George Hotz (@realgeorgehotz) published a July 12th blog post that reads like an intervention from inside the AI builder class: LLMs are useful, coding agents are improving, self-driving cars are real progress, and the industry is still over-selling doom because doom helps justify frontier-lab valuations.
Hotz is a useful messenger because he has spent the last decade building around the same forces he is criticizing. After his 2007 to 2014 run as a security hacker, he wrote that his career has been devoted to AI through comma.ai and tinygrad. That arc matters. Hotz is not attacking AI from outside the machine. He is arguing that the machine is powerful enough to matter, and ordinary enough to be commoditized.
In the post, Hotz says he is "giddy" about AI progress and points to new LLMs, Tesla's self-driving work, video generation models, and coding agents. His concrete example is local and practical: last week, he says, he set up a Linux box running opencode on a local GLM-5.2 model and asked it to install tmux with his configuration. It worked well enough for Hotz to joke that the Year of the Linux Desktop had arrived.
Then he draws the line. Hotz objects to what he calls "negative valence hype" around a closing window, a permanent underclass, and the idea that people outside San Francisco will fall hopelessly behind. He also rejects the jump from LLMs as better autocomplete, compiler, or search to sweeping claims about machines controlling the future. His bet is blunt: the progress is real, the fear layer is a sales apparatus, and the labs trying to own the narrative may struggle to own the economics.
The value-capture argument
Hotz's post ties together a thread he has been writing for months. In a June 21st post, he argued that some AI doom rhetoric functions as valuation support because current products alone do not carry the implied future value. In a February 19th, 2025 post, he made the cleaner economic version of the claim: AI may deliver large social value without letting any single company capture a proportionate amount of profit.
The July 12th essay is the same thesis in founder language. Hotz writes that the core of anti-open-source arguments is fear of commodification. Safety and China appear in the industry debate, but Hotz's claim is that the economic incentive underneath is simpler: if open models and cheaper compute keep improving, the scarcity premium around frontier labs gets harder to defend.
That is the same bet embedded in the tiny corp's May 24th, 2023 funding post. Hotz wrote then that the company had raised $5.1 million, described the business as a computer company, and stated the goal as commoditizing the petaflop. He argued that AI chip companies had underestimated the software layer, and that tinygrad would start by building a framework, runtime, and driver for AMD chips rather than taping out custom silicon. The investment details beyond the amount were not disclosed in the post.
Tinygrad's current site still presents that same operating theory. The project describes itself as a simple neural-network framework that reduces networks into three operation types, sells tinybox deep-learning machines, and says tinygrad is used in openpilot to run the driving model on a Snapdragon 845 GPU, replacing Qualcomm's SNPE. The site's FAQ says the framework has a PyTorch-like API, remains less stable while in alpha, and aims to leave alpha when it can reproduce a common set of papers on one Nvidia GPU 2x faster than PyTorch.
The hardware side makes the economics concrete. Tinygrad lists tinybox red v2 and green v2 Blackwell machines as shipping, with an exabox configuration available for preorder for 2027. The exabox spec is deliberately extreme: about one exaflop, 720 RDNA5 GPUs, 25,920 GB of GPU RAM, and a 600 kW power supply, according to the site. Hotz's argument against frontier-lab value capture is not theoretical punditry. He is trying to sell the picks and shovels for an AI market where compute gets cheaper, frameworks get smaller, and fewer developers need to rent their future from a hyperscaler.
Comma.ai is the older proof point
The other anchor in Hotz's post is comma.ai, the car-AI company he started before tinygrad. Comma.ai now sells comma four, which the company describes as an AI upgrade for existing cars. Its site says the device can run openpilot and supports lane centering, adaptive cruise, lane changing, dashcam recording, over-the-air updates, and 360-degree vision. Comma.ai also says comma four works on 325-plus car models from 27 brands.
Those numbers are company-reported, but they explain why Hotz's critique lands differently from a generic anti-hype post. Comma.ai's model has always been uncomfortable for incumbents: take hardware people already own, attach a device, ship software, collect driving data, and improve. That is a commoditization story too. The product sits in active driver assistance, not full autonomy, and the company frames setup in plain consumer terms: buy it, plug it in, engage.
Hotz's July 12th post also lands in a software market already learning the cost of AI-generated output. He concedes that he may have been too harsh about models not being able to program and now says programming is changing. He says he gets some boost from models, while warning that they can increase cognitive fatigue and that much of the generated software remains poor. That split tracks what engineering teams are seeing in production: AI agents can increase output, while validation becomes the new bottleneck. RuntimeWire reported in June on Greptile's work around AI pull request spam, where the important question is less whether code can be generated than whether teams can trust, review, and merge it.
Hotz is making a founder's version of that argument at market scale. LLMs are tools. Coding agents are tools. Better local models are tools. Tools change work, and the best ones usually become cheaper, more available, and less mystical over time. That path is good for builders and uncomfortable for companies priced as if they will control the bottleneck indefinitely.
The San Francisco pressure machine
Hotz's sharper cultural complaint is about geography and status. He rejects the idea that being outside San Francisco, away from the right rooms and parties, means missing a once-only AI window. The industry has heard that pitch before. It is useful for recruiting, fundraising, conference density, and insider mythology. It also makes founders feel late before they have had time to build anything durable.
His own setup undercuts the pitch. The example in the post is not a private frontier model inside a lab. It is a local GLM-5.2 model, running on a Linux box, controlled through an open-source coding agent. The work may be small, but the direction is the point: more capability is moving toward developers' desks rather than staying sealed inside a few model providers.
That does not mean frontier labs are irrelevant. Hotz's post does not deny that AI creates value, and it does not claim open-source models have already erased the lead of the largest labs. His case is narrower and harder to dismiss: value creation and value capture are different questions. If the underlying curve is driven by compute, software practice, open models, and general progress across the stack, the largest labs may keep shipping impressive systems while still facing margin pressure and weaker defensibility than their valuations imply.
For Hotz, that is the founder bet behind tinygrad and the long-running comma.ai pattern. Build closer to the metal. Push capability out of closed clouds and into machines people can own. Treat models as another layer in the computer revolution, rather than a priesthood around which founders must reorganize their lives. The final line of Hotz's post is almost deliberately anti-grandiose: "AI is the continuation of the computer revolution." Coming from a founder selling open AI software, driver-assistance hardware, and deep-learning computers, it is also a strategy statement. If AI is a continuation, the winners are unlikely to be decided by the loudest claims about destiny. They will be decided by who turns scarce systems into cheap tools first.