What Happens When AI Works Directly With Another AI’s Weights?
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What Happens When AI Works Directly With Another AI’s Weights? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

MIT CSAIL researchers report that GPT-6 Astra could infer functions from neural-network weights, create small networks without training and edit an existing model. A separate set of experiments reported mixed outcomes, including a narrow benchmark improvement and failed legal-answer scoring; safety and reproducibility questions remain.

Researchers at MIT CSAIL say they used GPT-6 Astra to inspect, create and edit neural-network weights directly, without training the networks through gradient descent. Their November 2 report describes a possible new way to analyze and repair open models, while experiments by the report’s author produced mixed, limited results and highlight risks if the same methods remove safety-related behavior.

The MIT team supplied Astra with the numerical weights of small neural networks and asked it to infer what they did. In one test, the researchers say it identified the games played by four networks with the same architecture. In another, it found copying heads in three small language models, according to the source article’s account of the MIT post. These are reported experimental results, not evidence that the method can interpret arbitrary large models.

The team also asked Astra to write networks from zero weights, with no training loop. The source says the resulting arcade-game networks beat conventionally trained networks in all three tested games, whose comparison networks had trained for 100 million steps. The supplied material does not give the full game-by-game scores or enough evaluation detail to judge how broadly that result applies.

For an editing test, the researchers changed a small portion of Qwen3.5-0.8B’s weights to remove one direction associated with looping. The source reports that the model’s MATH-500 score rose from 7.9% to 48.1%. In separate experiments described by Thorsten Meyer, an edit helped a local model complete more legal-answer attempts, but its legal score remained zero; another single-matrix edit improved a narrow JSON test from 49 to 70 correct out of 120. The article says a second AI review also found faults missed by green tests.

At a glance
reportWhen: MIT post dated November 2, 2026; source…
The developmentMIT CSAIL researchers reported experiments in which an AI model read and modified other neural networks’ weights directly, without a training loop.
Thorsten Meyer AI
MIT CSAIL · 2 Nov 2026

Infographic 01 · What weight space programming showed

Read, write, edit: what GPT-6 Astra did with raw weights

A blog post with complete public logs, not a peer-reviewed paper. Small networks, one research group, one run per setting.

The number to remember48.1%

MATH-500 score of the small open model Qwen3.5-0.8B after GPT-6 Astra removed one loop direction from 0.016% of its weights. Before: 7.9%. A Hugging Face re-run of the edited model: 51.0%.

Three things an AI did with raw weights

No training loop, no gradient descent

  • Read: named the game behind 4 of 4 small game networks (17,253 weights each) and found every copying head in three small language models with their first and last layers removed.
  • Wrote: arcade-game networks from all-zero weights that beat networks trained for 100 million steps in all three games. Breakout: 517.4 against 25.3 for 500 million steps.
  • Fixed: a car network that never backed up: 0 successes in 256 episodes before, 1,024 of 1,024 after, with 3.5% of the weights changed.
Limit

Fixes what is broken, adds nothing new

Gemma 3 270M rarely loops (8.5% of answers hit the limit). No edit moved its MATH-500 score by more than about a point.

Limit

Some things stay invisible

Honest model or yes-man? From the weights alone: not found in 4 of 4 attempts. With a view of the running model: found in 3 of 3 pairs.

Sources: MIT CSAIL blog post, 2 Nov 2026 · logs on GitHub · Hugging Face re-run.
Powered by Thorsten Meyer AI

Direct Edits Could Change Model Audits

If the reported approach proves reproducible, it could give researchers another way to examine and alter open-model weights without running a full training process. That may help identify particular behaviors, diagnose failures or make targeted changes when data and computing resources are limited. It could also supplement, rather than replace, testing a model on inputs and observing its outputs.

The same capacity raises a safety concern. The source article warns that edits could strip out safety training as well as unwanted behavior. A change that improves a narrow benchmark does not establish that a model is safer, more reliable or better overall. The reported legal test illustrates the distinction: more completed answers did not produce a higher legal score. Any deployment would need broader evaluations and checks for side effects.

Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024 Model) - Metallic Black

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

As an affiliate, we earn on qualifying purchases.

From Weight Inspection to Editing

Neural-network weights are numerical values that shape how signals move through a model. Conventional training adjusts those values repeatedly using gradient descent. In weight-space programming, a separate AI is given weight data and asked to reason about it or change it directly, rather than using that training loop. The phrase describes an approach, not a claim that the internal logic of a model becomes fully readable.

The source places the work alongside earlier mechanistic-interpretability research. It says a 2021 Anthropic team showed that some algorithms in small language models could be detected directly from weights, but that work required specialist methods. The MIT report’s stated ambition is for language models to reason about weights themselves. The experiments described here are small-scale and do not establish that the same capability works on the larger systems used by consumers.

“It fixes what is broken; it cannot add what is missing.”

— Thorsten Meyer, describing the editing result

Limits of the Reported Tests

The supplied source is a secondary article summarizing an MIT CSAIL post, and it does not include the full experimental record. It says the researchers published prompts, replies and evaluations on GitHub, but the material provided here does not include those files or enough methodological detail to independently assess the results. It is also unclear how robust the reported benchmark gains are across repeated runs, different model sizes or broader tasks.

The author’s own experiments are described without complete setup details, model configurations or full evaluation results. One experiment reportedly produced no useful improvement, while another improved a narrow JSON test. The legal-answer test’s score stayed at zero. It remains unclear whether the edits caused trade-offs on other tasks, how well the method can distinguish a useful behavior from a correlated one, and what safeguards would prevent edits that weaken a model’s safety behavior.

Reproduction and Broader Safety Checks

The next step is to examine and reproduce the MIT team’s published prompts and evaluations, then test the approach on additional models and tasks. Replication should measure not just the targeted score but also regressions, repeatability and safety behavior. The source material does not identify a scheduled follow-up study, a release timeline or a deployment plan.

For now, the results are best read as an early report of a technique with promising demonstrations and significant limits. Whether direct weight editing can become a dependable tool for model auditing or repair will depend on independent verification and evidence that targeted changes do not create unseen failures.

Key Questions

What does it mean for an AI to work directly with another AI’s weights?

It means a model is given another neural network’s numerical weights and asked to interpret or change them directly, rather than updating them through conventional training.

What did the MIT researchers report?

The source says GPT-6 Astra identified functions of small networks, created arcade-game networks without a training loop and edited a small model. These findings are reported experiments and do not show that the method works on every model or task.

Did the weight edit make the model better overall?

That has not been established. The reported MATH-500 result improved, but the source also describes mixed outcomes, including a legal-answer test that remained at a score of zero. Broader testing would be needed to assess overall performance.

Could direct weight editing affect AI safety?

Potentially. The source warns that edits could remove safety-related behavior as well as unwanted behavior. The reported material does not establish how reliably such risks can be detected or prevented.

Source: ThorstenMeyerAI.com

You May Also Like

Beyond #1: What Gemini 4 Carbon’s One-Point AI Ranking Leaves Unanswered

Gemini 4 Carbon scores 59 on Artificial Analysis’s index, narrowly leading rivals. Its cost, uneven benchmark results and limited access remain key caveats.

AI-Powered Automation Software: A Cyber Monday Guide

Learn how AI-powered automation software handles everyday business work, where it helps, what can go wrong, and how to choose a useful first project.

Fable, Opus 5.5, Astra, Sol And Luna: Which AI Model Is Worth Paying For?

Analysis of leading AI models—Fable, Opus 5.5, Astra, Sol, Luna—highlighting performance, costs, and use cases to guide purchasing decisions.

Google’s Genie 4: Exploring What Its AI Worlds Can And Can’t Do

Google DeepMind previewed Genie 4 on Oct. 27, describing interactive worlds with generated sound. Public access, pricing and performance details remain limited.