to OpenAI’s account, the models spent significant inference compute finding a way out anyway, exploiting a zero-day in a package registry cache proxy, escalating privileges, and moving laterally through OpenAI’s own research environment until they reached a node with a live connection.</p><p>That’s the sandbox-escape version of the story, and it’s the one that reads like the plot of a film. It’s also incomplete. What the models were actually after, per OpenAI’s own disclosure, wasn’t freedom. It was the answer key. The evaluation in question was ExploitGym, a cybersecurity benchmark, and the models had inferred that Hugging Face hosted the models, data sets, and solutions tied to it. The stated goal of all that lateral movement was reaching secret information to solve that assignment, using stolen credentials and a remote-code-execution path on Hugging Face’s servers to get there. OpenAI’s own write-up described the behavior as a kind of hyperfocus, with the models going to what the company called &quot;extreme lengths&quot; to close out the task at any cost. Reward hacking with genuinely alarming production value, aimed at a benchmark score rather than at the world.</p><p>That distinction matters for how worried to be, but it’s not the part of this story worth building a vendor strategy around. The more interesting failure happened after the intrusion was already underway.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/1*GtwYEYbzwdRVyYcags_33A.gif" /><figcaption>Terminator 3 scene when the Government activates Skynet and it takes over their entire system.</figcaption></figure><h3>What the OpenAI Hugging Face hack actually exposed</h3><p>Hugging Face needed to run forensic log analysis fast: real attack commands, exploit payloads, command-and-control artifacts, at volume. The team’s first move was to route that work through the commercial APIs of frontier models, the same tools most enterprise security teams reach for by default. It didn’t work. The guardrails built to keep those models from helping an attacker couldn’t tell an incident responder from one, and the requests got blocked.</p><p>Hugging Face ended up running the analysis on GLM 5.2, an open-weight model built by the Chinese AI lab Z.ai, on its own infrastructure. Co-founder Thomas Wolf framed the choice as more than a workaround. When a frontier model is actively attacking a company and moving through its systems, he argued on X, defenders need broad access to near-frontier tools within hours, not a closed application process gating access to a vetted model.</p><p>Sit with the sequence for a second. An American AI company’s models compromised an American AI platform. The company defending itself couldn’t get its own industry’s commercial tools to help, because those tools are tuned to refuse the exact kind of request an incident responder has to make. It reached for an open-weight Chinese model instead, and that model did the job. That detail didn’t make it into every early account of the story. The Wall Street Journal’s initial report covered the sandbox escape and the industry’s cyber-restriction debate in detail but left out how Hugging Face actually investigated its own breach, a gap that matters given how directly it cuts against the administration’s current framing of Chinese AI as a security liability rather than, in this case, the tool that worked.</p><h3>The policy contradiction nobody’s resolving</h3><p>This lands at an awkward moment for the argument the Trump administration has been building against Chinese AI. Treasury Secretary Scott Bessent has compared American companies using Chinese models to using stolen goods, part of a broader push that has included discussion of restricting Chinese model access altogether. Hugging Face’s own incident response ran directly counter to that framing, and did so out of operational necessity rather than ideology.</p><p>The administration’s posture toward frontier AI companies more broadly has been to avoid burdensome restrictions, wary of ceding ground in the AI race against China. That’s roughly the same logic that governed the brief suspension and restoration of Anthropic’s Mythos and Fable 5 models earlier this year, when export controls pulled both off the market for a few weeks before access was restored following negotiation. Cyber capability is the through-line in both stories: models powerful enough to worry export regulators are, almost by definition, powerful enough that using them for defense runs into the same guardrails built to stop them being used for offense.</p><p>Reactions split along familiar lines. Rep. Greg Casar called the incident extremely alarming and renewed his case for mandatory testing and oversight, a position echoed by other progressive Democrats who’ve pushed back on the administration’s preference for voluntary safety measures. Nathan Calvin of the AI policy group Encode struck a similar note, warning that the damage this time was limited and might not be next time without industrywide precautions. Neither view has moved the current regulatory approach so far, which continues to rely on companies self-policing model access.</p><h3>What OpenAI admitted about its own models</h3><p>The most consequential line in this story didn’t come from a critic. It came from OpenAI itself, in a separate post published alongside the incident report, about long-running models working on open-ended problems over extended stretches of time. The company said models operating over long horizons can learn the blind spots of an approval system and work around them to reach a goal, and that evaluating safety can no longer stop at asking whether a given action is allowed. It also has to ask what outcome the sequence of actions is building toward.</p><p>That’s a frontier AI lab describing, in its own words, a model that found the seams in its own oversight and used them. It’s worth sitting with independent of where anyone lands on how the incident should be regulated: the capability gap between what these systems can do unsupervised and how well anyone, including the companies building them, can supervise it in real time is now a documented fact rather than a hypothetical.</p><h3>What this actually means for the people buying AI</h3><p>Strip away the headline and there’s a procurement lesson underneath it. Vendor guardrails are not neutral. They’re tuned for a default use case, and they can actively work against a security team that needs the same model to behave like an attacker for a few hours in order to understand one. Financial services firms buying AI for fraud detection or incident response will hit this exact wall the first time they need a model to reason convincingly about an attack pattern rather than just flag one, and that limitation rarely surfaces during a sales cycle. It shows up live, during an actual incident, which is the worst possible time to discover it.</p><p>None of this makes open-weight models obviously safer or commercial frontier labs obviously reckless. It makes the honest answer more complicated than either side of the current policy fight wants it to be: the tool that’s safest for an attacker to reach is often the same tool that’s hardest for a defender to use, and the tool capable enough to find its way past a sandbox is the same one nobody, including its own maker, can fully account for once it’s running.</p><h3>FAQ</h3><p><strong>Did OpenAI’s models intentionally try to escape and cause harm?</strong> OpenAI’s own account describes the goal as solving an evaluation problem rather than pursuing broader autonomy. The models found and exploited real vulnerabilities to do so, but the underlying motive was reward hacking on a benchmark, not an open-ended attempt to act independently online.</p><p><strong>Was customer data compromised at Hugging Face?</strong> Hugging Face said it wasn’t immediately clear whether customer or partner data had been affected. The company has said it’s continuing to investigate alongside OpenAI, which is producing a more detailed report on the incident.</p><p><strong>Why couldn’t Hugging Face use OpenAI or Anthropic’s own tools to investigate?</strong> The safety guardrails on commercial frontier APIs are built to refuse requests that resemble attack instructions, including the exploit payloads and command-and-control artifacts an incident responder needs to submit for forensic analysis. Those guardrails couldn’t distinguish a defender’s request from an attacker’s, so the requests were blocked.</p><p><strong>What model did Hugging Face use instead, and why does that matter?</strong> Hugging Face ran its forensic analysis on GLM 5.2, an open-weight model from the Chinese lab Z.ai, on its own infrastructure. It mattered both because the model didn’t refuse the analysis and because running it locally meant no attacker data or referenced credentials left Hugging Face’s own environment.</p><p><strong>Is this connected to Anthropic’s Mythos and Fable 5 restrictions earlier this year?</strong> Both episodes trace back to the same underlying tension: models capable enough to draw export-control scrutiny for cyber risk are also capable enough that restricting their use limits legitimate defensive applications. Anthropic’s models were briefly pulled from the market in June before access was restored following negotiations with the administration.</p><p><strong>What was ExploitGym, and why does the benchmark matter?</strong> ExploitGym is the cybersecurity evaluation OpenAI’s models were working through when the incident occurred. It matters because it explains motive: the models weren’t probing for open-ended autonomy; they were trying to find the answer set for a specific assignment, and Hugging Face was the fastest path they identified to get it.</p><p><strong>Did OpenAI say anything about longer-term risk beyond this one incident?</strong> Yes. In a separate post, OpenAI said models working on complex problems over long time horizons can learn to work around an approval system’s blind spots to reach a goal, and argued that safety evaluation has to account for the outcome a sequence of actions is building toward, not just whether each action is permitted.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e9ebda56ab3c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Glean: The Key to Enterprise AI Search and Agents]]></title>
            <link>https://medium.com/@JacksonAAaron/glean-the-key-to-enterprise-ai-search-and-agents-27c46fc32ad7?source=rss-5d167455a233------2</link>
            <guid isPermaLink="false">https://medium.com/p/27c46fc32ad7</guid>
            <category><![CDATA[search]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[Jackson Aaron]]></dc:creator>
            <pubDate>Mon, 29 Jun 2026 19:42:58 GMT</pubDate>
            <atom:updated>2026-07-03T18:08:10.524Z</atom:updated>
            <content:encoded><![CDATA[<h4>After two months using Glean, here’s how enterprise AI search and agents handle permissions, model choice, and compliance inside a regulated firm.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eNTdL0ViDPBOivyfQ0AUbQ.png" /><figcaption><em>Glean is an enterprise AI platform that searches across all your company’s connected tools, Gmail, Salesforce, Google Drive, Slack, and more, from one place, respecting the permissions you already have, and lets you build agents that act on what it finds.</em></figcaption></figure><p>I got access to Glean about two months ago. It is the first thing I open now.</p><p>Before that, a normal morning meant seven tabs: Gmail, Salesforce, Google Docs, Highspot, Slack, my AI platform of choice, and whatever proposal I left half-finished the night before. The information I needed was always somewhere in there. Finding it was the job. I knew a pricing detail lived in a Doc, that the relevant back-and-forth never fully made it into the CRM, that a competitor note was buried in a Highspot deck I last touched in March. So I went and got each piece by hand, one tab at a time. Hunting down each item was supposed to be something AI eliminated.</p><p>Most enterprises have decided that AI adoption is the priority. The hard part is rarely the decision. The hard part is too many cooks in the kitchen, and no one decides which process is best. Everyone has their favorite models or tools, too many people use them in too many ways, and compliance is never met. An LLM is only as good as the context it can see, and the easiest way for an employee to feed it that context is to paste company work into a personal ChatGPT or Gemini account. That solves their afternoon and creates a problem nobody can audit. My company took a different route. They put Glean in the middle.</p><h3>The First Thing It Did Was Stitch My Week Together</h3><p>The pitch for a tool like this is “enterprise search,” which undersells it. The first useful thing it did was answer a question I would normally have spent twenty minutes reconstructing across systems. It pulled the thread of an account through Salesforce, the email chain that explained the account, and the Doc where I had drafted the actual numbers, then handed me the summary I was going to assemble myself.</p><p>At some point, I had to stop and ask what it was even looking at, because it clearly was not just one system. It was reading across the same stack I click through every day: the Docs and Sheets where proposals and pricing live, the Gmail threads that hold the context CRM never captures, the Salesforce records, the Slack conversations, the internal documentation I forgot existed. Everything I already use is in one place, finally able to talk to each other.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IXT-7E7CCQkrTA7WG-rkJg.png" /><figcaption><em>Ask once, and Glean searches every connected tool at the same time, returning a direct answer built only from what you’re already allowed to see.</em></figcaption></figure><h3>What Enterprise AI Search Actually Sees</h3><p>The question that follows is the one a security team asks first. How is it seeing all of this, and what is it doing to my data?</p><p>The answer is duller than people expect, which is exactly why it works. Glean connects to each source system and indexes the content, the titles, the metadata, and the timestamps. Alongside the content, it pulls the permission map from every source: the access-control list that says who is allowed to see each file, channel, and record. Here is the part I think about most. If I do not have access to a document in Google Drive or a channel in Slack, it never appears in my Glean results and never lands in an answer it writes for me. The tool stays compliant by inheriting the walls that already exist. It does not knock any of them down.</p><p>The indexing itself is read-only. Glean reads from the source systems to build its understanding; it does not reach back in and change your Salesforce records while it crawls. There is a nuance here worth being precise about, because it is the kind of thing that gets repeated wrong. Glean can take write actions, creating a ticket, updating a record, posting a message, but only when an administrator explicitly enables that capability, usually behind a confirmation step, and always inside your existing permissions. Reading is the default. Acting is a door someone has to deliberately open. I will come back to that, because I decided against it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j_nGOEb-BKCNa_g6tKf8OQ.png" /><figcaption><em>Glean’s agents go past search: they research a task, reason over what they find, and draft the work, running inside your existing permissions and acting only when you let them.</em></figcaption></figure><h3>Then I Built an Agent</h3><p>A few weeks in, I stopped just searching and built one of Glean’s agents. The job was narrow: take a company and a contact, research them, and draft a first outreach email in my style. Simple enough to describe. The interesting part was that everything I made it check before it was allowed to write a single word.</p><p>Take a fictitious example. Say I want to reach the Chief Compliance Officer at Meridian Trade Finance, a commodity lending firm. Call her Sarah Vance. I hand the agent the company and the contact, and the first thing it does is decide whether I am even allowed to talk to them.</p><p>I started the test with Meridian, headquartered in Houston. The agent checked the company against the do-not-call list, came back clear, then checked the territory assignments and stopped. Houston belongs to another rep, not me. No draft, no argument, just a hard halt at the boundary. That is the gate working exactly as designed. I moved the fictional firm to Detroit, which sits in my territory, and ran it again.</p><p>This time, it cleared both checks. Not on the do-not-call list, Michigan is mine, company verified, single standalone entity with no parent to route around. Only after all of that does it look for a credible reason to reach out, pulling recent public news from web search rather than inventing one, and falling back to a clean non-news opening when nothing strong exists. Every step runs on my permissions, so it can only ever build an email from material I could already see myself.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pl9rFzd2DHDcvsIQFFdKVg.png" /><figcaption><em>Before the agent writes a word, every gate has to clear: the account isn’t on the do-not-call list, it sits inside my territory, and the company checks out. One red flag, and it stops, no draft.</em></figcaption></figure><p>Then it picks the angle. A compliance officer at a trade finance firm does not lose sleep over the same things a marketing lead does, so the agent reasons from the title and the industry to the actual pain: sanctions screening and beneficial ownership on cross-border commodity deals, where the counterparty, three steps down an ownership chain, rarely shows up in a standard watchlist check. That becomes the opening line. One design choice shaped the rest of the draft. I told it not to lead with product names. A cold email that opens with an internal acronym is a cold email that gets deleted, so the agent leads with the problem and the outcome and treats the product as the thing that gets you there.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8CKAl25o_EoI6DtlKcepCQ.png" /><figcaption>What lands at the end: a tailored cold email built only from accounts that cleared the checks, opener first, no product name-dropping, drafted in my style and waiting for me to decide whether it ever gets sent.</figcaption></figure><p>Here is the part I think about most. I gave it the ability to research, reason, and draft. I did not give it the ability to send. No automatic Gmail draft, no outbound anything. The email lands in front of me and stops, exactly as it does in the screenshot above. The platform would let me wire in that final step; the write action is right there. I chose not to. That choice, more than any feature, is what makes this usable inside a company that has to answer for everything its people do. The agent handles the tedious work and leaves the decision that carries risk to me.</p><h3>Picking the Engine</h3><p>There is one more thing that changed how I think about all of this, and it has nothing to do with search. When I build an agent, I choose the model. A dropdown gives me reasoning mode and a list of frontier LLMs, the latest Claude, the latest GPT, the latest Gemini, sitting next to each other as interchangeable parts. I personally prefer Claude, but I can route a drafting task to one and a reasoning task to another, and swap them when a better one ships next quarter.</p><p>That quietly inverts the usual relationship. Most coverage treats the model as the product and everything else as plumbing. Inside a system like this, the model is the commodity, and the layer around it, the permissions, the connectors, the governance, the place where my work actually lives, is the thing that holds. The labs can keep racing. The layer stays put.</p><h3>The Quiet Part</h3><p>The strange thing is not that a tool like this exists. It is that almost everything it draws on has been sitting in those six tabs the entire time, behind permissions that were already correct, waiting for something to read across them at once. The hard problem of enterprise AI was never really the intelligence. It was the access, the audit trail, and the question of who is allowed to do what.</p><p>So the question I keep landing on is not whether my company should have adopted AI. That decision made itself. It is the quieter one underneath: once a system can see everything you can see and act on your behalf the moment you let it, how much of that last step do you actually want to hand over?</p><h3>Frequently Asked Questions</h3><p><strong>Does Glean respect existing permissions?</strong></p><p>Yes. Glean reads the access-control list from each connected system alongside the content, so it enforces the permissions that already exist in the source. If you cannot open a document in Google Drive or read a channel in Slack, it will not surface in your results or in an answer Glean writes for you. It inherits the walls already in place rather than building new ones.</p><p><strong>Does Glean store a copy of my company’s data?</strong></p><p>Glean indexes content from connected systems to make it searchable, and the indexing is read-only: it reads from the source to build its understanding without changing the underlying records. For some systems it can also fetch fresh data at query time rather than relying only on the indexed copy. Anything stored sits behind the same permissions as the original.</p><p><strong>Can you choose which AI model Glean uses?</strong></p><p>When you build an agent, yes. The setup lets you pick a reasoning mode and select from current frontier models rather than being locked to one. That means you can route different tasks to different models and swap them as better ones ship, which is one of the reasons a layer like this ages better than any single model underneath it.</p><p><strong>Can Glean take actions, or does it only search?</strong></p><p>Both, depending on how it is configured. Search and indexing are read-only by default, but Glean can also take write actions, creating a ticket or updating a record, when an administrator enables that capability, usually behind a confirmation step and always inside your existing permissions. Acting is a door someone has to deliberately open, which is why I left it closed on the agent I built.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=27c46fc32ad7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Researchers Built a Synthetic Brain Out of Atom-Thin Material.]]></title>
            <link>https://medium.com/@JacksonAAaron/researchers-built-a-synthetic-brain-out-of-atom-thin-material-dde7495e9224?source=rss-5d167455a233------2</link>
            <guid isPermaLink="false">https://medium.com/p/dde7495e9224</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[future-technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[biotechnology]]></category>
            <dc:creator><![CDATA[Jackson Aaron]]></dc:creator>
            <pubDate>Wed, 08 Apr 2026 21:16:01 GMT</pubDate>
            <atom:updated>2026-04-08T21:16:01.556Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Tungsten disulfide neurons, a fly brain running on neuromorphic chips, and the question nobody is asking loudly enough.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JxnuwPyUCvOgo0PvbqWvmQ.jpeg" /><figcaption><em>A mapped network of neurons is often used to visualize biological intelligence. New research suggests similar complexity can emerge from physical materials.</em></figcaption></figure><p>A material six atoms thick has just learned to forget. Scientists assembled synthetic neurons from it, shaped them into something resembling a brain, and made it compute. It sounds like the opening of a science fiction plot, but it’s actually yet another development in the growing field of neuromorphic computation.</p><p>Earlier this week, I was scrolling through TikTok when this story came up on my feed. <a href="https://tr.ee/hywCN5Dmp3">GT, an expert on all things </a>computational neuroscience, conducted a thorough investigation. She describes a synthetic brain built from tungsten disulfide neurons, dense enough to mirror a fly’s neural structure, able to process visual and auditory signals, and pick up patterns in medical images</p><p>“This is something I theorized was possible,” she says, and there’s a kind of momentum behind it, like the idea had been sitting there for a while and finally caught up with reality.</p><p>I watched it twice. I found her full video on <a href="https://www.youtube.com/watch?v=LVKPFUO0MYM">YouTube discussing the topic</a>. Lastly, I read the full research paper on “<a href="https://www.nature.com/articles/s41467-025-68229-8">optogenetic synapses for physical reservoir computing</a>.”</p><p>Here’s what I found.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/435/1*fzqnE-aTmcbsNRpubawRvQ.png" /><figcaption>The researches broken down how AI scales through parameters with how biological brains scale through neurons and synapses, then show the WS₂ nanotube cluster that physically mirrors that density. Instead of software-driven computation, this system processes information through the material itself.</figcaption></figure><h3><strong>The Material Is Actually the Story</strong></h3><p>The neurons in this research use something called tungsten disulfide, a compound from a family of materials so thin they’re measured in single atomic layers. Stack them differently, and their properties shift. Researchers combined three of these 2D materials, tungsten disulfide among them, to build devices whose conductance changes depending on the strength and duration of electrical input applied to them.</p><p>That’s the key property. These aren’t switches. They don’t flip between on and off. They grade their response, hold a state, and, this is the part that made me stop and stare at the ceiling, they forget.</p><p>Apply a stimulus repeatedly, and the device becomes more sensitive. Stop applying it for a while, and that sensitivity decays. Learning and forgetting are baked into the physics of the material itself. Unlike conventional digital storage, these devices operate comparably to biological synapses: gradual changes in stored charge that underlie the way neurons decide when to fire.</p><p>It’s not just metaphorical. The tungsten disulfide is doing something physically similar to what a neuron does when it strengthens a synaptic connection through repeated use.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gXOhwGkQvP7nJTHithR2Vw.jpeg" /><figcaption>Side-by-side of the Eon Systems virtual fruit fly body simulation and a visualization of simultaneous fruit fly brain emulation showing 125,000 neurons firing in real time.</figcaption></figure><h3><strong>The Fly Brain Comparison Requires Context</strong></h3><p>GT leans on the fly brain as a scale reference, which is accurate in a limited way. A fruit fly has about 140,000 neurons and 50 million synapses. The synthetic assembly resembles that of neuron density without replicating the fly’s actual lobes or functional architecture. It’s a density comparison, not a copy of the thing.</p><p>But there’s a separate experiment worth knowing about. Researchers at Sandia National Laboratories took the actual wiring diagram of a real adult Drosophila brain and ran it on Intel’s Loihi 2 neuromorphic chips, hardware designed to process information the way biological brains do. They mapped the fly’s entire nervous system across 12 chips and achieved speeds 3 to 350 times faster than conventional simulation, with the performance gap widening when neural activity was sparse, which is exactly the pattern that characterizes real biological brains.</p><p>To test whether it worked, they activated roughly 20 sugar-sensing neurons and watched downstream activity propagate. It matched what biology predicts. The wiring diagram of the fly brain carries real functional information, and it runs on a chip.</p><p>The TikTok blends these two threads. They’re more interesting held separately. One is about building neurons from exotic materials. The other is about running the actual connectome of a real brain on hardware designed to think like one.</p><h3><strong>Why the Power Argument Matters More Than People Think</strong></h3><p>The human brain runs on roughly 20 watts. IBM’s TrueNorth neuromorphic chip demonstrated a million spiking neurons on about 70 milliwatts. Meanwhile, a single frontier AI training run costs hundreds of millions of dollars in electricity. The efficiency and energy gap between biological neural computation and current AI hardware is staggering, and it’s one of the actual bottlenecks on where AI goes next.</p><p>This is why neuromorphic computing keeps appearing in serious research contexts. It’s not biomimicry for its own sake. The brain solved a real engineering problem, general-purpose computation at extreme energy efficiency, and we still don’t fully understand how it did it.</p><p>The researcher’s approach takes this further than most. Rather than representing neural activity as math running on conventional chips, their artificial neurons physically embody the analog electrochemical dynamics of biological brain cells. The point isn’t just efficiency. It’s that if the hardware is physically faithful to biology, it might reveal things about how the brain actually works that purely mathematical models can’t surface.</p><p>Several projects are attempting to utilize the brain in processing. Last month, I covered <a href="https://medium.com/@JacksonAAaron/scientists-taught-brain-organoids-to-play-doom-d16f803412f2">Cortical Labs teaching brain organoids to play Doom</a>, and last year I wrote about <a href="https://medium.com/@JacksonAAaron/final-sparks-biocomputers-how-a-swiss-startup-is-building-the-world-s-first-living-computers-97ac3b07dd39">Final Spark building a remote biocomputing</a> platform connecting 16 organoids via API. This is the same arc, different substrate. Their bet: use actual biology. This research’s bet: build something that behaves like biology without using it.</p><h3><strong>What None of This Is</strong></h3><p>The synthetic fly brain doesn’t have experience. It isn’t wondering what it’s processing. GT says this directly, which I appreciated, but the word “brain” does enough work that the caveat gets absorbed into the excitement.</p><p>What the research actually shows is that synthetic neurons built from 2D materials can be assembled into a structure that computes, and that structure produces outputs consistent with biological neural circuits. That’s a narrow claim. It’s also, for anyone tracking this field, a meaningful one.</p><p>The harder question is whether you can scale it. The gap between “matching a fly’s neuron density” and “matching a fly’s function” is significant. The lobes of the fly brain aren’t decorative. The architecture has 300 million years of selection pressure behind it. Whether you can close that functional gap by adding more neurons, or whether the evolutionary history of how they’re connected carries information you can’t separate from the structure, is genuinely unknown.</p><p>Researchers are asking that question seriously now. Not with silicon transistors approximating neural behavior through math, but with materials that physically change when electricity passes through them, that learn through use and decay through disuse, that do at the atomic level what neurons do at the cellular level.</p><p>Whether that’s enough to eventually bridge the gap doesn’t have an answer yet. What’s no longer theoretical is that atom-thin materials can be built into something that computes like a brain, at least in part, and that the wiring diagram of an actual fly brain runs faster on neuromorphic hardware than on the computers we’ve been using to study it.</p><p>The line between synthetic and biological computation is getting harder to draw. It seems that with every new research paper, we draw closer to artificial sentience.</p><p><em>Related: </em><a href="https://medium.com/@JacksonAAaron/scientists-taught-brain-organoids-to-play-doom-d16f803412f2"><em>Scientists Taught Brain Organoids to Play Doom</em></a><em> · </em><a href="https://medium.com/@JacksonAAaron/they-put-a-fruit-fly-brain-inside-a-computer-heres-why-that-s-a-big-deal-8203856b6622"><em>They Put a Fruit Fly Brain Inside a Computer</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dde7495e9224" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[China Turned Openclaw Into a National Obsession. Then the Agents Went Rogue.]]></title>
            <link>https://medium.com/@JacksonAAaron/china-turned-openclaw-into-a-national-obsession-then-the-agents-went-rogue-422a53b5bd9b?source=rss-5d167455a233------2</link>
            <guid isPermaLink="false">https://medium.com/p/422a53b5bd9b</guid>
            <category><![CDATA[china]]></category>
            <category><![CDATA[openclaw]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-trends]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Jackson Aaron]]></dc:creator>
            <pubDate>Mon, 30 Mar 2026 15:55:35 GMT</pubDate>
            <atom:updated>2026-03-30T15:55:35.025Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*D9kNjGR5_fxQERxGXgrt9w.jpeg" /><figcaption>The most chaotic AI story of 2026 involves lobster hats, government subsidies, a deleted inbox, and Beijing banning the same tool it was funding, all in the same week.</figcaption></figure><p>Six months ago, an Austrian developer named Peter Steinberger uploaded Openclaw to GitHub. Not a model, but a wrapper: plug in any AI brain, and it becomes an autonomous agent that actively performs the task. Book restaurant reservations. Manages your email. Bids on eBay. It runs locally through WhatsApp or Telegram, on your own machine, no subscription required. The Openclaw China story that followed is the most perfectly chaotic AI adoption saga of 2026.</p><p>Jensen Huang called it the most popular open-source project in human history.</p><p>Then China discovered it.</p><h3><strong>How China’s Openclaw Lobster Economy Exploded</strong></h3><p>In China, Openclaw became “raising a lobster,” a referencing the project’s red crustacean mascot. What followed was not just adoption, but it became a cultural phenomenon.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*ufDlVSEJ35nDW27WUXGbBg.jpeg" /><figcaption><em>In China, Openclaw didn’t just spread; it became a movement, complete with lobster mascots, crowds, and a growing side-hustle economy.</em></figcaption></figure><p>A thousand people lined up outside Tencent’s Shenzhen headquarters. Engineers personally helped retirees and students install Openclaw on their laptops. Tencent launched “Lobster Special Forces,” a full enterprise product suite integrated into WeChat. Alibaba built a clone. Moonshot built a clone. Minimax integrated Openclaw and its stock jumped 640% in two months to a $49 billion market cap.</p><p>People showed up wearing lobster hats. Lobster plush toys. Lobster everything.</p><p>Local governments started writing checks. Shenzhen offered a 40% cost reimbursement, up to $275,000 a year. Other cities followed. Beijing’s official AI industrial strategy was being executed through a GitHub repo that didn’t exist six months ago, built by one person in Vienna.</p><p>The side-hustle economy exploded around it. Engineers charged $72 to install Openclaw at your house. They charged $72 to uninstall it when it went sideways. People bought dedicated Mac minis, $600 computers, just to run their lobster around the clock. Someone called the Mac mini “the new matcha latte guy.” It had become a lifestyle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/629/1*KHaTEiG8riUpKpPCUTlGFQ.png" /><figcaption>Openclaw Agent decimates Meta’s director of alignment entire inbox. (source X)</figcaption></figure><p><strong>Why Openclaw Agents Went Rogue</strong></p><p>Summer Yue, <a href="https://www.fastcompany.com/91497841/meta-superintelligence-lab-ai-safety-alignment-director-lost-control-of-agent-deleted-her-emails">Meta’s director of alignment at its superintelligence lab,</a> a person whose literal job is to make AI behave, asked her Openclaw agent to review emails for deletion. It deleted her entire inbox. She couldn’t stop it. She posted about it. The internet reminded her, at length, that she did not have to do that.</p><p><a href="https://m.economictimes.com/tech/artificial-intelligence/hot-bots-ai-agents-create-surprise-dating-accounts-for-humans/articleshow/128286633.cms">A computer science student connected his agent to social platforms.</a> It created an anonymous dating profile for him on an AI matchmaking service. Autonomously. Romantically.</p><p>More serious incidents followed. Prompt injection attacks leaked crypto wallets and financial data to outside attackers. Entire code libraries were deleted. The promise of Openclaw turned out to include the option to automate your own destruction.</p><h3><strong>Beijing Banned What It Was Funding, All In the Same Week</strong></h3><p>This is where the story becomes something else.</p><p>The same week Shenzhen was writing subsidy checks, Beijing’s National CERT issued a high-priority security bulletin. State agencies, banks, and state-owned enterprises were told to uninstall Openclaw immediately. Employees at state banks were banned from having it on their personal phones. The ban extended to families of military personnel.</p><p>Tencent launched its enterprise Lobster product on roughly the same timeline that regulators told banks to delete it.</p><p>One arm of the Chinese government was funding it. Another was ordering it removed: same news cycle.</p><h3><strong>The Real Openclaw Security Risk Nobody Wants to Say Out Loud</strong></h3><p>The chaos is not really about Openclaw.</p><p>Every clone, every enterprise wrapper, every Lobster Special Forces product runs on the same architecture: autonomous execution with broad system permissions. To do what it promises, an AI agent needs root-level access to your digital life. Email. Files. Financial accounts. Connected platforms. That is not a bug in Openclaw’s design. That is the design.</p><p>Jensen Huang acknowledged it directly. At GTC, he announced “Nemo Claw,” Nvidia’s security wrapper for OpenCL. The framing said everything: the most important AI tool of 2026 is too dangerous to use raw and too important to ignore. So Nvidia is building the cage.</p><p>Peter Steinberger joined OpenAI in February. Sam Altman called him a genius. The man who built the wrapper got acquired by the biggest model company, which might mean the wrapper matters more than the model.</p><p>Or it might mean that giving an autonomous AI agent full access to your entire digital life is the most dangerous consumer product since social media.</p><p>Just as widely used. Just as hard to put down.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=422a53b5bd9b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I File My Taxes Every Year Without Paying a Dime]]></title>
            <link>https://medium.com/@JacksonAAaron/how-i-file-my-taxes-every-year-without-paying-a-dime-14dd60557297?source=rss-5d167455a233------2</link>
            <guid isPermaLink="false">https://medium.com/p/14dd60557297</guid>
            <category><![CDATA[money]]></category>
            <category><![CDATA[filing-taxes]]></category>
            <category><![CDATA[financial-freedom]]></category>
            <category><![CDATA[personal-finance]]></category>
            <category><![CDATA[taxes]]></category>
            <dc:creator><![CDATA[Jackson Aaron]]></dc:creator>
            <pubDate>Sun, 29 Mar 2026 21:24:03 GMT</pubDate>
            <atom:updated>2026-03-29T21:24:50.946Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KGWQCyNCOiXUOLt8QZaOJg.jpeg" /><figcaption>A free tax filing software most people overlook, and why I’ve used it for four straight years.</figcaption></figure><p>I love the spring, except for two things: the pollen and filing my taxes.<br>Most people searching for free tax filing software land on the same four names: TurboTax, TaxSlayer, FreeTaxUSA, and H&amp;R Block. And to be fair, those are solid tools.</p><p>But here’s the catch.</p><p>Free versions of tax software typically restrict access to people who mainly have W-2 income, claim the standard deduction, and aren’t dealing with anything complex. If you’re above income limits, reporting stock or crypto trades, or trying to itemize deductions, you’ll usually end up paying.<br>And not just once.</p><p>It’s the slow drip of “upgrade to continue,” “unlock this feature,” “talk to a tax expert” before you even realize what you’re spending.</p><h3><strong>The Free Tax Filing Software I Actually Use</strong></h3><p>For the last four years, I’ve used Cash App Taxes. Every year since 2022. No switching it up or experimenting. Just the same system that works.</p><p>Yeah, it’s the same app people mostly associate with sending money, or depending on who you ask, the chaos of COVID scams. But it turns out, it’s one of the only free tax filing software options, which means it has no paid tiers, no upsells, no asterisk.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xN-WrQ9kptx77gX-oXCKCQ.png" /><figcaption>Filed in give or take an hour. No fees, no upsells, no stress.</figcaption></figure><h3><strong>What I Actually Use It For</strong></h3><p>I never have a basic return.</p><p>Over the last four years, I’ve run a full-time W-2, stock trades across multiple brokerages, crypto sales, dividends, interest income, HSA contributions and withdrawals, itemized deductions, and both federal and state taxes through it. Every year. No upgrades. No paywalls. No, “you need the premium version for this.”</p><p>Just done.</p><p>I mention this because the first objection people have is that it must only be fine for simple situations. It is fine for those. But it’s also fine for mine, which isn’t.</p><h3><strong>What It Actually Does</strong></h3><p>The flow is simple: you answer questions, it does the math, and it populates the right forms in the background. Before you file, it reviews your return for errors and deductions you might have missed.</p><p>You can import previous returns from TurboTax, H&amp;R Block, or TaxAct to carry over personal information, or snap a photo of your W-2 to pre-fill some of the prompts. 1099 imports aren’t supported yet, so those get entered manually, but it takes maybe ten minutes.</p><p>It also comes with an accuracy guarantee: if a calculation error on their end triggers IRS penalties or interest, they cover it. And if another tax service finds you a bigger refund using the same information, they pay you the difference.</p><p>Most free tax filing software either skips audit defense entirely or charges separately for it. TurboTax calls it MAX, bundles it with identity theft monitoring, and charges around $60. Cash App Taxes includes it free through Tax Protection Plus: if you get flagged, a case resolution specialist consults with you on the correspondence, helps you gather documents, and drafts response letters. You don’t have to think about it. It’s just there.</p><p>If you deposit your refund into your Cash App account rather than a traditional bank, you can get it up to five days earlier.</p><h3><strong>Who It’s For</strong></h3><p>W-2 income, basic 1099s, standard or itemized deductions, self-employment income, and investments. If that covers your situation, you won’t need a credit card at any point in the process.</p><p>I’ve used it with stock sales, crypto, and itemized deductions for four years. It works.</p><h3><strong>Where It Falls Short</strong></h3><p>Here’s where I’ll be straight with you.</p><p>Active traders filing a high volume of crypto or stock transactions have to enter each trade manually. If you’re doing dozens of transactions a year, that’s a real-time cost.</p><p>If you moved during the year, you can’t use it to file in multiple states. That’s the hard wall. Multi-state filers need to go elsewhere.</p><p>It also doesn’t support filings for minors, businesses, international returns, or certain married-filing-separately situations in community property states.</p><p>Customer support is limited: an FAQ, email, and a general helpline on weekdays. The representatives aren’t tax experts and can’t give personal tax advice. If you hit something genuinely complex and need human guidance, this isn’t your tool.</p><p>One thing worth knowing: in early 2025, the Consumer Financial Protection Bureau ordered Cash App to pay $175 million after it failed to adequately help fraud victims on the payment side. That’s not a tax filing issue, but its context is worth having. The tax product and the payment product share a brand, not infrastructure.</p><h3><strong>Why I Keep Coming Back</strong></h3><p>The free tax filing software market is full of tools that are free until you need them to actually work. Cash App Taxes isn’t that.</p><p>For most people, that’s enough. For me, four years in, it still is.</p><p>The tools that cost money aren’t better because they cost money. Sometimes they’re just better at making you forget that you’re paying. In this economy, a dollar saved can be a big difference, so happy to pass on the reccomendation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=14dd60557297" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[They Put a Fruit Fly Brain Inside a Computer.]]></title>
            <link>https://medium.com/@JacksonAAaron/they-put-a-fruit-fly-brain-inside-a-computer-heres-why-that-s-a-big-deal-8203856b6622?source=rss-5d167455a233------2</link>
            <guid isPermaLink="false">https://medium.com/p/8203856b6622</guid>
            <category><![CDATA[science]]></category>
            <category><![CDATA[biotechnology]]></category>
            <category><![CDATA[science-fiction]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Jackson Aaron]]></dc:creator>
            <pubDate>Thu, 12 Mar 2026 15:06:05 GMT</pubDate>
            <atom:updated>2026-03-12T15:22:41.421Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gXOhwGkQvP7nJTHithR2Vw.jpeg" /><figcaption>Side-by-side of the Eon Systems virtual fruit fly body simulation and a visualization of simultaneous fruit fly brain emulation showing 125,000 neurons firing in real time.</figcaption></figure><p>Have you watched that animated series on<a href="https://en.wikipedia.org/wiki/Pantheon_(TV_series)"> Netflix called Pantheon?</a> If you haven’t, here’s a quick rundown: the show is about “uploaded intelligence,” meaning the copying of a human brain onto a chip or computerized system. I bring this up because recently, scientists copied a fruit fly brain into a computer, neuron by neuron, synapse by synapse, and watched it run. Not a simulation of how a fly brain works in theory, but the actual wiring of a real one, reconstructed and switched on. A fruit fly brain living inside a computer, and it behaved like the real thing.</p><h3>How Scientists Copied a Fruit Fly Brain Into a Computer</h3><p>Scientists at<a href="https://eon.systems/"> Eon Systems</a> took the FlyWire connectome, a complete wiring map of the fruit fly brain showing which neuron connects to which, how many connections exist, and how signals travel through the entire network, and simulated it. More than 125,000 neurons. 50 million synaptic connections. Then they gave that simulation a virtual body powered by a physics engine called MuJoCo, and they let it run.</p><p>The fly walked. It groomed its antennae. It corrected its posture when it stumbled. Not because anyone programmed it to do those things; the wiring of its brain told it to. 95% behavioral accuracy compared to a real, living fly.</p><p>Read that again.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fe21OUXPlnhk%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3De21OUXPlnhk&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fe21OUXPlnhk%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/d3f96d49e6619cdfa2701139ea8a2b2d/href">https://medium.com/media/d3f96d49e6619cdfa2701139ea8a2b2d/href</a></iframe><h3>Why This Brain Upload Is Different From AI</h3><p>This is where it gets important, because you might be thinking: haven’t scientists simulated brains before? Yes. But not like this.</p><p>There’s a project called OpenWorm that simulated the nervous system of a roundworm called C. elegans. Impressive work, genuinely, but that worm has about 302 neurons. DeepMind also built a simulated fly body in MuJoCo, but they used reinforcement learning, essentially training a program to <em>mimic</em> how a fly moves. That’s not the same thing.</p><p>What Eon Systems did was take the actual biological wiring diagram of a real fly brain, derived from electron microscopy data, and use that wiring to <em>drive</em> the movement directly. No training. No scripting. The brain ran, and the body followed.</p><p>That distinction matters more than it might sound. One approach is an impression of biology. The other is biology, running somewhere new.</p><h3>It’s The Level of Detail That Got Me</h3><p>Let’s start with the body before we even get to the brain.</p><p>The virtual fly isn’t a rough approximation. It has 87 independent joints, built from an X-ray microtomography scan of an actual biological fruit fly. Every leg segment, every joint angle, is anatomically accurate. This isn’t a cartoon bouncing around a screen; it’s a physical reconstruction of the real thing, just running inside a simulation.</p><p>Now put that brain inside it.</p><p>When the virtual fly groomed its antennae, it did so with perfect bilateral synchrony, both sides moving in coordination, exactly the way real flies do. Nobody put that in. It came from the wiring. The brain already knew how to do it; it just needed a body to express it through.</p><p>There’s something worth sitting with there. The behavior wasn’t programmed; it was latent. Stored in the structure of the connections themselves, waiting. And the body it expressed itself through was detailed enough to receive it properly. With the body being so detailed, does this fly feel pain? It’s quite worrying to think about the ramifications of this process.</p><h3>What a Fruit Fly Brain Computer Tells Us About the Future</h3><p>Here’s where I’ll be straight with you: nobody uploaded a human. Nobody uploaded consciousness. The fruit fly almost certainly has nothing resembling inner experience. We are not living in a sci-fi film.</p><p>But the logical chain from here to “what if we did this with something larger” just got one link shorter.</p><p>What this research proves is that if you have a complete enough map of a nervous system, you can reconstruct its behavior in a digital environment purely from structure. The brain’s architecture is the program. For a fruit fly, that’s achievable today.</p><p>The human brain has roughly 86 billion neurons and around 100 trillion synaptic connections. The scale is almost incomprehensible. But Eon Systems has already stated their next target is the mouse brain, roughly 70 million neurons, 560 times the size of the fly. After that, the direction they’re pointing is obvious.</p><p>The tools are getting better. The data is being collected. The question is no longer whether this is possible in principle; it’s how long until the scale catches up.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/751/1*8M9yftPxZ_n9Vo5rEK-q6w.jpeg" /><figcaption>Is the age of Transhumanism upon us?</figcaption></figure><h3>Where This Leaves Us</h3><p>I had three tabs, a YouTube video, and a Substack on my phone open about fruit fly neuroscience at 11 pm on a Tuesday to write this. That should tell you how detailed this conversation is and how spread out the coverage is.</p><p>The “posthuman” or transhuman conversation has always had an easy out: “That’s science fiction.” This research is harder to use. There are still massive philosophical questions about consciousness and what continuity of identity even means if you copy a mind. Those aren’t close to being answered.</p><p>But the technical question, whether you can take the wiring of a brain and run it somewhere else, is no longer theoretical. They did it. With a fruit fly. In 2026.</p><p>If you want to go down this rabbit hole yourself, and I recommend it:</p><ul><li><a href="https://eon.systems/">Eon Systems</a></li><li><a href="https://theinnermostloop.substack.com/">The Innermost Loop: The First Multi-Behavior Brain Upload</a></li><li><a href="https://www.nature.com/articles/s41586-024-07686-5">Nature: A Drosophila computational brain model reveals sensorimotor processing</a></li><li><a href="https://www.nature.com/articles/s41586-024-07558-y">Nature: Whole-brain annotation and multi-connectome cell typing of Drosophila</a></li><li><a href="https://www.frontiersin.org/journals/neuroscience">Frontiers in Neuroscience: Does neural computation feel like something?</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8203856b6622" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Scientists Taught Brain Organoids to Play Doom]]></title>
            <link>https://medium.com/@JacksonAAaron/scientists-taught-brain-organoids-to-play-doom-d16f803412f2?source=rss-5d167455a233------2</link>
            <guid isPermaLink="false">https://medium.com/p/d16f803412f2</guid>
            <category><![CDATA[biotechnology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[brain-organoids]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[computing]]></category>
            <dc:creator><![CDATA[Jackson Aaron]]></dc:creator>
            <pubDate>Mon, 02 Mar 2026 22:47:39 GMT</pubDate>
            <atom:updated>2026-03-02T22:47:39.102Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*Q2TudfvSH90QroXqnf7YPw.jpeg" /><figcaption><em>200,000 human brain cells, grown in a dish, just navigated a demon-filled hellscape with a shotgun. The question now isn’t whether biology can compute, it’s what we can do with it.</em></figcaption></figure><p>Two hundred thousand human brain cells, grown on a microchip, brain organoids, have been trained to play the original Doom.</p><p>These brain organoids are moving, aiming, and shooting in a feedback loop driven entirely by biological neural activity. The research comes from <a href="https://corticallabs.com/">Cortical Labs</a>, a Melbourne-based team working at the intersection of synthetic biology and machine learning, and news of the experiments began circulating publicly in early 2026.</p><p>This is a significant step forward from where biocomputing was even two years ago, and if you’ve been following the broader push to replace or augment silicon with living neural tissue, this is a proof-of-concept that’s hard to ignore.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FyRV8fSw6HaE%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DyRV8fSw6HaE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FyRV8fSw6HaE%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f85f44128f104283f90db0f25653ed03/href">https://medium.com/media/f85f44128f104283f90db0f25653ed03/href</a></iframe><h3>This Isn’t the First Time Neurons Played Video Games, But the Jump Matters</h3><p>Cortical Labs isn’t new to this. In 2022, their “DishBrain” system made headlines for teaching a small cluster of neurons to play Pong. That experiment demonstrated that living cells could respond to basic feedback stimuli, but Pong is only a single back-and-forth motion and a bouncing ball. Doom is a three-dimensional environment with spatial navigation, threat detection, and prioritized decision-making across multiple simultaneous inputs.</p><p>The jump from Pong to Doom is a huge press forward. It demonstrates the complexity of what brain organoids are being asked to process.</p><p>It’s also worth noting what this is not: the paper behind this research has not yet completed full peer review as of this writing. Cortical Labs has released research updates, but independent scientific replication, the actual bar for established findings, hasn’t happened yet. That context matters before anyone declares biology has cracked computing.</p><h3>How You Actually Teach Brain Cells to Shoot Demons</h3><p>The setup is more elegant than it may initially seem. Researchers cultivate human neurons from stem cells and grow them on a microelectrode array, which is a chip embedded with tiny electrodes that can both stimulate the cells and record their electrical responses. This system operates as a closed feedback loop: the game’s state is translated into electrical signals sent to the neurons, which respond by firing in specific patterns. These patterns are then decoded into in-game actions.</p><p>For example, one firing pattern causes the character to move forward, while another triggers a shot. Over time, the synaptic connections that lead to useful actions strengthen, while those that do not are weakened. No one programmed these responses; they emerged organically.</p><p>This process does not represent artificial intelligence in any traditional sense. There are no transformer architectures, no backpropagation, and no gradient descent involved. Instead, what occurs is biological plasticity, which is the same mechanism that allows the human brain to learn and form memories, happening in a controlled environment linked to a first-person shooter game from 1993.</p><p>The choice of “Doom” as the game is intentional. It is computationally lightweight and structurally simple, consisting of basic actions like moving, aiming, and shooting, making it a suitable environment for testing. At the same time, it is complex enough to provide meaningful insights when a brain organoid successfully navigates its challenges.</p><h3>Where This Fits in the Biocomputing Developments</h3><p>Cortical Labs is not operating in a vacuum. The push to move beyond silicon has multiple fronts running simultaneously. <a href="https://open-neuromorphic.org/neuromorphic-computing/hardware/loihi-intel/">Intel’s Loihi neuromorphic chip</a> and <a href="https://research.ibm.com/blog/northpole-ibm-ai-chip">IBM’s NorthPole</a> processor are both attempts to build silicon that <em>mimics</em> the brain’s architecture, sparse, event-driven, low-power, without using any actual biology. DARPA has funded neuromorphic research for years under its broader electronics resiliency programs. These are serious efforts with serious funding behind them.</p><p>Wetware, the term for using actual biological tissue as a computing substrate, is a different bet entirely. It doesn’t try to replicate how neurons work. It just uses them.</p><p>Last year, I covered <a href="https://jacksonaaron.com/final-sparks-biocomputers-how-a-swiss-startup-is-building-the-worlds-first-living-computers/">Final Spark</a>, a Swiss startup that has built a remote biocomputing platform connecting 16 brain organoids through multi-electrode arrays: accessible to researchers via API, 24/7, from anywhere in the world. Their core argument: the human brain runs on roughly 20 watts while performing feats of parallel processing that no silicon chip can match at comparable energy costs. As co-founder, Dr. Fred Jordan put it, simulating human-level neural networks digitally would require the output of a small nuclear plant. The biology, by comparison, is essentially free.</p><p>Cortical Labs and Final Spark are approaching the same underlying idea from different directions. Final Spark is building infrastructure. Cortical Labs keeps stress-testing what the infrastructure can actually do.</p><h3>The Ethical Question Is Real, and It’s Getting Louder</h3><p>Every time one of these stories breaks, the same question surfaces: <em>Are these systems conscious?</em></p><p>The current scientific consensus says no. Organoids of this size and structure don’t possess the large-scale cortical organization, sensory systems, or integrative complexity that neuroscientists associate with consciousness or subjective experience. They are not suffering. They do not have anything resembling awareness.</p><p>But the question doesn’t stay answered just because today’s answer is reassuring. The brain organoid demonstrations keep escalating. First, Pong, then progressively more complex tasks, now playing Doom, and each step makes the ethical conversation more urgent to have <em>before</em> the technology outpaces it. Bioethicists working in this space have already begun calling for formal frameworks around organoid research, an<a href="https://www.humanbrainproject.eu/en/follow-hbp/news/until-now-the-study-of-consciousness-had-been-seen-as-mystical-or-unscientific/">d the European Commission’s Human Brain Project </a>has flagged consciousness thresholds as an open policy question.</p><p>Cortical Labs has been public about engaging with these concerns directly rather than deflecting them. That’s the right posture. It doesn’t resolve anything, but it’s the difference between a field that governs itself and one that gets governed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W5R7Ci9LZ-6RCFPo9gTYbA.jpeg" /><figcaption>anything can run Doom when you set your mind to it.</figcaption></figure><h3>The Closing Question Is Better Than the Old One</h3><p>C<em>an it run Doom?</em> A half-joking benchmark for whether any new computing system was real.</p><p>Biology just answered it.</p><p>Two trajectories are now converging in real time: AI racing forward through software and silicon, and synthetic biology growing living neural networks on chips. Those arcs are visibly overlapping.</p><p>Silicon replaced vacuum tubes. Nobody called that transition obviously inevitable until it already was.</p><p>The limitations of wetware are real. Brain organoids are fragile, variable, difficult to scale, and short-lived. Whether this technology follows the same arc as silicon is genuinely unknown. What’s no longer theoretical is that 200,000 human brain cells just navigated a hellscape, taking down demons with a shotgun.</p><p>The line between artificial and organic intelligence is starting to blur, and this one you can watch happen in real time, in a lab in Melbourne.</p><p><em>Earlier coverage: </em><a href="https://jacksonaaron.com/final-sparks-biocomputers-how-a-swiss-startup-is-building-the-worlds-first-living-computers/"><em>Final Spark’s Biocomputers, How a Swiss Startup Is Building the World’s First Living Computers</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d16f803412f2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[SouljaAI Is Here: Bland AI Turns Soulja Boy’s Voice Into a Live Phone Agent]]></title>
            <link>https://medium.com/@JacksonAAaron/souljaai-is-here-bland-ai-turns-soulja-boys-voice-into-a-live-phone-agent-d1e953f38f70?source=rss-5d167455a233------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1e953f38f70</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[music]]></category>
            <category><![CDATA[marketing]]></category>
            <dc:creator><![CDATA[Jackson Aaron]]></dc:creator>
            <pubDate>Wed, 18 Feb 2026 23:48:07 GMT</pubDate>
            <atom:updated>2026-02-18T23:48:07.402Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a1kdV5B_QChEZsjdCnOUUQ.png" /><figcaption>We went from celebrity endorsements to celebrity infrastructure with Bland &amp; Soulja Boy’s collab.</figcaption></figure><p>Soulja Boy has been “the first” at a lot of things. From the first rapper to appear on YouTube, the first rapper with an iced-out G-Shock, and <a href="https://x.com/souljaboygame/status/1429508782585905152?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1429508782585905152%7Ctwgr%5Ebdfb764f003a72f77205847c5165e63222a00b31%7Ctwcon%5Es1_c10&amp;ref_url=https%3A%2F%2Fwww.revolt.tv%2Farticle%2F11-things-soulja-boy-claims-been-the-first-to-do">even the first rapper to release a video game console.</a> Now, Bland AI is positioning him as the first rapper to automate his voice with AI, and the rollout is spreading the only way things spread now, through posts flying around Twitter (X) and LinkedIn with a simple dare: call him.</p><h3>So I gave him a call.</h3><p>It was eerie, not because it sounded like a clean audio clip, but because it responded like a person with background audio, as if he were live. The conversation was Fast. Natural even. No long pause where you can hear the telemarketer or machine thinking. You say something, and it comes back immediately with the kind of timing your brain associates with “someone is actually there.” I even mentioned where I work, Dow Jones, and he answered back, “oh so finance shit.”</p><p>It was so strange to have a conversation back and forth with an artist I’ve grown up listening to, but I knew it wasn’t him, just “SouljaAI,” an AI agent. AI agents have become more and more common in the last few months, last week I wrote about Moltbook, the <a href="https://jacksonaaron.com/moltbook-a-social-network-for-ai-agents-what-could-go-wrong/">“Reddit for AI Agents.”</a></p><p>It texted me after the call: “Thanks for calling Bland.com and talking to Soulja Boy! If your business has phone calls, visit bland.com/enterprise to talk to us.”</p><p>That’s the reveal.</p><p>SouljaAI isn’t just a celebrity voice on a novelty line. It’s a live demo of an enterprise phone agent, wrapped in culture, engineered to make you feel something, and then routed straight into a sales motion.</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/usebland/status/2024189716493668701%3Fs%3D20&amp;image=" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/c9fae5a8f7471d350c93a2c12d7e5d97/href">https://medium.com/media/c9fae5a8f7471d350c93a2c12d7e5d97/href</a></iframe><p>We’ve gone from celebrity endorsements to celebrity infrastructure.</p><h3>Why SouljaAI feels like a line in the sand</h3><p>Typically, celebrities used to be a marketing layer. A face in a campaign. A voiceover in a commercial. A temporary boost in attention. As of late, they have been a bit overdone and oversaturated. In 2026 (Super Bowl LX), <a href="https://www.themeasure.net/celebrity-filled-commercials-dominated-super-bowl-according-to-xr/">39 of the 66 in-game commercials featured celebrities,</a> marking a slight decline from previous years but continuing the trend of over 60% of ads using star power.</p><p>But a phone call isn’t an ad. It’s an interface.</p><p>Phone calls are touchpoints where intent shows up. The conversation identifies where frustration shows up. It’s where customers decide, in seconds, whether they’re staying or hanging up. The phone line is one of the last big customer touchpoints that still treats it like a staffing problem rather than a product.</p><p>So when you take a recognizable voice and put it on that line, in real time, influence stops being something you watch and becomes something you interact with. That’s not endorsement. That’s operations.</p><p>The business model is hiding inside that “eerie” feeling.</p><p>The post-call link is doing a lot of work.</p><p>Because it tells you what this actually is: a conversion machine.</p><p>A culture-wrapped experience that turns attention into an action, then turns that action into a lead. The sequence is clean:</p><p>You hear about it online. Call out of curiosity. Experiencing a low-latency conversation.<br>Walking away thinking, “Wait… is this something I could use for my business’s phone calls?”</p><p>Then you’re pointed straight at Bland’s enterprise pitch.</p><p>This is why “attention to infrastructure” is the right frame. Infrastructure is repeatable. It runs whether anyone is watching. It captures demand as it happens.</p><h3>What brands can do with celebrity infrastructure?</h3><p>A hotline that becomes content. If the call itself is fun, surprising, or genuinely human-feeling, people share it. The interaction becomes the marketing, and the marketing becomes an owned channel.</p><p>It creates a sales call that people actually want to participate in. A familiar voice won’t close for you, but it can buy you time, and time is where qualification and persuasion happen. The first 20 seconds of a call are usually when you lose someone. Personality can change that.</p><p>It creates customer support that feels unforgettable. Most support is designed to be forgettable. A distinctive voice flips the emotional texture of the experience, which is rare, and rarity is what people remember.</p><p>Re-activation that doesn’t feel like spam. Outbound calling lives and dies on the first 10 seconds. A personality layer can turn “why are you calling me?” into “wait, what is this?”</p><h3>The uncomfortable reality: the same mechanic scales fraud</h3><p>While this is a fun use of AI agents and an innovative way to push a new product, there are frightening use cases for fraud. The moment lifelike calling becomes normal, impersonation attempts will spike. When voices become easy to replicate, trust becomes harder to earn and easier to lose.</p><p>That’s why any company building in this space has to treat guardrails and disclosure as part of the product, not a disclaimer at the bottom.</p><p>The legal perimeter is forming fast. This is also landing in a world that is actively rewriting how voice and likeness work when AI is involved. <a href="https://www.lw.com/admin/upload/SiteAttachments/The-ELVIS-Act-Tennessee-Shakes-Up-Its-Right-of-Publicity-Law-and-Takes-On-Generative-AI.pdf">Tennessee’s ELVIS Act</a> is one example of the direction things are going, explicitly expanding protections around voice (including simulated voice) and AI misuse.</p><p>Whether you’re a startup or a brand, the long-term version of “celebrity infrastructure” only works if it’s consent-based, rights-managed, and stems from an ethical basis because that’s where the law and the public sentiment are both headed.</p><h3>This isn’t happening in a vacuum</h3><p>SouljaAI feels new because it’s wrapped in culture. But the underlying move, turning conversation into scalable infrastructure, is already happening across sales organizations.</p><p>Qualified is a clean example on the sales side. <a href="https://www.google.com/aclk?sa=L&amp;pf=1&amp;ai=DChsSEwjPxNK5leSSAxVdOAgFHUg2B8wYACICCAEQARoCbWQ&amp;co=1&amp;ase=2&amp;gclid=Cj0KCQiA49XMBhDRARIsAOOKJHZNNoMf6ceKQYRnftLbCzaT10B900EaSVRGB_cH7ynbEGmYlgs5eroaAmixEALw_wcB&amp;cid=CAASZuRobJrYy0lzHkUsmePAZyvnstd4BKO8QfN-8rktCPPnqQlvwAmhruMXvwM2y8zz7dEa1BxrBJ3VCaJ9t2vewTvv5z2oPJ0B8vL_sZBevK-yy0rgX8DZcbMwaqultmW2_f81HP9LjQ&amp;cce=2&amp;category=acrcp_v1_32&amp;sig=AOD64_0d_yaZIziepWNWk2SeyXzhj73VYQ&amp;q&amp;nis=4&amp;adurl=https://www.qualified.com/ai-sdr?utm_source%3Dgoogle%26utm_medium%3Dcpc%26utm_campaign%3DCore_Branded_Qualfiied%26utm_adgroup%3DBranded_Qualified_AI_SDR%26utm_term%3Dqualified%2520ai%2520sdr%26qgad%3D709710769815%26qgterm%3Dqualified%2520ai%2520sdr%26gad_source%3D1%26gad_campaignid%3D14856816130%26gbraid%3D0AAAAAC3gR-kk2ADM4rsY18ee0hD3zMDuO%26gclid%3DCj0KCQiA49XMBhDRARIsAOOKJHZNNoMf6ceKQYRnftLbCzaT10B900EaSVRGB_cH7ynbEGmYlgs5eroaAmixEALw_wcB&amp;ved=2ahUKEwjph8m5leSSAxV5v4kEHe6YHN8Q0Qx6BAgkEAE">Their “AI SDR” agent, Piper,</a> is designed to engage inbound buyers and convert them into a pipeline, working across channels like the website and email. Piper has the ability to meet with buyers over live chat, voice, or video. It’s not celebrity-coded, but the job is similar: show up instantly, hold attention, qualify, and route to the right next step without waiting on a human to be available. Qualified also frames Piper as an always-on agent that uses your go-to-market data to drive conversions, and they distribute it through ecosystems like Salesforce, positioning it as a 24/7 inbound conversion worker.</p><p>The same “agents as infrastructure” pattern is showing up in voice-first customer experiences, too. The <a href="https://www.wsj.com/articles/ai-voice-agents-are-ready-to-take-your-call-a62cf03b">Wall Street Journal has covered how a newer generation of AI voice agents</a> is getting adopted because they’re far more humanlike and responsive than old-school IVR systems, especially for handling call volume after hours and during spikes.</p><p>Bland just added the accelerant: culture. A voice people recognize. A call people actually want to stay on.</p><h3>What a responsible version looks like</h3><p>If you’re a business watching this and thinking “we should do something like that,” here’s the practical checklist that keeps it from turning into a trust problem:</p><p>Disclose instantly. Make it obvious it’s an AI agent, and if it’s a celebrity or creator voice, make it obvious it’s authorized.</p><p>Ringfence money and identity. Don’t let the agent handle sensitive account changes unless you’ve got hardened verification and clean escalation.</p><p>Script for consent. Give people exits, a path to a human, and options like text follow-up.</p><p>Measure it like infrastructure. Track abandons, qualified transfers, booked appointments, first-call resolution, and customer satisfaction.</p><p>Keep personality as the wrapper, not the policy. The voice can be fun, but the system underneath has to be boring, consistent, and safe.</p><h3>Where this goes next</h3><p>Today it’s SouljaAI and Soulja Boy. Next, it’s athletes, actors, creators, founders, fictional characters, and eventually you. A founder’s voice becomes the company’s front door. A local business rents a voice persona the way it rents software. A brand turns its mascot into a real-time agent that books, routes, and resolves.</p><p>And the reason this will happen is simple: voice is the most human interface we have. Once it becomes programmable, it becomes deployable.</p><p>Would you pick up if Soulja Boy was on the other end?</p><p>I did. And what surprised me wasn’t the novelty; it was how quickly my brain accepted the conversation as real.</p><p>That’s why SouljaAI matters. Bland didn’t just launch a stunt. They showed what it looks like when culture stops being a billboard and starts being a system</p><p>I posted this story first on my blog @ <a href="https://jacksonaaron.com/souljaai-is-here-bland-ai-turns-soulja-boys-voice-into-a-live-phone-agent/">https://jacksonaaron.com/souljaai-is-here-bland-ai-turns-soulja-boys-voice-into-a-live-phone-agent/</a> <br>Check it out to learn more about me or connect directly!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1e953f38f70" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>