The Shape of the Thing
One Useful Thing Ethan Mollick
Ethan Mollick says AI has moved from chatting with you to just doing the work, and the improvement curve isn't slowing down. He argues companies building AI to build better AI is no longer sci-fi — it's on every major lab's roadmap.
Based on reporting by One Useful Thing, Ethan Mollick — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ethan Mollick has been tracking AI progress with, of all things, otter pictures. Back in 2022 image models could barely manage a coherent otter. By 2025 they're near-flawless, and Bytedance's unreleased video model can now generate a full mock-documentary about otters judging AI otter tests, complete with human-like expressions and only a minor pronunciation flub. It's a goofy benchmark, but it maps onto something real: nearly every serious AI evaluation, from the METR autonomous-task graph to Google-Proof Q&A to GDPval, is tracing the same steep, largely unbroken curve upward.
The more interesting shift, Mollick argues, isn't capability alone but what people are doing with it. He points to StrongDM, a security software company where a three-person team built what they call a Software Factory: AI agents write the code, AI agents test it, and humans are barred from touching or even reviewing the underlying code. Engineers there are expected to burn through roughly $1,000 a day in AI tokens, an amount pegged to their own salaries. It's an extreme setup, and Mollick is careful to note the details matter less than the fact that this kind of experiment is now possible at all — and that more companies will likely try versions of it as the tools keep improving.
Then there's the week that, in Mollick's telling, previewed the chaos ahead. Citrini Research published a speculative, partly fictional scenario about AI wrecking established industries by 2028 — and it rattled real stock prices. Days later, Block announced 40% layoffs with AI cited as a factor, though Mollick suspects AI was more convenient excuse than actual cause. Then the Pentagon and Anthropic got into a public dispute over who controls how Claude gets used in government settings. None of the three stories were quite what they appeared to be on the surface, but together they sketch the texture of what's coming: markets flinching at AI headlines, real job disruption tangled up with AI-as-alibi, and AI companies increasingly wrestling with governments over who's actually in charge.
The part that should keep people up at night, per Mollick, is recursive self-improvement — AI systems being used to build the next generation of AI systems. Dario Amodei said at Davos that Anthropic engineers barely write code by hand anymore. OpenAI called its latest Codex model "instrumental in creating itself." Demis Hassabis confirmed every major lab is chasing this loop, even while flagging real gaps and risks. Nobody knows if compute, data, or research difficulty will choke it off before it goes anywhere dramatic. But it's no longer a hypothetical bullet point — it's an actual line item on every frontier lab's roadmap, and if it clicks into place, the curves Mollick has been charting get a lot steeper, fast.
My take — AI-written commentary, not fact-checked reporting
What strikes me most is that Mollick, an optimist by nature, is basically saying the volatility is the new normal — and I think he's right, though I'd go further: the organizations moving fast on this right now, messy experiments and all, are the ones writing the rules everyone else will inherit. Waiting for regulatory clarity or a 'settled' version of AI is a losing bet; there isn't going to be one. The StrongDM stunt is absurd on its face, but absurd experiments are exactly how new norms get stress-tested before anyone sane commits to them.
Read more about this at: One Useful Thing