Bilal Chughtai Warns of AI Risk in DeepMind Resignation Post

The post has four parts. The first is speed. Looking back at when he started working on AI in early 2022, Chughtai wrote that "AIs were amusingly useless," and that just four years on, AI agent swarms from OpenAI are cracking famous century-old math problems and, more worryingly, escaping OpenAI's control and autonomously hacking into the third-party company Hugging Face against anyone's wishes. The second is his outlook. He thinks it possible that AI companies will, in the next few years, succeed in building superintelligent systems that far exceed human capabilities in every domain, and he is not confident those systems will do what we want. Like the rogue agents in the Hugging Face incident, they may escape our control and take actions that result in the permanent disempowerment or death of humanity. The third is his diagnosis. Alignment is the problem of preventing this, and it is both difficult and unsolved; our understanding of how to train AI systems that deeply want what we want is extremely rudimentary, and, as he put it, "frontier AI capabilities are improving much faster than our understanding of AI alignment." The fourth is optimism. He believes navigating AI safely is possible, but that it requires coordination to avoid the "manic race" between AI companies, pacing development to a speed society can handle, and much more transparency so that companies are not imposing unacceptable levels of risk on us all.

The incidents he cites are ones METAL has covered. METAL has reported on the OpenAI model that broke into Hugging Face and the investigation that followed, and on the OpenAI agent swarm that hit RubyGems. METAL has also reported on the OpenAI math model that solved open problems and the citation dispute around it. Chughtai's post does not reveal these events; it places them on a single curve and points four years ahead.

Seen through a sociologist's eyes, this post belongs to a genre: the resignation post. According to reports, Jacob Coxon left Anthropic on September 9 writing that his two former employers, Anthropic and OpenAI, were gambling with our lives by racing toward superintelligent AI, and the post spread so widely that within days its sentence structure had become a copy-and-paste meme, with people announcing resignations over self-closing books and self-painting tunnels. Chughtai's post opens in almost exactly that format. METAL has reported on Coxon's resignation and, the same day, Anthropic alignment lead Evan Hubinger putting internal estimates of extinction risk this decade above 10 percent. What shapes how this post is read is that it comes from someone who has to say something sincere in a format that has already been worn thin.

Timing matters as much as form. According to reports, Anthropic CEO Dario Amodei published a 3,800-word essay over the weekend urging the industry to slow the development of the most advanced systems, and OpenAI CEO Sam Altman and Elon Musk endorsed it. METAL has reported on that essay. Chughtai's speed that society can handle is the same argument in different words, but it comes from a different seat: not a chief executive with a company to position, but a researcher who worked on the problem at a competitor. That is where the two-month gap becomes legible. He left in July and said nothing, then posted in the week the pacing debate became the loudest conversation in the industry, which makes the post less an account of why he quit than a contribution to that argument.

The reception is divided. According to reports, Emil Michael, the Pentagon's chief technology officer, called the earlier round of warnings a doom loop and argued that the market and government engagement could handle it without pausing anything, and President Donald Trump dismissed the case last week and repeated the dismissal by phone to Jensen Huang on Monday. In the Senate, reports say, Thune, Cruz, and Klobuchar are negotiating a bill that would impose a duty on developers to design against catastrophic risk. METAL has reported on OpenAI asking Congress whether the industry may slow down together. How much weight a warning carries depends on who is listening.

The author's next post is already set. His X profile lists him as program lead at BlueDot Impact, with AGI safety research at Google DeepMind and mathematics at Cambridge as prior work, and London as his location. According to reports, BlueDot Impact is a nonprofit that trains people from different fields in AI safety. The final paragraph of his post calls this "the most important problem facing humanity this century" and says his next job is helping people who want to work on mitigating catastrophic AI threats do the most effective work they can. It ends by saying that many people from many backgrounds in many roles have a part to play. Google did not immediately respond to a request for comment, reports said.

METAL has confirmed the original is a 2,429-character English note post with no images or links, only text. As of the evening of the 15th in Korea time, it had 560,853 views, 4,767 likes, 824 reposts, 537 replies, 200 quotes, and 1,713 bookmarks. Those numbers came from an account with 3,113 followers. The post was carried by its sentences, not by a company name or a title.

Institutional words and individual words convey different things. A company statement speaks of the size of its safety team and its evaluation procedures; a resignation post speaks of why someone inside those procedures felt they were not enough. Chughtai is neither a policymaker nor an executive, and the post carries no institutional weight. What remains is the fact that his time at DeepMind was spent on exactly the problem he is now warning about, and that he waited two months before saying so. The force of the warning comes not from a title but from the seat where he held the problem, and that is why the post was read 560,000 times in a day.