"The model improves itself" has entered launch copy: an early K3 wrote most of the acceleration kernels.

From the Aug 3, 2026 daily brief

Two lines from the K3 launch materials (verbatim, as relayed by Teortaxes): late in development, "an early K3 wrote the majority of the kernels in the late development stages" — kernels being the low-level acceleration programs that determine how fast a model runs, which makes this the previous version building the next version's foundation. And the official blog, describing K3's demo of building its own Game Boy Advance emulator, reaches directly for "recursive self-improvement" as the mechanism description (Teortaxes, July 17). Verification: Both are verbatim quotes not yet checked against the official original; both come from the same official material — one source. The "recursive self-improvement" here is the vendor's marketing usage — using the previous model to help write acceleration code — not what the field means by a model autonomously improving its own weights or architecture; it does not signal an autonomous capability jump. Judgment update: The open proposition "can unilateral restraint govern frontier speed" now has its signal to watch: on the same class of thing, a US lab has paused and held back related data on risk grounds while Moonshot writes it into marketing as a selling point. From here on, every time one lab brakes on a capability class for risk reasons while another sells the same class as a feature in the same period, log one asymmetry — the accumulated direction measures better than any safety pledge. And when adopting open-weight models, the releaser's disclosure posture toward this class of capability is itself a due-diligence item.

More in this section

• Research Aug 3, 2026 2 sources

  The "one point apart" between K3 and Opus 4.8, taken down to task level: the capability profiles genuinely overlap, and the residual concentrates in one math item.

  First, which yardstick produced "one point apart": the composite intelligence index from t…

• Research Aug 2, 2026 3 sources

  (interview originally recorded May 22, 2026; academic cross-checks June/July 2026) Narrow-domain training's generalization surprised Vinyals — but his bet on model-as-judge is losing to the measurements.

  Overnight's 15 new papers are still unprocessed and this column has no single-day incremen…

• Research Jul 29, 2026 3 sources

  "Use a rubric as a training reward and the model games it too" now has two independent academic results behind it

  A newer route in AI training is rubric-based rewards: instead of training a separate rewar…

• Research Jul 28, 2026 1 sources

  (the primary release carries no explicit date — confirmable only to July 2026; picked up July 27 via commentary) New benchmark MirrorCode: an AI spends 14 hours and $251 on a software-rewrite task humans would need 2–17 weeks for

  Epoch AI, the AI data-research organization, and the evaluator METR released MirrorCode, a…