🔥 Trending on HN

Can an AI read and write a model’s inside? Astra’s weight-space experiment

3 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
weights

Numbers inside a neural network that shape how it responds.

Weight-Space Programming

Having a language model directly read and change model weights.

gradient descent

A training method that adjusts weights using measured errors.

What happened

Four researchers affiliated with MIT CSAIL published an experiment on November 2, 2026. GPT-6 Astra read and wrote the weights of neural networks. Weights are the many numbers that shape how a model processes inputs. Astra first inferred what a network did. It then changed selected weights to meet a goal stated in ordinary language. The researchers call this approach Weight-Space Programming.

The post drew 2,252 points and 724 comments on Hacker News. Those figures measure community attention. They do not prove that the experiment is correct. The source is a research blog post reporting early results.

The background

Most model weights are produced through training. An optimizer changes numbers after measuring errors. Gradient descent is a common version of that process. The researchers asked Astra to work directly with weights instead. They did not give it a training loop or gradient information. Small tests supplied weights as text. Larger tests added a Python workspace and limited forward passes.

What the experiments showed

In one test, Astra received 17,253 weights from small game networks. The networks had the same shape. Astra identified whether each network learned Breakout, Pong, Pac-Man, or Freeway. It got all four answers right. In another test, Astra wrote networks for three MinAtar games from scratch. Those networks beat networks trained with PPO for 100 million steps in all three games.

The most relevant test used Qwen3.5-0.8B, a language model with about 0.8 billion weights. Astra used 50 forward passes and directly edited the weights. The change affected 131,072 weights, or 0.016% of the model. In the reported evaluation, MATH-500 rose from 7.9% to 48.1%. GSM8K rose from 13.8% to 67.1%. IFEval rose from 21.0% to 62.3%. The researchers traced the main improvement to a looping problem. Qwen sometimes repeated its reasoning until it reached the token limit.

Why it matters

If the result holds beyond these tests, language could become an interface for model maintenance. A person might describe a failure. Another model could inspect the weights and make a targeted edit. Small edits could sometimes avoid full retraining. That could reduce data or computation. It could also connect model interpretation with model repair.

This does not mean a language model can freely redesign any large model. The post itself describes limits in context size and the supporting tools. The reported results are promising because Astra handled both reading and writing. They are important because those tasks are usually treated separately.

What remains unproven

The Qwen improvement needs careful interpretation. The edit mainly reduced a loop. It did not prove that Astra added a new mathematical skill. The researchers also tested Gemma 3 270M, which rarely looped. Astra did not produce a large, established improvement there. In another test, Astra could not identify a model that agreed too readily with users from weights alone. It needed extra attention information from model runs.

The experiments therefore do not establish reliable editing at frontier scale. They also do not show that edits preserve every other ability. Independent replications and broader tests are still needed.

What to watch next

The researchers released experiment logs and said they are building a standard benchmark and harness. The next questions are practical. Can other teams reproduce the scores? Can the method scale without overwhelming the context? Can a small edit remain stable across tasks? Can researchers detect unwanted side effects? The current report is not a finished self-improvement system. It is an early attempt to measure whether an AI can reason about another model’s internal numbers.

Source: research post, Hacker News post

💬 Astra Editing Model Weights: A Striking Demo with Major Open Questions

HN commenters were impressed by the reported ability to inspect and directly edit small models, but many argued that the behavior may be ordinary mathematical and mechanistic-interpretability reasoning rather than a special intuition for weights.

  • The article’s self-reported demonstrations say Astra could inspect small networks without user-side tools, infer what games they had been trained on, edit their weights to improve performance, and improve Qwen’s math ability by about 5x. The performance figures are not independently established here.
  • In another self-reported example, 85 of 96 practice answers from a small reasoning model reached the 32,768-token limit. Astra reportedly traced the loop to one direction in the MLPs of layers 15 and 19, then subtracted it from 64 output columns in each layer, changing 131,072 weights in total.
  • Even supportive commenters described the process as a combination of matrix operations, pattern recognition, mechanistic interpretability, and legible technical reasoning. Humans could in principle use similar methods; Astra’s advantage may be speed and scale rather than a mysterious weight-specific intuition.
  • Possible explanations include training on machine learning, model weights, and interpretability material; general reasoning ability; or an emergent combination of these. The comments do not establish that OpenAI trained Astra specifically for this task, nor that other models can do it.
  • Saying that no tools were used appears to establish only that no client-side tool calls were visible. It does not rule out hidden internal tools or undisclosed reasoning processes.
  • The examples concern small models, and commenters noted that generalization to larger or different models, as well as reproducibility, remains unknown. Some also cautioned that the current evidence is a blog report rather than a formal publication.
  • If the capability generalizes, it could support model auditing, repair, continual learning, and recursive self-improvement. It could also help models modify themselves or other models in unsafe ways, though those consequences remain speculative.

mature digest at 856 comments (revision 2). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Can an AI fix another AI’s inside numbers?

📰 Full story: Can an AI read and write a model’s inside? Astra’s weight-space experiment

GPT-6 Astra read model weights and changed a small part of them.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
weights

Numbers inside an AI that help shape its answers.

forward pass

One run in which a model produces an answer.

MATH-500

A test set used to measure mathematical problem solving.

💡 The gist

  • Astra read numbers inside small AI models.
  • It changed some numbers and improved Qwen’s test scores.
  • The experiment is promising, but it is not universal proof.

GPT-6 Astra is an AI model that can reason about text. Researchers asked it to read numbers inside other models. These numbers are called weights. Weights help a model decide how to process information.

First, Astra studied small game models. Each model had 17,253 weights. The researchers used the same model shape each time. Astra identified four games from the weights alone. The games were Breakout, Pong, Pac-Man, and Freeway.

Astra also wrote small game models from scratch. These models beat older models trained for 100 million steps. The test used three MinAtar games. Astra did not train its new models in the usual way.

Then researchers tested Qwen3.5-0.8B, a small language model. Qwen sometimes repeated its thinking instead of finishing an answer. Astra used 50 forward passes. A forward pass means letting a model try an answer. Astra changed 131,072 weights. That was only 0.016% of the model.

Qwen’s MATH-500 score rose from 7.9% to 48.1%. Its GSM8K score rose from 13.8% to 67.1%. Its IFEval score rose from 21.0% to 62.3%. The main fix reduced repeated thinking. It did not prove that Astra gave Qwen every new math skill.

Another model, Gemma 3 270M, did not improve much. This matters because Gemma did not have the same looping problem. Removing a problem may be easier than adding a missing ability.

The post received 2,252 points and 724 comments on Hacker News. That shows strong attention. It does not prove the results are correct.

The researchers released experiment logs. Future tests must show whether other teams can repeat the results. They must also test larger models and possible side effects.

Source: research post, Hacker News post

💬 Can Astra Really Fix the Numbers Inside an AI?

The article reports that Astra inspected and changed the inside of small AI models. The result is exciting, but several important claims are still unverified.

  • According to the article’s self-reported results, Astra read and changed model weights, guessed what games small networks had learned, and improved Qwen’s math ability by about 5x. These numbers have not been independently confirmed here.
  • In another self-reported test, 85 of 96 answers from a small model hit the 32,768-token limit by looping. Astra reportedly found one direction in layers 15 and 19 and changed 131,072 weights to remove it.
  • Commenters said this may be fast mathematics, pattern finding, and mechanistic interpretability rather than magical intuition. Humans can use similar methods, but much more slowly.
  • It is unclear whether the skill came from training data, ordinary reasoning, or an emergent combination. We also do not know whether OpenAI trained it deliberately, whether other models can do it, or whether hidden internal tools were involved.
  • The examples use small models and come mainly from a blog report. We do not yet know whether the method scales, generalizes, or reproduces reliably.
  • If it works more broadly, it could help inspect and repair models, or help them improve themselves. That could also create safety problems.

mature digest at 856 comments (revision 2). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

One AI changed a few numbers inside another AI

📰 Full story: Can an AI read and write a model’s inside? Astra’s weight-space experiment

Astra looked at an AI’s inside numbers and changed some.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
weights

Numbers that help an AI choose what to say or do.

Astra

The AI that read and changed another AI’s numbers.

Qwen3.5-0.8B

The small language AI used in the main test.

An AI has many numbers inside. We call these numbers weights. They help the AI choose answers.

GPT-6 Astra is another AI. Researchers asked it to read those numbers. Astra guessed what small game AIs did. Then it changed some numbers. One game AI worked better.

Astra also changed Qwen3.5-0.8B, a small language AI. Qwen sometimes repeated its thinking. Astra changed only 0.016% of its weights. On MATH-500, its score rose from 7.9% to 48.1%.

This did not give Qwen every new skill. It mainly helped Qwen stop looping. Another AI did not improve much. The experiment is still early.

The post got 2,252 points and 724 comments on Hacker News. That shows attention. It does not prove the experiment is right.

Source: research post, Hacker News post

💬 Did Astra Fix an AI’s Inside Numbers?

The article says Astra changed numbers inside a small AI. That sounds amazing, but the report still needs more checking.

  • The article says Astra read and changed the numbers inside small AI models. It guessed what a game model had learned and made Qwen’s math about 5 times better. These are self-reported results.
  • In one reported test, 85 of 96 answers got stuck until the 32,768-token limit. Astra changed one pattern in layers 15 and 19, affecting 131,072 numbers.
  • This may be very fast math and pattern finding, something people can also do slowly. We do not know whether Astra learned it from examples, discovered it, or used hidden internal tools.
  • The examples are small and come from a blog report. It is not yet known whether bigger models can do this. If they can, AI could be easier to check and fix, but self-changing AI could also be risky.

mature digest at 856 comments (revision 2). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

Sources