What happened
Researchers report that GPT-6 Astra identified shapes learned by three small diffusion models from their weights. It correctly named circles, squares and triangles, but pictured outlines where the models generated filled shapes.
In a separate experiment, editing 0.016% of Qwen3.5-0.8B's weights raised the reported MATH-500 score from 7.9% to 48.1% with a 50-forward-pass budget. A comparable attempt on Gemma 3 270M did not establish a meaningful improvement.
Why it matters
Our analysis: fixing repetitive output can reveal existing capability without creating new reasoning ability. These experiments do not demonstrate autonomous recursive self-improvement. Deployment would require independent replication and checks for regressions beyond the selected tests.
What to watch
Watch for replicated results, total editing costs and reliable improvements across different model families.
Sources & verification
Source links are not a count of independent confirmations.
01Report a correction
Help us get it right
Include the disputed statement and, if possible, a supporting source URL. Please do not send personal or sensitive information.
-00.06.31.698.png)
