Silicon Valley has started punishing the AI doomers
This week's existential-warning cycle did something the earlier ones didn't: it gave the industry's biggest names permission to mock the messengers out loud.
An ex-Anthropic researcher's warning that AI builders are "gambling with our lives" drew open ridicule from Nvidia's CEO and an investment fund head this week — and one customer reaction that cost Anthropic actual deployment. Jacob Coxon, 27, who worked at OpenAI before joining its chief rival, resigned on Tuesday saying people building AI believe the technology could destroy humanity, that "these will soon be superhuman systems that can hack anything." Anthropic team lead Evan Hubinger echoed him on X: "We really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade." The response from the money side was not awe. Jensen Huang — who has called the extinction scenario "complete nonsense" — reportedly dismissed Coxon's claims in front of a conference crowd in San Francisco, while Altimeter Capital's Brad Gerstner called them "ridiculous hyperbole" and Hugging Face's Clément Delangue compared quoting Coxon to "asking your AC guy about climate change."
The commercial consequence is the part that lands. Grindr CEO George Arison called the warnings "dangerous" and evidence of "an anti-civilisational worldview at Anthropic," and said he told some of his engineers to stop using Anthropic's technology — while offering the theory that hurts most: that doom-saying is "a great way to gin up more investor support," since the only way to justify a $965 billion valuation is to claim the AI takes every industry. The shift matters more than the insults. Safety pessimism was the field's shared marketing language for two years — the half-of-entry-level-jobs line, the pacing essays, the auditor pledges we tracked this week in Amodei's pacing plan puts outside auditors inside Anthropic. It has now crossed into being a procurement liability, with a vendor-loss story attached and a sitting president saying he has "no" extinction concerns, only a China race. Watch for the correction, not the argument: both Gerstner and Delangue went soft on the idea within a day of Amodei publishing something actionable.
A single navigation model now drives a humanoid, a quadruped, a drone and a wheeled robot with no per-body tuning — and it learned almost none of it from real robots. Light Origins released LightNav-0 this week with open weights, code and a technical report: a Real2Sim2Real data engine turns 2,000+ internet-sourced real-world scenes into reusable simulated worlds, generating 4,000+ hours of aligned vision-language-action experience, then three post-training stages (embodied-reasoning mid-training, supervised fine-tuning, online RL) convert that into behaviour. The model ranks first in all ten monocular navigation comparisons and holds first in four when broader multi-camera methods are included, and transfers zero-shot across four physical embodiments. Two details are worth more than the leaderboard: reasoning is expressed as two image-space point tokens per decision (object point plus affordance point), which lifted mean success rate 8.4 points and SPL 5.7 without adding depth, localization or a map; and the ablation found adding new environments beats adding model capacity or squeezing more episodes out of the same scenes. That is the open question for the whole Physical AI wave — we covered the demonstration-learning version of it in Skild's S1 learns brand-new robot tasks from a single video — no retraining, and LightNav-0's answer is that environment diversity, not scale of any single axis, is what generalizes. The company shipped two companion systems in the same stretch: LightParkour, which distills walking plus three contact-heavy skills into one 50 Hz onboard policy, and Light REACT, which infers its own damage state from recent history rather than a fault label — upright behaviour after injury rose from about 42% to 76% once the policy was aligned on when to use each skill.
The most concrete outcome of the doomer week is legislative, and it is not from the White House. Vermont Senator Bernie Sanders co-sponsored the Ban Artificial Superintelligence Act this month, which would impose a temporary pause on advanced AI development — "When scientists tell you there is a chance that it could have a cataclysmic impact on humanity, you've got to be a moron not to say, slow it down." The bill is a long shot; the industry's own framing of it as a duopoly-protection play is the tell that it now registers as a competitive variable.
What to watch: whether Anthropic's next raise prices the safety brand as an asset or a discount.
Is a lab that tells you its product might end the world easier or harder to buy from? Tell us in the comments.
