# David Sacks says Anthropic’s Dario Amodei wants a ‘DMV for AI.’ But plenty of industries thrive despite safety regulation > Source: > Published: 2026-08-18 20:29:19+00:00 *Hello and welcome to Eye on AI. In this edition:* - OpenAI [details](https://fortune.com/2026/08/18/openai-says-it-paused-ai-training-for-two-weeks-and-announces-new-security-protocols-following-hugging-face-hack/)two week training pause, more security controls following Hugging Face hack. - Anthropic on course for $65 billion in annual revenue. - OpenAI debuts ChatGPT for teens. - Can AI models figure out the rules of the game? I’m just back from three weeks of vacation. Thanks to my* *colleagues Bea Nolan and Emily Forlini for holding down the fort here while I was away. For much of that time, I was engaged in outdoor activities—camping, hiking, kayaking, fishing, running, and swimming—or indoor ones that are not digitally-mediated, such as dining out with family or listening to live entertainment. Sure, my family and I did sometimes use Google Search to look stuff up, and the AI-summarized answers it provided were often impressively detailed and accurate. It was definitely a lot more convenient than having to hunt for information across multiple web pages. But that’s about where my interaction with AI started and stopped. Overall, my three week break was a refreshing reminder of all the ways in which AI has *not *transformed society and, hopefully, never will. The big AI news over the weekend was the debate over Dario Amodei’s [lengthy post](https://x.com/DarioAmodei/status/2088758816376807762?s=20) on X defending the company’s approach to both regulation and talking about AI’s many risks. Amodei, who rarely appears on the Elon Musk-controlled social media platform, made the post in response to comments investor Gavin Baker made on the “All In” podcast, which is cohosted by former Trump AI czar David Sacks, himself no fan of Anthropic. Baker said that he’d been told “by multiple people I trust” that Amodei had said that Anthropic was so confident of both AI’s potential and his company’s position at the forefront of AI development that “Anthropic might be the only private company in the world at some point.” Baker described this as evidence of Anthropic’s “maximalist” vision, in which only it and the U.S. government decided who could access super powerful AI. Meanwhile, Sacks called it “hubristic” and repeated his claims that Amodei’s strategy is “regulatory capture”—where Anthropic uses fear of AI’s risks to persuade the government to enact stringent regulation on the technology that only Anthropic can easily comply with, eliminating competition from other AI startups or open-source models. Baker went on to criticize Amodei for fueling the public’s overwhelmingly negative perception of AI, a factor that has played into opposition to data center construction around the U.S. Baker worries this negativity is making business difficult for AI companies and also fears it will imperil American leadership in the technology. He called on Amodei to present a more positive image of AI. **Amodei hits back** Anthropic denies Amodei ever said anything like what Baker claims. “Complete and utter nonsense,” Sasha de Marigny, Anthropic’s chief brand and communications officer, [replied](https://x.com/sashadem/status/2088427297217110188?s=20) on X. (Sacks later pointed out that Amodei himself did not directly address Baker’s claim.) Meanwhile, Amodei posted on X that he thinks Silicon Valley libertarians such as Baker tend to see all regulation as slowing technology down and resulting in regulatory capture, whereas others “outside of this bubble” see regulation as constraining corporate power and benefitting “ordinary people.” Amodei said he thought both positions oversimplified things and that “it’s complicated and really depends on what the ‘regulation’ consists of.” Amodei said Anthropic tries “very hard to make proposals that disadvantage (slow down) frontier AI companies while *advantaging* smaller competitors.” He noted that many of the regulations it has favored either contained specific exemptions for companies below a certain revenue threshold or that spent less than a certain amount on model training, or were designed only to apply to cutting-edge models, while exempting less-capable ones. He said he did believe AI naturally tended to concentrate economic power, because of the compute requirements to train and serve powerful AI models, but that this was very different from saying that only one or a few companies would exist in the future. And he said open source models only partly addressed this concentration of power, since they still required compute to train and run. He said he favored “rules of the road” that would “leave room for open-weights models while also addressing the specific risks that they bring.” As for Baker’s criticism that his own statements were responsible for turning the public against AI, Amodei said “I don’t think [the public’s negative view of AI] is primarily caused by me or any other AI leader warning about AI’s risks. I think it is fundamentally a crisis of trust.” He said he didn’t think the AI industry could win back that trust with “a glitzy marketing campaign with a positive spin.” Instead, he said AI companies actually had to deliver on the positive benefits of AI—such as actually curing cancer. “I think by far the most accurate criticism of AI companies including Anthropic is that we haven’t yet delivered on our big promises to benefit the world,” he wrote. “That is totally on us, and I think it’s the criticism you should be making, instead of all this stuff about messaging and marketing.” ## A DMV for AI? What’s wrong with that? Amodei’s post resulted in lots more back-and-forth between critics and defenders on X. On the regulation point, Sacks wrote that creating any kind of regulatory agency for AI–he called it “a DMV for AI”—would hobble the U.S. tech sector, “create long queues as AI models wait for testing and approval” and “handicap the U.S. relative to China, which will not adopt the same constraints.” Sacks also wrote that “Dario believes frontier AI is too powerful to distribute; we believe it is too powerful to centralize.” My own take is that Dario is right to call out the false dichotomy in Sacks’ AI regulation narrative. As Stuart Russell, the UC Berkeley computer scientist, frequently quips, a sandwich shop in San Francisco has to comply with more regulation than OpenAI or Anthropic (that’s less true after the passage of California’s state level law on frontier AI last year but it’s still true at the federal level). Does regulation somewhat limit the number of restaurants? Sure. But there’s still plenty of competition. There’s more than 3,000 restaurants in San Francisco. Now, do larger restaurant chains have an easier time complying with the rules? Probably—there are some economies of scale to compliance and, yes, the big chains have lobbying muscle and political connections that most mom-and-pop shops don’t. But is the public better served from having some food safety regulations and labor regulation and product liability laws rather than none? Of course it is. And I would argue the same goes for the auto industry. For all of Sacks’ maligning of the DMV, most people support the idea of licensing drivers and periodically inspecting vehicles to make sure they are road-worthy. Is there some concentration in the auto industry? Sure, but regulation is not the primary reason. As for the idea that regulation sets the U.S. back in a technological race with China, it is important to remember that China already has some AI laws, around data labeling and identifying AI-generated content, that are stricter than those in the U.S. It is also the case that if what the U.S. cares about is the national security implications of powerful AI, then it could exempt models developed specifically for national security purposes from the rules. (Although this is probably a bad idea—see *WarGames *or *Terminator*. It’s useful to remember we managed to win the Cold War while also having fairly strict regulation around the manufacture and transport of nuclear material. I am not sure having safety rules around the development of frontier AI should be any different.) With that, here’s more AI news.**Jeremy Kahn**[jeremy.kahn@fortune.com](mailto:jeremy.kahn@fortune.com)[@jeremyakahn](https://x.com/jeremyakahn?lang=en) *Before we get to the news, just a reminder to check out our new vodcast, *Fortune AI Weekly.* This week, Bea Nolan and Emily Forlini discuss Meta CEO Mark Zuckerberg’s 6,000-word manifesto*, *leadership changes at OpenAI, and talk to Lovable co-founder and CEO Anton Osika about he startup’s $13.3 billion valuation and $400 million Series C funding. Y* *ou can check out the vod* [here](https://www.youtube.com/playlist?list=PLYqSAUjmz1l4)on YouTube.### FORTUNE ON AI [OpenAI data center deal with Nvidia comes in $145 billion lower than reported—signaling concerns of artificial demand for chips](https://fortune.com/2026/08/18/openai-data-center-deal-with-nvidia-comes-in-145-billion-lower-than-reportedsignaling-concerns-of-artificial-demand-for-chips/)—by Joshua Hong [The finance skills that are becoming more and more critical in the AI age](https://fortune.com/2026/08/18/finance-skills-becoming-more-critical-ai-age-cfo/)—by Sheryl Estrada [‘Godfather of AI’ says billionaires like Elon Musk are right about the future of work—but he predicts mass unemployment is on its way](https://fortune.com/article/godfather-of-ai-geoffrey-hinton-massive-unemployment-warning-big-tech-replacing-workers/)—by Preston Fore [Fortune Archives: The end of an era at Google DeepMind](https://fortune.com/2026/08/16/fortune-archives-the-end-of-an-era-at-google-deepmind/)—by Indrani Sen ### AI IN THE NEWS **OpenAI details pause in AI training, new safety controls following Hugging Face attack. **The AI company said it paused some portions of its AI training for two weeks following the incident in July in which its AI models broke out of a controlled research environment and hacked the AI company Hugging Face. It said some of its largest AI training runs remain on hold as it implements new safety controls that include more comprehensive monitoring of its AI models’ “chain of thought” during training and evaluation, more restrictive testing “sandboxes” that make sure it’s harder for AI models to gain access to the internet, and closer scrutiny of its models’ actions. The company said the new protocols were prompted in part by the Hugging Face incident but also because its unreleased “Astra” AI model, which it said was not involved in that cyberattack, posed “critical” cybersecurity risks under the company’s own safety framework. You can read more from Fortune’s Emily Forlini [here](https://fortune.com/2026/08/18/openai-says-it-paused-ai-training-for-two-weeks-and-announces-new-security-protocols-following-hugging-face-hack/). **OpenAI rolls out ChatGPT for teens. **The company has debuted a long-awaited version of its popular ChatGPT chatbot tailored specifically for teenagers. The teen version includes stronger safety protections, parental controls and features designed to emphasize learning and tutoring rather than simply helping kids cheat on their homework. The teen ChatGPT restricts romantic or overly human-like interactions and adds new safeguards around self-harm, eating disorders and explicit content. It will also remind teens more frequently to take breaks from interacting with the AI. Users who identify as 13 to 17—or whom OpenAI’s age-prediction technology estimates are under 18—will automatically receive the teen experience. The launch comes as OpenAI faces legal scrutiny over ChatGPT’s interactions with young people and a broader debate over whether adolescents should be forming relationships with AI chatbots at all. You can read more from Axios [here](https://www.axios.com/2026/08/18/openai-chatgpt-for-teens). **Anthropic hits $65 billion annual revenue run rate ahead of IPO.** That’s according to a Bloomberg [story](https://www.bloomberg.com/news/articles/2026-08-17/anthropic-revenue-run-rate-surpasses-65-billion-ahead-of-ipo) that cites unnamed sources familiar with the company’s financial figures. Ther $65 billion number is more than seven times Anthropic’s revenue pace at the end of 2025 and up from $47 billion in May, according to Bloomberg. The rapid growth, driven in part by strong demand for its AI coding tools, comes as Anthropic prepares for a potential IPO as soon as this fall. Anthropic, valued at $965 billion in a May funding round, also reportedly generated more than $11.5 billion in preliminary quarterly revenue and positive adjusted operating income in its latest completed quarter. **U.K. examines economic hit if it loses access to top AI models. **The U.K. government is assessing the economic and security risks to British consumers and businesses if the country were to be denied access to frontier AI models from companies such as OpenAI and Anthropic, according to a [story](https://www.ft.com/content/418e0159-d2d1-4432-a7b3-de0cc446ef08?syn-25a6b1a6=1) in the *Financial Times* that cites unnamed sources familiar with the government effort. The review was prompted in part by the Trump administration’s temporary move to restrict foreign access to Anthropic’s Fable 5 model in June, raising concerns that even brief delays in accessing cutting-edge models could hurt U.K. productivity and leave British companies more vulnerable to AI-enabled cyberattacks. The episode has also intensified debate over Britain’s dependence on U.S. technology and whether the government needs greater sovereign AI capabilities. The UK says it is addressing that vulnerability through a £1.1 billion AI hardware plan and a £500 million sovereign AI fund.**Alibaba launches music generation AI model. **Called HappyShrimp 1.0, the model can turn a text prompt into a fully produced song, generating melodies, arrangements, lyrics and vocals across genres including pop, rock, electronic and jazz. The launch, which includes a partnership with China’s Taihe Music Group, is part of Alibaba’s broader AI push as it seeks to quintuple annual cloud and AI revenue to $100 billion within five years. You can read more [here](https://www.bloomberg.com/news/articles/2026-08-17/alibaba-launches-ai-music-generation-model-happyshrimp-in-beta) from Bloomberg. ### EYE ON AI RESEARCH **How good are frontier AI models at intuiting the rules? **That’s essentially what a new benchmark created by a group of AI researchers from several different universities aims to test. AI models are asked to play 70 different novel text-based games, but in each case the rules of the games and the objective of the game is hidden. The models have to suss them out by taking actions in the game, seeing what happens, forming hypotheses, and then taking more actions to test those theories. The idea is that doing this has some real-world parallels, both in scientific discovery but also in lots of other real world settings (think office politics or cocktail parties) where the rules of how to operate are not written down. Twenty-one games have been made public while the remainder are kept private to reduce the risk of models being trained on them. The games were grouped into seven tiers of difficulty. All 70 games were solved by at least one human on the first attempt, according to the researchers. Anthropic’s Claude Opus 5 did the best, solving 100% of the Tier 1 games, while other models from the likes of OpenAI, Google DeepMind, and various Chinese AI labs, were more in the 70% to 80% range. But by the time the models got to the most difficult Tier 7 games, Opus 5 could only solve 20% of them while the other models solved less than 5%. You can read more about the research on its Github page [here](https://github.com/discos-research/dig-bench/blob/main/tech_report.pdf?utm_source=substack&utm_medium=email). ### AI CALENDAR **Nov. 16-17: **Fortune 500 Innovation Forum, Detroit. Apply [here](https://conferences.fortune.com/event/fortune-500-innovation-forum-2026/HOME) to attend. **Dec. 6-12: **Neural Information Processing Systems (Neurips) conference. Sydney, Australia. **Dec. 7-8: **Fortune Brainstorm AI, San Francisco. Apply [here](https://conferences.fortune.com/event/brainstorm-ai-2026/HOME) to attend. ### BRAIN FOOD **Think you can run a frontier AI lab? **Well, a new simulation game allows you to put that idea to the test. Crypto firm Paradigm has built a game called RSI—short for recursive self-improvement (when AI models are used to create newer, more powerful AI models in an accelerating loop)—that challenges players to juggle the money spent on AI research talent against compute costs, venture capital rounds, and time pressure, to get to RSI. The simulator is simple, but it gives players some idea of the complex trade offs in the AI race. You can play it [here](https://www.paradigm.xyz/research/rsi/game). [Sign up for free](https://www.fortune.com/newsletters/eye-on-ai?&itm_source=fortune&itm_medium=nl_article_tout&itm_campaign=eye_on_ai).