Three things crossed my desk today and they all point at the same question. Everybody building this stuff is starting to get nervous, but nobody's willing to be the one who slams on the brakes first. Let's get into it.
OpenAI Asks Congress If Slowing Down Is Even Legal
OpenAI went to Congress this week and asked a question that tells you a lot about where things stand. They wanted to know if the big AI labs coordinating to slow down frontier development would count as illegal collusion under the Sherman Act. Sam Altman told his own staff he's open to easing off the gas, but only if everybody else eases off too.
Here's the thing nobody's saying out loud. When a company that's been sprinting for three years straight suddenly asks a lawyer if it's allowed to stop sprinting, that tells you the sprinting part is scaring even the people doing it. OpenAI's chief scientist basically said the same thing in a blog post, calling voluntary slowdowns the likely near term reality until everybody agrees on where the safety line actually is.
My take: this is what it looks like when an industry realizes it's stuck in a game of chicken nobody actually wants to be playing. Antitrust law was built to stop companies from teaming up to screw over customers on price. It was never built for a handful of labs trying to figure out whether to slow down before something breaks. Congress ought to sort this out fast, because right now the law is basically silent on it, and silence just means everybody keeps their foot on the pedal by default.
China's Labs Are Strip Mining Claude
Anthropic dropped a threat report showing just how much Chinese AI companies have been leaning on Claude to build their own models. We're talking Alibaba, Moonshot, DeepSeek, and Xiaomi, all caught running what Anthropic calls distillation campaigns. Alibaba alone generated something like 151 million exchanges with Claude between May and July, using thousands of fake accounts, all aimed at making its Qwen models smarter off Claude's back.
Distillation isn't some exotic hack. It's just training a cheaper model to mimic a more expensive one by feeding it tons of that model's output. Companies do it to themselves on purpose sometimes. The difference here is scale and the fact that nobody asked permission. Nearly 190 million exchanges across five separate campaigns is not a curious researcher poking around. That's an industrial operation.
My take: this was always going to happen the second frontier models got good enough to teach other models for free. You can build the best model in the world and somebody with less money and fewer rules will find a way to copy your homework at scale. The labs pouring billions into training runs need to start planning for this the same way you'd plan for theft, because that's basically what it is.
Mistral Just Cashed Europe's Biggest Tech Check Ever
Mistral closed a 3 billion euro raise this week, pushing its valuation up to 21 billion euros. That's the largest private tech fundraise in European history, and it comes right as the European Commission is hinting it might loosen up rules around big corporate mergers to keep European AI competitive with the US and China.
Europe has spent the last few years watching American and Chinese labs eat its lunch on AI while its own regulators mostly played defense. This raise is Europe finally putting real money behind the idea that it wants a seat at the table instead of just writing rules for everybody else's table.
My take: money alone doesn't win this race, but you can't even enter without it, so this is a real step. Whether Mistral turns 21 billion euros into a model that actually competes with the top of the pack is a different question, and the answer to that one is still wide open. I'll believe Europe's back in the game when I see it in the benchmarks, not the funding rounds.
That's the roundup. Slow down or don't, get copied or don't, raise the money or don't. Same story every week, just a new set of names attached to it.