A less discussed angle on Dean Bell's opinion regarding open-weight models, entering a new era of mcarthyism.

# The Rise of Digital McCarthyism Source: [https://russwilcoxdata.substack.com/p/the-rise-of-digital-mccarthyism](https://russwilcoxdata.substack.com/p/the-rise-of-digital-mccarthyism) Dean Ball, days into his new job as Head of Strategic Futures at OpenAI,[posted his assessment of Kimi K3](https://x.com/deanwball/status/2078133895766114412), the new model from the Chinese lab Moonshot AI\. Kimi is an open\-weight model, which means its inner workings are published for anyone to download, run, and modify; the newest version’s weights ship later this month\. It had just matched the best publicly available models of the first quarter\. The post made three claims\. Open\-weight models are “inherently decelerationist\.” A probable outcome of an open\-weight\-dominant world is “full AI communism,” a future of AI as public good and digital public infrastructure that strikes him as a “dystopian hellscape\.” And the United States government will at some point realize that its best strategy is to create large amounts of “regulatory risk” around the use of open\-weight Chinese models\. I have met Dean Ball\. We have disagreed before\. This time the disagreement is public, and so the response should be too\. Two of his sentences are true, and they deserve to be said plainly before the rest is examined\. Kimi K3 is hungry:[independent benchmarks](https://artificialanalysis.ai/articles/kimi-k3-achieves-3-in-the-artificial-analysis-intelligence-index-comparable-to-opus-4-8-and-gpt-5-5)put it near a dollar per task, roughly half the cost of the leading closed model but far from the bargain that earlier Chinese releases were\. And the question of what happens when powerful models circulate beyond any institution’s ability to recall them is a real question\. It deserves a serious answer, and it will get one\. Ball is not wrong about the model’s appetite or the governance problem, but he is wrong about everything he builds on top of them, and the shape of the wrongness matters more than the wrongness itself: not a personal failure, a frame doing exactly what frames do\. Let us begin with the word everything else in the post stands on: “inherently decelerationist\.” The field Ball now works in exists because of open diffusion\. The transformer architecture behind every modern model was a published paper\. PyTorch, the software most of the field builds with, is open source\. The benchmarks, the training techniques, the research commons that made OpenAI’s breakthroughs possible were open before the company closed the door behind itself\. Open weights are simply that same commons moved one level up the stack: what the published paper was to the last decade, the published model is to this one, and closed labs themselves now depend on open models to do reproducible safety research\. To call open weights inherently decelerationist is to stand at the top of a ladder and declare ladders inherently opposed to climbing\. The evidence for the opposite sits inside his own post\. He concedes that Kimi matches the best public models of the first quarter\. When DeepSeek forced the same comparison eighteen months ago, the result was a price war, a surge in infrastructure spending, and a capability sprint across every lab, including the one that now employs him\. An open model is no slow lane but the version of this technology that a city government can run on its own machines, in its own jurisdiction, without asking a San Francisco pricing page for permission\. So the word “decelerationist” is doing something other than measuring, and it helps to say precisely what\. There are two clocks in this debate\. One measures the frontier: on that clock, closed models lead open ones by somewhere between four and twelve months on standard benchmarks\. The other measures reach: who can run the technology, where, at what price, under whose control\. On the first clock Ball’s employer is winning; on the second, openness is the entire race\. And the second clock is where most people live\. A dollar a task is a rounding error to a laboratory and a locked door to a night\-school student, a rural clinic, a small\-town newsroom, and when the frontier stays behind a paywall, the working class pays the toll\. His mechanism concedes this: open weights deter capital spending, he argues, because no one will pay for closed models below the frontier\. That admission rewards a slow reading\. The closed\-lab capital structure depends on scarcity, on the frontier staying fenced\. In his usage, acceleration means revenue continuing to flow to institutions that sell access, and deceleration means capability reaching people without payment\. The word is not measuring the speed of artificial intelligence\. It is measuring the speed of OpenAI, and borrowing the vocabulary of progress to do it\. There is nothing inherent in any of this\. It is a business model wearing the costume of a law of nature\. The second claim does heavier lifting with weaker materials\. A public good, in the language of economics, is a shared thing that one person can use without using it up, and that no one can practically be fenced out of, the lighthouse being the textbook case\. Open weights are, as a matter of plain description, exactly that\. This is not an ideology but a property of the artifact, the same property that describes Linux, the open operating system running the servers of every frontier lab, and the protocol stack, the shared plumbing of the internet, on which OpenAI’s entire business sits\. The county agent system that once carried agricultural science into every rural county in America was public knowledge infrastructure too, and none of these things were communism\. Calling the category a “dystopian hellscape” is not an argument about the category\. The word was chosen to end the conversation before the category could be examined, offered between a false pair of alternatives, market product or state hellscape, the frame’s only two categories, with no shelf for most of the infrastructure the modern world runs on\. The strongest witnesses against the post are OpenAI’s own documents\. In May 2015, in[the email that founded the company](https://www.techemails.com/p/elon-musk-and-openai), Sam Altman proposed to Elon Musk “a Manhattan Project for AI” structured so that the technology would belong to the world through a nonprofit\. In August 2025, OpenAI released its first open\-weight models since GPT\-2, and[Altman explained the reasoning](https://www.cnbc.com/2025/08/18/openai-altman-china-ai.html): if OpenAI did not release open models, the world was going to be built mostly on Chinese ones\. And here inconsistency alone should not carry the verdict, because hypocrisy proves only that someone is wrong once, not which time\. What settles it is what the release demonstrated\. In May of this year,[RAND researchers reviewed](https://arxiv.org/pdf/2606.19890)every open\-weight model family above a high compute threshold released since January 2025, thirty\-seven in all, and coded them against the strictest evaluation standard in the literature: testing models with their safeguards stripped, retrained by simulated bad actors, against worst\-case misuse\. Exactly one family met the full standard\. It was GPT\-OSS\. Ball’s employer is not merely inconsistent with his post but the certified best practitioner of the responsible openness the post says cannot exist\. The company built the working alternative to his proposal, spent seven figures in computing hours stress\-testing it, and shipped it nearly a year before he arrived\. Which leaves the third claim, the one[Ball has since described](https://x.com/deanwball/status/2078619513575137330)as cold analysis rather than advocacy\. Let us take him at his word and read it as pure prediction: the government’s best move is to manufacture regulatory risk around Chinese open\-weight models\. The difficulty is that the government already examined this question and answered it in writing\. The administration’s own[AI Action Plan](https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf), published in July 2025, contains a section titled “Encourage Open\-Source and Open\-Weight AI” and directs the federal government to create a supportive environment for open models, citing their value for startups, research, and geostrategic leadership\. And Ball would know, because[he was the plan’s lead writer](https://80000hours.org/podcast/episodes/dean-ball-ai-policy-governance-white-house/)\. His prediction about what this administration will realize was falsified a year before he made it, by a document he himself drafted\. And even as pure analysis, the sentence refutes his first claim, because no one needs to manufacture uncertainty around a technology that is inherently losing\. Suppression is what you predict for a competitor you expect to win\. The predicted instrument is not a safety standard, not an evaluation regime, not anything addressed to actual harms, but manufactured ambiguity, and his own example names the mechanism: the Federal Reserve issuing warnings about potential backdoors in Chinese models\. A central bank publishing fear notices about software is uncertainty deployed as a chilling instrument, and the chill would benefit the closed incumbent labs, one of which now employs the analyst\. He rejects an outright ban, to his credit, as among the dumber ideas in the policy debate; what remains, by design, is fog\. And the reading is not mine alone: within days,[the White House’s own AI adviser publicly asked](https://www.axios.com/2026/07/19/pentagon-openai-dean-ball-trump-ai)whether the proposal amounted to regulatory capture for the benefit of Ball’s new employer, and by Sunday a senior Pentagon official was attacking him by name\. The reading has left the building\. This is not an accusation of bad faith, and warnings can be true even when the warner profits from them\. Nor is it a question of ability, because Ball knows the serious method as well as anyone alive: five months before the post,[he co\-authored a landmark paper](https://arxiv.org/abs/2601.11699)on rigorous third\-party auditing of AI companies, alongside Bengio and Bommasani, and he lists it on his own site\. The capacity for rigor is documented, which is exactly why the post reads as the institution’s voice rather than the analyst’s limits\. The problem is that no mechanism inside these institutions lets the rest of us tell the warning from the pitch: the same balance sheet funds both, the investors reward fear, and the discourse takes the shape of what the investors reward\. A strategist at a closed lab warning that open weights lead to communism is the bomb shape producing bomb\-shaped discourse: the Manhattan Project reflex, bred into American AI thinking from the beginning, to concentrate a powerful force, race for it, and treat whatever cannot be concentrated as a threat\. None of this makes the seed innocent\. The seed is the opposite reflex, the one China chose: distribute the force through existing terrain rather than trying to contain it\. That architecture has costs, and the serious answer starts by stating them at full strength\. Channeling a force through terrain requires the terrain be legible to whoever digs the channels, and legibility at that scale is surveillance\. Release is irreversible: a model strong enough to matter, once published, can never be recalled\. The stripping of safeguards is industrial\. One popular platform hosts more than eight thousand models modified specifically to remove refusal behavior, and the research finds the gap between a model with its safeguards and the same model stripped grows as the models grow stronger\. These facts are real, and they do not become less real because an incumbent invokes them in self\-interest\. What they describe, though, is a decentralized, ungovernable commons, thousands of variants in thousands of hands\. Ball half\-sees this himself: in the same thread he compares open\-weight users to the hill peoples of James Scott’s anthropology, communities defined by their escape from every state that tried to govern them, and then, two points later, predicts the same tools end in state\-provided communism\. His two futures, the ungovernable and the monopolized, cannot both be where this architecture leads, and his own citation votes for the first\. So does the empirical record\. Meanwhile the serious literature, including the RAND work above, responds to these exact dangers not with suppression but with proportional evaluation: specify the threat, measure the harm release adds to a world that already holds libraries, test the worst case\.[The government’s review](https://www.ntia.gov/sites/default/files/publications/ntia-ai-open-model-report.pdf)reached the same method, and so did the academic consensus, and so, in practice, did his own employer\. The debate Ball believes he is starting already exists, it is rigorous, and every serious participant is on the other side of it\. That frame is the real subject here, larger than one post and one strategist\. It is powerful and it is blind in a specific direction\. It cannot see distribution as anything but leakage, cannot see infrastructure as anything but a rival’s plot, cannot see a public good except as the road to a hellscape\. So when a strong open model arrives, the frame does not analyze it\. It reaches for the oldest word in its drawer\. But America is also the civilization that built the county agent, the land\-grant university, the internet’s shared plumbing, the open commons its own frontier labs are built on, and it built them in the same century it built the bomb\. It invented the concentrated project and the distributed channel, and its best decades ran both at once\. A single keeper of the frontier is also a single point of failure, one door for every thief and every tyrant to try, and the sturdiest defense against a dangerous technology has never been the locked room but the nation that holds its own tools\. The question Kimi K3 poses is not whether to suppress the seed\. It is whether America still remembers that it planted the first one\. #### Discussion about this post ### Ready for more?