Unverified50% confidenceFactExact time
Four MIT researchers (led by Rikia Takehi and Jaden Teo) published a paper proposing the concept of 'weight-space programming,' i.e., handing the raw weights of a small network to GPT-6 Astra and letting it read, repair, or write new weights from scratch.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/4/2026
First Seen
Valid until: 2/2/2027
Sources
Related Entities
Related Claims
UnverifiedIn a study titled "Sparks of Weightspace Programming," the MIT team had OpenAI's GPT-6 Astra directly read and modify another model's weights without training or gradient descent.73% similarUnverifiedGPT-6 Astra's ability to directly read model weights without tools currently covers only small networks, with an upper scale limit of about 170,000 weights.63% similarUnverifiedIn the optimization workflow of web-edit 0.1a1, GPT-6 Astra was responsible for diagnosing performance bottlenecks, Claude Opus 5.5 was responsible for rewriting the code, and Simon Willison was responsible for reviewing, merging and releasing the results.61% similarUnverifiedThe researchers gave 17,253 weights to GPT-6 Astra, and all four networks were correctly identified as their corresponding games.61% similarUnverifiedMIT found that frontier models can directly hand-write neural networks, and the Breakout network that GPT-6.2 Astra wrote from scratch scored 20 times that of the PPU-trained version.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1135777API
curl https://kongchang.com/api/v1/knowledge/claims/1135777MCP
get_claim(id=1135777)