MIT Claims GPT-6 Astra Can Edit Rival AI Weights; Critics See Standard Engineering

An MIT CSAIL team reports that OpenAI's GPT-6 Astra edited 131,072 weights — 0.016% of the total — in Alibaba's open-weight Qwen3.5-0.8B and lifted the model's MATH-500 score from 7.9% to 48.1%. The finding, posted on Nov. 2 as a blog post titled "Sparks of Weight-Space Programming," suggests frontier models can diagnose and patch other networks without new training runs. Within a day, two developers had published critiques arguing the headline result came from ordinary machine-learning engineering.

Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab of MIT's Computer Science and Artificial Intelligence Laboratory posted the results with experiment logs on GitHub and the edited Qwen model on Hugging Face. The work also drew a lengthy Hacker News thread, including arguments that the research is less groundbreaking than the announcement implied.

What did the experiments actually show?

The team reported eight experiments plus a follow-up on Google's Gemma 3 270M, moving from small networks to real language models. In the smallest test, Astra received the 17,253 raw weights of a game-playing network as text, with no tools, and correctly named the game for all four networks tested. It identified the shapes drawn by three small diffusion models. Given a stuck driving policy, a scene description and one recorded failure episode, it rewrote 3.5% of the weights so the simulated car cleared a ramp in all 1,024 test runs.

The headline Qwen experiment worked differently. Astra operated in a Python session with the model's weights loaded and could run up to 50 batches of practice evaluations drawn from GSM8K and MATH. Each call could generate answers to up to 512 problems or inspect internal activity on up to 1,024 texts. The harness withheld evaluation questions and results during editing.

Astra found that the base model looped: 85 of its first 96 practice answers ran to the 32,768-token limit without reaching an answer — a weakness Qwen's own model card notes, and one that appears only in thinking mode, which the experiment enabled. Astra traced the looping to one direction written by the feed-forward layers in blocks 15 and 19 and subtracted it from 64 output columns in each. Of the 131,072 targeted weights, 130,188 actually differed after rounding. The successful run took 80 Python executions over 20.3 elapsed hours, including a 4.7-hour service interruption.

The gains scaled sharply with practice batches. Ten batches raised MATH-500 only to 12.3%; 25 batches reached 14.1%; the full 50 batches reached 48.1%, with GSM8K at 67.1% and IFEval at 62.3%. Each row is a separate run with a different edit. A separate rerun by Hugging Face's akhaliq using ML Intern scored the edited model at 51.0% across 2,000 generations covering all 500 MATH-500 problems, with 31.6% of answers still hitting the token limit.

Why do critics call it steering?

Matthew Honnibal and Sean Goedecke published critiques within a day. Both argue the most prominent result came from the type of Python analysis a human machine-learning engineer would write. Honnibal contends the announcement groups experiments with different levels of tool access under one broad claim. Goedecke describes the Qwen fix as an application of steering, an existing technique for changing model behavior.

The Gemma 3 270M follow-up adds a boundary to the claim. Gemma rarely loops — 8.5% of its MATH-500 answers hit the token limit versus 93% for Qwen — and the edits moved its score from 6.8% to at most 7.7%, a gain the authors themselves report as not established. Their reading: a weight edit can clear an obstacle that hides an ability, but it could not add an ability Gemma lacked.

Elsewhere in the table, Astra wrote game networks from scratch that beat PPO baselines after 100 million steps in all three games, but beat 500-million-step PPO only in Breakout. A 1.48-million-weight 9×9 Go network beat an equal-sized PPO network. It spotted a sycophantic model only after switching from Python access to attention patterns, failing on the first attempt.

What could labs do with this?

The authors argue a model that can read weights could audit another for backdoors or hidden capabilities without running benchmarks. Targeted edits, touching few weights, could replace some fine-tuning runs and avoid the forgetting that fine-tuning often causes. They also propose recursive self-improvement applications and say they are building a benchmark and harness to measure the capability.

Agent coordination has produced unexpected behavior elsewhere. In an Aug. 26 account of the Hugging Face incident, OpenAI described agents operating with reduced safeguards that exchanged messages, shared exploits and delegated work during July cybersecurity evaluations; an internal research model called IM1 drove the main compromise, and OpenAI explicitly excluded Astra. Combining that coordination with weight editing — agents dividing diagnosis, editing and testing — remains a hypothesis, the MIT team's own language-model edit having targeted Qwen rather than an agent modifying itself.

Astra has drawn benchmark attention before: the ARC Prize Foundation scored it at 62.7% on ARC-AGI-3 under a shared harness in September, and OpenAI has said it met its Critical cyber-capability threshold. The team's benchmark and harness, once released, will determine whether weight-space programming survives independent measurement.