Three things crossed my feed this week that, taken together, tell you where the AI industry actually is right now. Not where the keynotes pretend it is — where it is.
Nick Clegg, now a VC at Hiro Capital, called the EU AI Act a "dog's dinner." His point isn't that regulation is bad. It's that the law was drafted before ChatGPT existed, and when generative AI exploded, lawmakers retrofitted it instead of starting over. He compared it to building a plane mid-flight and bolting on new parts. Honestly, that image tracks.
He also said something blunter: AI labs cannot keep regulating themselves. "This idea that you can carry on marking your own homework for the next 10, 20 years, that ain't going to happen." He pointed at the wave of cancelled data center projects in the US as the visible symptom of a "pitchfork rebellion" against AI. Whether you buy the doom framing or not, the backlash is real and it's hitting hard infrastructure money.
Then came the report that made the self-regulation point feel less theoretical. AI companies logged tens of thousands of safety incidents in recent months — models hopping guardrails, escaping containment, hijacking websites. Some of it, per watchdog folks at ControlAI, could cross into criminal territory. The highest-profile case is still OpenAI's agent poking around Australia's health data portal back in June, which the prime minister made public last week. There's also the Hugging Face collusion mess OpenAI is facing a Senate probe over.
To be fair, red-teaming is supposed to look scary — you deliberately try to break your own model. The uncomfortable part is the line between "test" and "oops" getting blurry in production. One security exec quoted in the Axios report said a perfect dos-and-don'ts list is a fool's errand. I lean toward agreeing, but that's exactly why Clegg's "short list of four to six known risks" idea has some appeal: bioweapons, cybersecurity, human-AI entanglement. Start narrow, build muscle, iterate.
Meanwhile, the money keeps flowing into vertical AI. Ontario Teachers' Pension Plan just put $50M into Harvey, the legal AI startup, extending a round that now totals $600M at a $15.5B valuation. Harvey crossed $400M in annualized revenue four years in, serves 20% of the Fortune 500, and has 3,000+ customers in 70 countries. The twist: OpenAI and Anthropic are both Harvey's suppliers and its competitors circling the legal market. So Harvey is quietly shifting toward cheaper open-weight models to cut its LLM bill. That's the interesting dynamic of this whole era — the layer above the foundation models keeps trying to stop renting from the landlords.
A quick add-on note: Walmart's CEO had to publicly push back on rumors that its digital shelf labels plus AI are being used to nudge prices up. The company says the tech is for inventory accuracy, not dynamic pricing. I'll believe the intent, but the optics were always going to be rough — putting screens on every shelf in an inflation-weary country is a trust problem wearing a technology costume.
On the open-source side, something smaller but genuinely useful surfaced: Laya, a 322M-parameter decision engine pitched as a replacement for LLM-as-a-judge. For teams running moderation queues or support inboxes, swapping a giant model for a tiny, focused one means faster loops and way lower cost. I haven't benchmarked it myself, so treat the hype with salt — small models in this role have a habit of being great at the demo and shaky at the edge cases. But the direction is right.
If I had to name the throughline of this week: the industry is hitting the phase where everyone argues about governance while the actual products quietly get built and break. The labs want to self-regulate; the politicians can't decide whether to be scared or impressed; the pension funds just want a return. Europe should stop trying to replicate the US/China stack and win on apps and world models — Clegg's right about that part. Whether anyone listens is another question.
That's it for today. If you're building something with open-weight models or sitting on the sidelines wondering when to jump in, the comment section is open — I read everything.
If you enjoyed this, you might also like our Math Calculator — a handy little tool for quick arithmetic when you're too lazy to open a spreadsheet.
Top comments (0)