When your safety-first AI lab sounds more like a doomsday cult than a tech company, maybe the problem isn't the AI.
The Summary
- OpenAI allegedly spent $23M in compute to win a $1M math prize, possibly training on work from researchers who were using their tools — sending a clear message about IP risk in the LLM era
- An Anthropic engineer publicly states >10% probability of AI killing all humans within a decade, while admitting the company has no plan to prevent it and isn't on track to develop one
- The real story: two leading AI labs revealing completely different pathologies, competitive desperation at OpenAI and apocalyptic groupthink at Anthropic
The Signal
OpenAI's Navier-Stokes play looks like garden-variety academic misconduct scaled to billion-dollar budgets. They burned $23 million in compute to claim a breakthrough on a Clay Mathematics Institute problem, potentially using models trained on the very work mathematicians Tristan Buckmaster and Levent Alpöge had been developing. Those researchers were using OpenAI and Anthropic tools for months. The prize was $1 million. The compute cost 23 times that.
The economics make no sense until you realize this wasn't about the money. It was about the headline. OpenAI needed a win that looked like AGI progress, something that could justify their valuation and distract from the fact that GPT's improvements are plateauing. So they optimized for spectacle, not science.
"The message here seems to be: Don't use Codex or ChatGPT unless you're OK with OpenAI stealing your work if it's of interest to them."
But the Anthropic situation is weirder and more troubling. Jacob Coxon quit publicly, claiming the company was becoming reckless despite internal consensus that LLMs could end civilization. Evan Hubinger, still at Anthropic, confirmed this worldview: greater than 10% chance AI kills everyone within ten years, no solution to alignment for superintelligence, not clearly on track to find one.
If you genuinely believe that, the rational responses are:
- Stop building the thing
- Work exclusively on alignment until you have a solution
- Sound every alarm you can find
What you don't do is keep shipping models while posting your P(doom) estimates on Twitter like you're sharing fantasy football stats. That's not whistleblowing. That's brand management for the apocalypse. It's a way to position yourself as the responsible player in a race you refuse to stop running.
Key contradictions in Anthropic's position:
- Claims to be safety-first while racing to ship Claude updates
- Admits no alignment plan exists but continues scaling anyway
- Engineers publicly state civilization-ending risk while staying employed there
The real tell is how both companies talk about these issues. OpenAI stays silent on the Navier-Stokes mess because admitting the truth destroys their credibility. Anthropic talks constantly about existential risk because it differentiates their brand. One company cheats at math contests. The other treats doomsday scenarios as marketing copy.
Neither behavior suggests these organizations are equipped to build the technology they're building. OpenAI is optimizing for hype cycles and competitive wins, willing to burn researcher trust for headlines. Anthropic has convinced itself it's in a movie where the fate of humanity rests on them beating everyone else to superintelligence, even though they admit they don't know how to make it safe.
The Implication
If you're using these tools for anything proprietary, assume it's training data. OpenAI has shown they'll optimize models on whatever flows through their systems if it gives them an edge. The license agreements won't protect you. The PR statements about data privacy won't either.
For anyone building agent systems or automation workflows, this matters practically. You need to assume anything you feed into frontier LLMs is potentially being used to train their next model. That means keeping sensitive IP, novel approaches, or competitive advantages away from these systems entirely. Use them for commodity tasks. Keep the edge cases and the breakthroughs local.
And if you're watching the AI safety debate, notice who's still cashing paychecks while predicting the apocalypse. Real concern looks like action. This looks like people who've found a comfortable way to feel important while doing exactly what they'd be doing anyway.