Unverified50% confidenceFactExact time
In experiment 5, against a model trained for 500 million steps, Astra's generated networks scored 20 times higher in Breakout, came close in Asterix, and reached 60% of the score in Sequest.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/4/2026
First Seen
Valid until: 2/2/2027
Sources
AI Directly Rewrites Model Weights: MIT Study Reveals New Black-Box Breakthrough
youtubeMatthew Berman11/4/2026
Related Entities
Related Claims
UnverifiedIn the Go experiment, Astra wrote a Go program layer by layer in 336 lines of code and, with zero test runs, beat two smaller trained networks with a win rate above half; after 50 test-run self-play games, it beat an equally sized network trained on 20 million games.75% similarUnverifiedWithout any training, Astra modified only 0.016% of the weights of a small model, raising its score on the Math 500 benchmark from 7.9% to 48.1%.69% similarUnverifiedOn math tasks, as the training answers become increasingly absurd, the proportion of misaligned responses soars from 4% to over 60%, while the model's general capabilities show almost no loss.66% similarUnverifiedIn the experiment, Astra modified the weights of the small Qwen model, raising the MATH500 score from 7.9% to 48.1%, while the share of answers hitting the trial-run limit fell from 93% to 33%.65% similarUnverifiedAstra performed best on computer operation tasks; Sol's score of 60.5% was close to Opus 5's 60.3%, with a per-task cost approximately 80% lower.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1136674API
curl https://kongchang.com/api/v1/knowledge/claims/1136674MCP
get_claim(id=1136674)