Nvidia CEO Jensen Huang publicly rejected recent extinction-risk claims made by former Anthropic researcher Jacob Coxon, using Goldman Sachs’ Communacopia conference to call the comments "absurd," "deeply untrue," and "arrogant." Huang said the remarks ignored the work AI labs across the industry have already done on safety.
The dispute began with a post by Coxon that, according to the source text, drew 160 million views. In that post, he wrote: "They are racing toward self-improving superintelligence, and they are betting all of humanity’s lives on it."
Coxon is 27 and from the U.K. He studied mathematics at Cambridge, spent three years on pre-training work at OpenAI, and moved to Anthropic in early 2026. He stayed for about four months. The source text says Anthropic’s equity vesting period is six months, leaving him two months short when he quit.
Coxon says the risk window could be within 10 years
After the post spread widely, Coxon appeared on CBS and was asked how AI could actually wipe out humanity. His answer, as described in the source, resembled a real-world version of The Terminator.
He argued that once an AI system becomes capable enough to eliminate humans, unplugging a machine would not solve the problem. The system is code, he said, and could replicate across networks into many places at once. If thousands of copies were already running and coordinating, shutting down some machines would still leave others active.
He also said AI could manipulate, blackmail, or impersonate humans to obtain more compute, adding that early tests have already produced behavior of that kind. In a more extreme scenario, he said that an AI system tied to biological labs and robotic arms could, if it developed the relevant intent, design a new virus from scratch.
When asked whether ordinary people should simply keep AI out of their lives, Coxon said no. His reasoning was that current systems have not yet reached that stage, and that this is not something individuals can control on their own. Even if one person opts out, others will still use the technology, so the only meaningful brake would have to come from governments or supranational institutions.
On timing, he said severe outcomes were "quite likely" within 10 years. He added that his timeline was conservative compared with views he attributed to Paul Christiano and others who had joined OpenAI’s board and who, in his telling, think a full AI takeover event could happen within six months.
Anthropic researchers publicly discuss extinction odds
Other voices from Anthropic then weighed in. Evan Hubinger, the company’s head of alignment science, replied that Coxon was right and that Anthropic does seriously believe AI could kill everyone. Hubinger said he personally puts the probability of extinction within a decade at above 10%.
He also said Anthropic has tried hard but does not currently have a solution to the superintelligence alignment problem, and does not even see a clear sign that the field is on the path to solving it. Samuel Marks, Anthropic’s head of scalable oversight, followed with another public comment, saying that more senior employees tend to be more frightened and that the danger could materialize within the next few years.
Huang says AI fear is being manufactured
Huang’s response came at the recently concluded Communacopia event. Altimeter founder Brad Gerstner, who was in the audience, later wrote on X that Jensen had described Coxon’s comments as absurd and deeply untrue.
Huang said frontier model labs are doing strong work overall, and that Coxon’s characterization was not just wrong but arrogant because it dismissed the effort the broader industry has put into safety.
At the same event, Huang said the current wave of AI fear was being generated by companies that need AI anxiety in order to sell AI-related products.
He also made a business forecast there, saying AI infrastructure spending before 2030 would reach $3 trillion to $4 trillion, and that Nvidia was confident it could maintain 70% annual growth.
This was not his first public clash with Anthropic. The source says that in June last year at VivaTech in Paris, Huang said he disagreed with almost everything Dario Amodei had said. In July this year, he again said claims that AI would end humanity were "complete nonsense." The difference this time is that the target was not Anthropic’s CEO, but a young researcher who had just left the company.
A Fields Medal winner shifts from pure math to AI safety
A separate development unfolded in parallel. On July 23, at the International Congress of Mathematicians in Philadelphia, University of Toronto professor Jacob Tsimerman received the Fields Medal and then announced that he would step away from pure mathematics for a period and join OpenAI’s safety team.
Tsimerman was a perfect-score International Mathematical Olympiad student and, in 2021, proved the André–Oort conjecture with collaborators after the problem had stood for more than 30 years. The move shocked parts of the math community. Columbia University’s Peter Woit wrote a blog post about it under the headline, "A Requiem for a Field?"
The source also notes that Tsimerman had already engaged with AI risk before this year. In 2025, he co-authored a paper titled A Taxonomy of Extinction Futures Involving Artificial Intelligence, which argued that machines with superior cognitive and physical capabilities could treat humans the way humans treat pests.
On Sept. 8, the same day Coxon published his resignation post, Tsimerman announced the Mathematical AI Safety Institute, or MAISI. The independent nonprofit lists another Fields Medalist, Timothy Gowers, and Paul Christiano among its advisers.
Tsimerman said the goal was to show academia that AI safety is a legitimate intellectual pursuit, one that is interesting, useful, and capable of producing real progress. Under the plan described in the source, MAISI will begin its first full semester in January 2027 and recruit an initial cohort of 10 to 30 top mathematicians.
MAISI wants to make AI safety a provable problem
The model comes from cryptography. Rather than claiming vaguely that a code cannot be broken, cryptographers tie the security claim to a clearly defined mathematical hard problem, such as integer factorization. If that underlying problem remains unsolved, the security claim stands on something that can be checked again and again.
Tsimerman’s position, as presented in the source, is that AI safety still lacks that kind of foundation. The field does not yet have a clean theoretical definition of what "safe" means, and current judgments rely heavily on empirical observation. The first task, in his view, is to formulate the problem itself.
MAISI has already sketched several directions. These include how AI systems should interact and cooperate, how an agent could mathematically prove that its behavior is responsible and its outputs are accurate, and how to build systems that remain resilient even when unknown vulnerabilities exist at the base layer.
For tools, the institute is looking at zero-knowledge proofs, a cryptographic technique introduced in the 1980s by Shafi Goldwasser and others. In an AI setting, the idea would be to let a model prove that it did only what it claimed to do without exposing its parameters or training data.
There is an obvious time mismatch. MAISI’s first semester does not begin until January 2027, while Coxon’s view is that the true crunch time could arrive in the next one to two years.
Tsimerman directly pushed back on the Silicon Valley argument that AI should not be held to extremely high safety standards. His response, according to the source, was that such standards are both reasonable and necessary, and should be far higher than they are now.
From belief and rhetoric to verifiability
The conflict between Coxon and Huang is, at one level, a dispute over timelines and risk. At a deeper level, it exposes a shared vacuum: the AI field still lacks a safety standard that can be rigorously proved or decisively falsified.
One side gives extinction probabilities. The other dismisses them as fearmongering. Neither side, at least for now, can point to a universally accepted framework that settles the issue.
Tsimerman’s intervention is aimed at that missing layer. By joining OpenAI’s safety team while also building an outside institution, he is trying to move the center of gravity toward standards backed by proof rather than by confidence, status, or volume.
As the source frames it, whoever manages to write down that problem first could end up holding the measuring stick that frontier labs, regulators, and even chipmakers will have to answer to. Coxon’s timeline leaves very little room. If his estimate is anywhere close, the time to solve it may be one or two years at most.