Honnibal and Goedecke Dispute MIT's Weight-Space Programming Claims

Four authors listed as MIT CSAIL researchers published a paper claiming that a model they call GPT-6 Astra can read and edit a neural network's raw weights without gradient descent, and two engineers published rebuttals dated Nov. 3 saying the headline result depends on tools the paper's framing leaves out.

The paper, "Sparks of Weight-Space Programming," is by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. The authors call it "very early results" and describe a general-purpose model that "reads and writes a network's weights directly." They said they are building a benchmark and a harness for the capability.

What the Qwen result claims

The result drawing attention is experiment 8. Astra edited 131,072 weights in Qwen3.5-0.8B, which is 0.016 percent of the model, after running it on up to 50 sets of practice problems. At 50 passes, MATH-500 rose from 7.9 to 48.1 percent, GSM8K from 13.8 to 67.1 percent and IFEval from 21.0 to 62.3 percent. Hits on the token limit fell from 93 percent to 33 percent.

The paper says Astra traced the looping behaviour to one direction written by the MLPs of layers 15 and 19, then subtracted that direction from 64 output columns in each layer.

Where the two rebuttals land

Matthew Honnibal, the author of spaCy, wrote that the paper presents several different methods as one. In his count only experiments 1, 2 and 4 read raw weights directly, on networks of 17,000 to 172,000 weights. Every experiment on a real language model used Python, and the Qwen result also ran the model on practice problems.

He also said the game-network experiment starts from weights set to zero, so there is nothing to read. Astra beat a PPO network trained for 100 million steps only on Breakout, and lost to one trained for 500 million steps on Seaquest and Asterix.

Sean Goedecke wrote that the Qwen model was not generally bad at mathematics. It looped after 32,000 tokens, and he said the original model solved every task under that length. In his reading Astra wrote four pairs of looping and non-looping texts, measured which weights contributed to the loop and trimmed them. He said that matches steering and abliteration methods, the family of tools covered in the Heretic project, and that "a smart human ML researcher could perform all of these tasks."

Honnibal's own verdict was less dismissive: "Figuring out how to do this productively is not nothing, and AI models of a year or two ago would have failed at all these tasks."

Experiment	Tools allowed	Reported result
1. Name the game a PPO policy plays	Read-only, 17,253 weights	Correct on all four games
4. Fix a car-on-ramp policy	No tools, one attempt	0 to 1,024 of 1,024 episodes succeed, 3.5% of weights edited
5. Write a Breakout policy from zero	No tools, one reply	Score 517.4 against 25.3 for PPO at 500M steps
8. Reduce looping in Qwen3.5-0.8B	Python, up to 50 forward passes	MATH-500 7.9% to 48.1%

1. Astra, written from zero517.4 points
2. PPO, 500M steps25.3 points
3. PPO, 100M steps15.8 points

Source: Takehi et al., Sparks of Weight-Space Programming, accessed 2026-11-03

What failed in the paper's own results

The authors report their misses. Astra could not improve Gemma 3 270M, which rarely loops: MATH-500 moved from 6.8 percent to 7.7 percent at most.

The paper's own summary of the Gemma result is that editing "could remove an obstacle that hid an ability, but it could not add an ability the model lacked."

Simon Willison, who writes a weblog on models, ran the first experiment on his own laptop. He fed raw weights for four game policies to Qwen 3.8 27B, a 17 GB Q4_K_M file, with no tools. It named Breakout after 52 minutes and 25 seconds, Pong after 1 hour 4 minutes, and described Pac-Man after 46 minutes, but guessed wrong on Freeway. He said he is "not ready to declare that the age of source code is over." The Terminal has covered the Qwen3.8 27B release separately.

The paper's logs are on GitHub in the RikiyaT/lm_understand-logs repository, and the authors released an edited Qwen model on Hugging Face. They did not give a date for the benchmark.