The MIT experiments are real, and an outside engineer re-ran the headline test and got a similar number. Yet the sixfold maths gain came from deleting a repetition habit, not from new skill, and the cleanest reading results came from small networks built to be readable. The strongest claims rest on tests the authors designed themselves, so their planned benchmark will say more than the blog post does.
Table of Contents
On 2 November 2026, four researchers at MIT CSAIL reported that OpenAI’s GPT-6 Astra can read and rewrite the weights of other neural networks without any training run. The best-known result is a small open model, Qwen3.5-0.8B, whose score on the MATH-500 benchmark rose from 7.9% to 48.1% after Astra changed 131,072 of its 0.8 billion weights, or 0.016%. The edit removed a repetition loop and added no mathematical ability, as the authors say themselves, and an independent re-run landed close to their figure. That is a narrower claim than the headlines that followed, and it is the one the evidence supports.
A video from the channel Two Minute Papers, titled “OpenAI’s Astra Just Opened The Black Box Of AI”, is the example this article tests. It lists three results (reading, fixing and improving a network), then spends a long stretch on common misreadings: that Astra taught the model maths, that all three feats happen at once, and that only OpenAI’s model can do them. Most of those corrections hold up. The title promises more than the experiments show, and the video closes with a promotion for the GPU cloud provider Lambda that has no bearing on the findings.
Reading, fixing and writing were separate experiments
The post describes eight experiments, and the video is right that they do not form one pipeline. In the first, four small game-playing networks with 17,253 weights each, trained on grid versions of Breakout, Pong, Pac-Man and Freeway, were handed to Astra as raw numbers. With no list of possible games, it named all four games correctly.
In the second, a car controller kept stalling on a ramp in a physics level. Given the weights, a description of the scene and one recorded failure, Astra edited 3.5% of the weights in a single reply, and the car then cleared the ramp in all 1,024 test episodes, against none in 256 before. The third is the Qwen repair. Each result came from its own setup, mostly on small networks, which is the point the video makes when it warns that nothing here happens “at the same time” in one system.
The 48.1% came from stopping a loop
Qwen3.5-0.8B has a documented habit. The authors point to its own model card, which warns that the model is prone to thinking loops, and on MATH-500 93.0% of its answers ran to the 32,768-token limit without finishing. Astra traced the looping to one direction written by the feed-forward layers 15 and 19 and subtracted it from 64 output columns in each. The share hitting the limit fell from 93.0% to 33.1%, MATH-500 rose to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval, an instruction-following test, from 21.0% to 62.3%. A gain on a test that is not about maths fits a general loop fix and not a maths lesson, which is this article’s inference from the numbers.
The control run settles it. Gemma 3 270M rarely loops, and Astra could not improve it: MATH-500 moved from 6.8% to at most 7.7% across four budgets. The authors summarise the lesson in one line, saying weight edits could remove an obstacle but “could not add an ability the model lacked”. VentureBeat reached the same reading.
Fifty test runs and twenty hours stand behind the patch
The video says Astra tested the other model repeatedly, and the numbers show how much that mattered. With 10 test runs on practice problems, MATH-500 reached 12.3%; with 25, 14.1%; only with 50 did it reach 48.1%. The share of answers hitting the token limit barely moved at 10 and 25 runs (87.2% and 86.4%). IFEval went 49.0%, 39.9%, then 62.3%, so the curve is not smooth, and Andreas Maier’s review notes that each budget was run once, leaving the spread between repeats unknown.
VentureBeat, citing the team’s experiment documentation, reports that the strongest attempt took 158 conversational turns, 80 Python runs and 20.3 hours, 4.7 of them spent waiting out a service outage. A small patch needed a long investigation. The patch itself is public, and Hugging Face’s Ahsen Khaliq rebuilt the checkpoint, confirmed that only two tensors differ from the base model and re-ran the full MATH-500 set with four samples per problem. The edited model scored 51.0% in an independent re-run, with 31.6% of answers hitting the limit.
Reading weights worked best where the weights were easy to read
Matthew Honnibal, author of the spaCy language-processing library, argued on 3 November that the post blends different techniques. The setting where Astra sees raw floats with no tools appears only in experiments 1, 2 and 4, on networks of 17,000 to 172,000 weights. Every experiment with a real language model uses Python, and the Qwen fix also ran the model on up to 50 batches of practice problems. Sean Goedecke adds that the game networks had labelled input channels, so a few lines of Python produce a heatmap showing where the player sits, and he describes the Qwen repair as the standard steering playbook a human researcher would follow.
The video’s claim that smaller open models can do the reading too holds, with caveats. Simon Willison ran Qwen 3.8 27B on a laptop using the authors’ prompts. It named Breakout at 88% confidence after 52 minutes, Pong at 75% after about an hour and Pac-Man at 72% after 46 minutes, and it failed on Freeway. That is independent confirmation, and it is slow and incomplete.
Writing a game network from scratch means compiling a hand-made program
The video says writing a gameplay network from scratch could upend how such networks are trained. Astra received the rules and source code of three MinAtar games and produced the weights as text in one reply, without rollouts. In Breakout it scored 517.4 points against 25.3 for a network trained for 500 million steps; in Seaquest it scored 53.1 against 88.9, and in Asterix 53.2 against 55.2. The post explains the method: Astra designs a small program, such as computing where the ball will land and moving the paddle there, and converts it into network layers.
Maier points out that the comparisons are against the authors’ own training runs, with no baseline of the same program written as ordinary code, and Goedecke notes that humans already know how to express any simple program as a network. In 9×9 Go, a 1.5-million-weight network written by Astra beat an equal-sized trained network in 70 to 79% of games with 50 test runs, against 26 to 43% with none, according to Maier. These are real results about turning designs into weights. They say little about understanding weights that training produced.
The black box on the other side of the experiment
The video’s title says Astra opened the black box of AI, and the experiments open a much smaller one. In the test closest to a safety audit, telling an honest Gemma 3 270M from a sycophantic copy, Astra failed in all four attempts from weights alone. It identified the sycophant in all three pairs only after being allowed to see attention patterns on texts it wrote, and Maier notes that three correct binary picks happen by chance one time in eight.
The reader is also hard to inspect. OpenAI released Astra on 3 September 2026, and TechCrunch described it as controversial for its use of opaque recurrence, a technique that obscures the chain of thought researchers use to audit decisions. OpenAI has downplayed how much the model relies on it, and chief scientist Jakub Pachocki said “monitorability is getting more challenging”. PauseAI chief executive Maxime Fournes has warned that successful weight editing could speed dangerous progress, a prediction no experiment in the post tests.
Old ideas, a blog post and a benchmark still to come
Jürgen Schmidhuber called the approach “old hat” and pointed to his work from the 1990s, and Omar Khattab replied that the new part is how well current language models translate between human language and a network’s numbers, according to The Decoder. Direct weight editing also has recent precedent in ROME and MEMIT, which changed factual associations in language models. The result is a blog post, not a peer-reviewed paper. It uses one model, most conditions were run once, the tasks were chosen by the authors and earlier prompt versions are absent from the logs, per Maier.
For teams running small open models, one practical reading follows from the Qwen case: when a model fails by looping or truncating and not by being wrong, a targeted edit may be worth testing before fine-tuning. Let’s Data Science adds that the same tooling can strip safety behaviour, so any edited model should be diffed against the base. The evidence supports narrow model repair today. A benchmark with tasks the authors did not design, several runs per condition and comparisons with ordinary decoding fixes would show whether it is more than that, and the authors say they are building one.
Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency
This article is an original analysis supported by the sources cited below
OpenAI’s Astra Just Opened The Black Box Of AI
The video examined here, whose transcript lists the read, fix and improve results and the caveats about separate experiments, Gemma and open models.
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
The authors’ own post, with all eight experiments, the Qwen results table, the Gemma control and their statement that edits cannot add missing abilities.
RikiyaT/Qwen3.5-0.8B-weight-space-programming
The model card for the edited checkpoint, showing two changed tensors and an independent MATH-500 re-run at 51.0%.
OpenAI launches Astra, its powerful (and controversial) new model
Established the 3 September 2026 release of Astra and the dispute over opaque recurrence and monitorability.
‘Like gene editing for neural networks’: MIT researchers used GPT-6 Astra to edit Qwen’s open weights directly, boosting math performance
Independent reporting on the 20.3-hour Qwen run, the prior ROME and MEMIT work and the safety objection from PauseAI.
GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights
Reported the Go result, the Gemma failure and the exchange between Schmidhuber and Khattab over novelty.
The ‘Astra can read and write weights to improve Qwen’ post is misleading
Critique sorting the experiments by technique and showing that raw-weights-only reading appears in just three of them.
LLMs are not superhuman at reading and writing neural network weights
Argued that labelled inputs made the game reading easy and that the Qwen fix follows the standard steering playbook.
Qwen 3.8 27B on my laptop can interpret model weights, too
Independent replication of the game-reading test on a laptop, with confidence figures, run times and a failed Freeway attempt.
Weight-Space Programming: What Is Measured and What Is Anecdote
Sorted the experiments into measured and weak evidence and flagged single runs, non-monotonic results and missing baselines.
GPT-6 Astra Edited 0.016% of Qwen’s Weights. Its Math Score Rose Sixfold.
Collected the post, the Hugging Face re-run and the critics’ replies within a day of publication, with practical cautions on edited models.
| Citing this article? Brief excerpts are welcome. Please credit Webiano.digital, name the author where stated, and include a link to https://webiano.digital and to this original article. Full or substantial republication requires prior written permission. Read our Copyright and Content Use Policy. |
This article was prepared with the assistance of artificial intelligence tools. The content underwent expert human review, and Webiano Digital & Marketing Agency assumes editorial responsibility for its final version and publication.
