OpenAI's reported $1.25T valuation meets Altman's self-improvement claim

A $1.25 trillion valuation for a company that, a few years ago, was a research lab with an awkward governance structure should make us ask what exactly investors think they are buying. OpenAI has reportedly pulled in $11 billion, with the round possibly nearly doubling in the coming weeks. I find the number less interesting than the story attached to it, because the same week, Sam Altman is telling us that the era of recursive self-improvement has begun.

Altman points to AI agents that already rack up days of work and models that find new results on their own. OpenAI has just released 377 new math findings, which is unsettling the field. I think this matters, but I'd separate two claims that are getting blurred. One is that models are producing genuinely useful results in narrow domains like mathematics, which looks real and worth taking seriously. The other is that this amounts to a self-accelerating loop, which is a much bigger statement, and unsurprisingly critics are piling on from many directions. A company raising money at a trillion-dollar-plus valuation has an obvious incentive to describe its technology in the most transformative terms available. That doesn't make Altman wrong. It does mean the claim deserves the same scrutiny we'd give any founder's pitch.

The sharpest counterweight comes from Yann LeCun. He and Randall Balestriero have published a paper using JEPAs to prove and improve the Platonic Representation Hypothesis, and the implication is that simply scaling up may not be the only path forward. If they're right, the capital being poured into ever-larger training runs is at least partly a bet on one architectural philosophy among several. By the way, this is where Jim Fan's new VFlash fits in. It's a vision model built to parse the visual world in parallel, outputting entire results at once rather than step by step, aimed at robotics and Physical AI. That is a different design instinct from the autoregressive models dominating the headlines, and it suggests the field is less settled than the funding numbers imply.

The money keeps flowing regardless. Thinking Machines Labs has launched Inkling 2 Small, a model designed to automate AI training tasks, while reportedly seeking around $5 billion. Notice the theme: models built to improve how models get built. If that sounds like Altman's argument in a more modest, shippable form, I'd say that's about right.

Meanwhile, the physical and legal constraints are catching up. Texas's AI boom is colliding with shifting energy policy, and I suspect power, not parameters, will decide where the next wave of compute lands. On the regulatory side, Hamburg's data protection commissioner has published a practical case study on deploying LLMs in line with data protection rules. For anyone building with these systems in Europe, that is more useful than most of the week's grand pronouncements.

Demis Hassabis told Reid Hoffman that society is not ready for what's coming and that a more diverse group of experts needs a seat at the table. If even the optimists agree on that, the open question is who actually gets invited, and whether anyone with a trillion-dollar valuation will make room.