Follow-up: newer models are much better at 'writing weights' directly

After yesterday's MIT 'weight-space programming' result, Google researcher Alex Krentsel ran a similar test across model generations on a tiny game-playing network (about 20,000 weights). Older models scored near zero; the newest ones produced networks scoring around 18 to 22 times a trained baseline. It's one researcher's own run on a toy task, and open models struggled.

Why it mattersStill research, not a tool. It's a sign that the newest models can do kinds of engineering that older ones couldn't, so retest tasks a model failed six months ago.

Curated by Sara Simeone as part of our human editorial coverage. The original sources are linked below; curation is distinct from a Hype or Fact? assessment.

SHARE THISPost on X ↗Share on LinkedIn ↗Copy link

How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.