<?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 Madhuban on Medium]]></title>
        <description><![CDATA[Stories by Madhuban on Medium]]></description>
        <link>https://medium.com/@madhuban_67926?source=rss-7c2710636fe7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by Madhuban on Medium</title>
            <link>https://medium.com/@madhuban_67926?source=rss-7c2710636fe7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 10:39:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@madhuban_67926/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[Ramblings about Frontier models IP, Energy, open source and brain fog]]></title>
            <link>https://medium.com/@madhuban_67926/ramblings-about-frontier-models-ip-energy-open-source-and-brain-fog-6bb6c004c184?source=rss-7c2710636fe7------2</link>
            <guid isPermaLink="false">https://medium.com/p/6bb6c004c184</guid>
            <dc:creator><![CDATA[Madhuban]]></dc:creator>
            <pubDate>Sun, 19 Jul 2026 20:52:17 GMT</pubDate>
            <atom:updated>2026-07-19T20:52:17.643Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Zt1iPsQJVau4lLH6fIBTPg.png" /><figcaption>Visual by Claude Design</figcaption></figure><p>Back when Windows first started, one recalls how the Windows operating system began to eat into business, initially with partners like Lotus 123 and WordPerfect, but eventually built in-house through the convenience of Excel, Office, and the operating system, which sat outside the company’s data layer. Then Google came along and, through search, indexed and helped you find the best sites for information. It studied your search history and created a closed loop by setting up properties in shopping, flights, and more. It had court cases in Europe against this practice. Today, less than half of the search comes from offsite properties. Again, it sat outside your data layer. Amazon, similar story, building its own brand by tracking those it enabled through its fulfilment business.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/851/1*smITa3Y6EwzsnjopGGRDTA.png" /><figcaption>Visual developed by Claude design</figcaption></figure><p>These examples of past technology shifts are deeply relevant to today’s AI landscape. Each of these companies initially served as an entry point to data, creating significant value by operating outside core business processes and gradually building closed ecosystems that curtailed competition and influenced user behaviour. Now, we are witnessing a similar inflection point with foundational AI models: they are easy to use, sit directly in the intelligence layer of the company and your brain, and threaten to pull IP, judgment, and alpha inward. This shift from external tools to embedded intelligence increases risk for tech professionals and businesses. It is no longer simply about adopting technology, but deliberately considering how much intellectual property, decision-making, and unique advantage are being handed over to the model providers. For companies, the issue is the alpha, the cost of using these models, and the energy impact on the business and society.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/919/1*1dtPjVZImxS49BKXJHup7w.png" /><figcaption>Visual developed by Claude design</figcaption></figure><p>Foundational models, be it Anthropic, OpenAI, Grok, Gemini, DeepSeek, or Kimi 3.0, are super easy to use and have helped close the gap between information and intelligence, and more. I started to see people dump all their data and slides to get more intelligence or quick, easy information as needed. It now sits in your data layer, hoovering up IP, intelligence and possibly thoughts (more on that later). It’s silently taking your alpha, that competitive differentiation and turning the business into a commodity. Today, Anthropic has built the Application layer based on all the dumped data for code and design and is extending it to healthcare, financial services, legal, and life sciences. The code was based on the traction it saw from Cursor. Its CPO served on Figma’s board three days before the launch of Claude Design. So worthy of mentioning.</p><p>Digging deeper and listening to <a href="https://www.linkedin.com/company/allinpod/">AllIn</a>, they discuss this in detail on their podcast. Chamath makes an interesting point that 50% of companies lack the ability to grow beyond the current long-term cost of capital, which is between 8% and 11%. Low-return companies with 1–3% are 1–7 companies globally. So, what is the outcome? They will need a productivity enhancer. And along came the world of Frontier models, which only further commoditised their businesses into one undifferentiated blob. It sucked alpha and served it back as undifferentiated goop, while companies paid for tokens to increase productivity and the employees, possibly a little dumber or at least a little zombified. MIT did some research on this <a href="https://www.bbc.co.uk/news/articles/cd6xz12j6pzo">here</a> and called it ‘cognitive atrophy’. In this MIT study, the authors found that employees who rely heavily on AI tools for decision-making can experience a measurable decline in their analytical thinking and problem-solving abilities. The studies found that over time, frequent use of generative AI made professionals less likely to recheck AI outputs, increasing the risk of passing errors down the line and making teams more passive. The main implication for tech professionals is to stay actively engaged in reviewing AI-generated solutions, to use these tools as collaborators rather than replacements, and to be mindful of retaining the distinct skills and expertise that set teams apart from the competition.</p><p>I digress, but it’s worth bringing up the world of open source. Kimi released its new model at 1% of Anthropic’s cost. What does that mean for closed models that cost almost 99% more, and where does this lead with the data centre rush? Kimi’s K3 is a 2.8-trillion-parameter model that the company describes in its <a href="https://www.kimi.com/blog/kimi-k3">technical blog</a> as the world’s first open 3T-class system and the largest open-weight AI model to date. While it’s still behind Anthropic’s Claude Fable 5 and OpenAI’s GPT 5.6 Sol on overall performance, it outperformed every other model in the company’s evaluation suite, including Claude Opus 4.8 and GPT 5.5, across coding and agentic <a href="https://www.tomshardware.com/uk/tag/benchmark">benchmarks</a>. The model has a 1 million token context window, native vision, and activates just 16 of its 896 experts per token, roughly 1.8% of the pool. Local models like Qwen run on your laptop and are similar to Claude Sonnet and Bonsai 27B, with a 27-billion-parameter model that runs on your smartphone. Distributed training with Bittensor used 70 GPUs to train a 72-billion-parameter model over the internet. So, while locals have a long way to go, open is almost here, and at 1% of the cost, not all jobs need to be token maxed.</p><p>It’s important to compare where open or local models are preferable now and where closed models still offer unique <a href="http://variousresearch/">value.</a> Open or local models shine when cost, data privacy, and intellectual property control are key concerns. For routine work such as code migration, internal data analysis, prototyping, or workloads that can tolerate slightly lower accuracy or speed, open models and local deployment offer meaningful benefits in terms of cost savings and customisation, while still requiring one’s brain. However, closed models such as GPT-5.6 Sol or Claude Fable 5 maintain a lead in state-of-the-art capabilities, handling highly complex, nuanced, or state-of-the-art language tasks with greater reliability and access to the latest research. For regulated industries with sensitive use cases, leveraging validated closed systems may still be warranted; however, they may still capture your IP. Overall, the strategic decision boils down to considering cost, privacy, performance, and business risk: open and local for most business-as-usual or privacy-sensitive workflows, closed when you need the absolute latest intelligence and accuracy for mission-critical or high-stakes tasks.</p><p>Further, with local and open, it will not drive the energy consumption through the roof at the data centre where your query is processed. Most importantly, using local and open will help keep costs down and definitely protect your alpha for businesses currently token-maxing and leaking alpha. Energy consumption by foundational models is going through the roof, and we, the public, may be subsidising it through non-commodity costs passed on to our bills (I won’t digress, but this is another topic).</p><p>So, let’s understand the type of job, why you need the cutting-edge model and why you cannot use your brain?</p><p>A task involving Typical old Enterprise code migration to a new framework may not be alpha-driven or time-critical, but the data it contains is.</p><p>Why not try local mode, which may be cheaper and slower, but protects your Alpha and keeps your token and energy costs down</p><p>Software Factory + Claude 1.4X cheaper and 1.5X faster</p><p>Frontier open source and Software factory 16.2X cheaper and 3.2X slower</p><p>Today, we have</p><p>Large hub: clusters of capital advantages (OpenAI, Anthropic, etc.) for training models, and we dump everything into these large black boxes, but there is a tiered approach of</p><p>Large spokes: clusters to deploy models with inference (neocloud and hyperscaler, etc.)</p><p>Medium Hubs: Enterprise models with local prem</p><p>Smaller Hubs: cluster for training proprietary models using their own data</p><p>Distributed spokes: servers to run models locally, with their own inference and weights workflows run locally, value of on-prem</p><p>Not every task needs to be done at the speed of light, so an open-source system running on your local hard drive and processed locally may save the day while saving your company’s Alpha, tons of token costs, energy, water and, in a small way, your brain as you work on the code and handcraft the outcome sought. To put this in perspective, running a l<a href="https://medium.com/@arsen.apostolov/how-much-does-it-actually-cost-to-run-a-local-llm-per-million-tokens-measured-4a90a7f31a48">ocal model</a> like Qwen or Llama 3 on a standard workstation can cost as little as $0.05–$0.10 per hour in electricity, compared with cloud-based API charges that can exceed $2–$3 per 1,000 input/output tokens for advanced closed models. One financial services team reported reducing monthly <a href="https://intuitionlabs.ai/articles/deepseek-inference-cost-explained">AI inference costs</a> from over $8,000 (using closed APIs) to less than $100 by moving to open-source models on on-prem GPUs. In terms of IP, local deployment keeps proprietary data within the company’s firewall, stopping accidental data leaks into third-party servers or model providers. On the energy front, some studies have shown that running LLMs locally can use up to 20 times less energy per query than centralised hyperscaler APIs, since you skip networking and data centre load. Act local and think local may help not only with using foundational models to build your creativity but also, if you extrapolate, with being cheaper, safer, and better for the community and for energy. A 100X loss in profits is a high cost for already struggling companies that have now also lost their IP and been commoditised, facing further price challenges, as technology has always been a price compressor.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/920/1*dezs4vny-GbHX956afnM3A.png" /><figcaption>Visual by Claude Design</figcaption></figure><p>With local hard drives, you own your GPUs to process local data in local data centres. A thought, more now than ever before: look at services that truly think local and act local. So, build in your environment, keep the weights and harness local, and build on open source. Own the model, the data and the weights with the harness. This will save your IP; it’s miles cheaper and, more importantly, keeps control local. I leave with more questions than answers.</p><ul><li>So, what is the answer and how does human intelligence stop being commoditised?</li><li>What happens when intelligence and ideas are commoditised?</li><li>How do we preserve ideas and intelligence?</li></ul><p>Is this part of that new world order where you own nothing, including your brain, and slowly move to the pre-frontal reacting, while the hippocampus outsources its processing? Letting that sink in means we are left with an outer shell while our inner self is outsourced, but maybe not today! In the meantime, there will be a period of market correction as we head from summer to autumn, and the hubris of all the listings and liquidity gets absorbed.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6bb6c004c184" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>