Four researchers at MIT CSAIL report that OpenAI’s GPT-6 Astra, editing Alibaba’s open-weight Qwen3.5-0.8B directly, with no gradient descent and no training, raised its MATH-500 score from 7.9% to 48.1% with the same sampling settings. The edit changed 131,072 of the model’s 0.8 billion weights, or 0.016%, according to a blog post published on 2 November by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab.
By our arithmetic, 48.1% is 6.1 times the unedited score. The jump came only in the run that allowed 50 forward passes: runs allowed 10 and 25 passes reached 12.3% and 14.1%. The authors say the edit fixed a looping failure that Qwen’s own model card warns about, rather than adding math ability.
Business Pill · Fine-Tuning Explained
A short Business Engineer explainer on fine-tuning: how a general AI model is adapted to one job. It teaches the general idea only and says nothing about the study in this piece.
The key insight: By our arithmetic, GPT-6 Astra’s 0.016% weight edit took Qwen3.5-0.8B’s MATH-500 score to 6.1 times its base, but only in the run allowed 50 forward passes; runs allowed 10 and 25 passes added 4.4 and 6.2 points.
The authors call the approach “Weight-Space Programming”: a general-purpose LLM reads and writes a network’s weights directly, not through an optimizer. Their post says “frontier LLMs are surprisingly good at natively reading and writing neural network weights”, and describes its results as “very early results”.
For the Qwen experiment, the released logs say Astra ran through OpenAI’s Codex CLI at reasoning effort “max”, with Codex’s own tools switched off. It got a Python kernel over the weights (CPU only, no GPU, no internet), 7,469 practice problems from GSM8K and 7,275 from MATH, none from the test sets, and a budget of 10, 25 or 50 forward passes. Gradient descent and any training were forbidden, and Astra never saw the benchmarks.
The post says the base model loops: “85 of its first 96 practice answers ran to the 32,768-token limit”. Astra traced the looping to one direction written by the MLPs of layers 15 and 19 and subtracted it from 64 output columns in each, 131,072 weights in all.
With 50 forward passes, the post reports, MATH-500 rose from 7.9% to 48.1%, GSM8K from 13.8% to 67.1% and IFEval, an instruction-following test, from 21.0% to 62.3%. The share of MATH-500 answers hitting the token limit fell from 93% to 33.1%. The logs add AMC23 (0.5% to 23.4%), MMLU-Pro (40.3% to 45.5%) and GPQA Diamond (22.7% to 36.6%).
Qwen3.5-0.8B’s scores after OpenAI’s GPT-6 Astra edited its weights with no training, by the number of forward passes each run was allowed, as reported by MIT CSAIL researchers in a blog post and released logs on 2 November 2026. Each budget is a separate run with a different edit. The 6.1x and the point gains are our arithmetic.
Why the 50-Pass Run Is the Result
Each budget was a separate run that found a different edit, the logs show. The 10-pass run removed a loop direction from 30 output matrices in layers 8 to 22; the 25-pass run switched off attention head 7 of layer 15; the 50-pass run made the MLP edit described above. By our arithmetic, the first two added 4.4 and 6.2 points on MATH-500, and the 50-pass run added 40.2 points. In the 10- and 25-pass runs, 87.2% and 86.4% of MATH-500 answers still hit the token limit.
The 50-pass run took 158 turns, 80 Python runs and four conversations over 20.3 hours of wall-clock time, the logs say, including 4.7 hours when the model could not be reached. A full generation pass took up to about half an hour.
As we read it, the scarce input in this experiment was the feedback budget, not the number of weights changed: the same model, given five times as many forward passes, found an edit that moved the score by 40.2 points instead of 4.4.
Reading, Fixing and Writing Smaller Networks
The post reports seven other experiments. From raw weights alone, Astra named the right game for all four small game-playing networks. In one shot, it edited 3.5% of a failing car-control network’s weights and the car succeeded in all 1,024 episodes. Writing game networks from scratch with no training, it scored 517.4 in MinAtar Breakout against 15.8 for a PPO network trained for 100 million steps.
Not every test worked. From the weights alone, Astra could not tell which of two Gemma 3 270M copies had been trained to be sycophantic, in all four attempts; it picked the sycophant in all three pairs once allowed to see attention patterns. On Gemma 3 270M, a model that rarely loops, no edit moved MATH-500 by more than about a point. The authors write: “Fixing loops, however, is not the same as making a model better at math: it lets the model finish answers it could already work out.”
They add: “Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked.”
What It Means for the AI R&D Loop
The authors list possible uses: auditing a model from its weights without running benchmarks, cheaper optimization, adding or unlearning a capability without data, model merging across architectures, and recursive self-improvement. They frame each as a possibility and write that their results are “exciting but still limited, specifically due to inefficiencies in context length and harnesses”. They say they are building a benchmark and a harness for the task.
As we read it, the business lever is who can change a model after release: here a closed frontier model acted as the editor of an open-weight model, inside a harness that rationed how often it could run that model. On the same day, Anthropic’s amended S-1 said its internal estimates indicate it “may imminently meet” its Responsible Scaling Policy threshold for AI-accelerated progress, as we reported. The MIT work does not involve Anthropic, and Astra edited a different, much smaller model, not itself.
As we read it, the experiment adds a step to the AI R&D loop that sits between prompting a model and retraining it: one model diagnosing a failure inside another model’s weights and editing it directly.
As we read it, the binding input was feedback, not weights: the 50-pass run touched 131,072 weights, while the logs measure its effort in 158 turns and 20.3 hours of wall-clock time that include the forward passes.
As we read it, the authors’ own Gemma result sets the boundary of the claim: the edit removed a failure that hid an ability, and on a model without that failure it found nothing comparable to fix.
Takehi, Teoh, Zhang and Khattab (MIT CSAIL), blog post, 2 Nov 2026
“Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked.”
Three Implications
FOR AI LABS AND MODEL BUYERS The authors report a targeted fix to a released open-weight model with no retraining: 131,072 of 0.8 billion weights changed, which they say can leave the rest of the model intact.
FOR THE COST OF MODEL IMPROVEMENT The 50-pass run used 158 turns and 80 Python runs over 20.3 hours, the logs say; the documents we read give no dollar or total compute figure.
FOR SAFETY AND EVALUATION The authors list auditing a model from its weights as a possible use, but in their own test Astra could not identify a sycophantic model from weights alone.
The framework opens: “The transformer is a commodity now; what changes between paradigms is where intelligence accumulates and who can capture it.”
As we read it, this experiment places some of that capture in the harness: the gain came from the forward-pass budget and tools wrapped around GPT-6 Astra, not from new data, and the open-weight model was the material it worked on. The framework is the owner’s argument and does not discuss this study.
The results come from a blog post and its released logs; we read no peer-reviewed paper. There is one run per forward-pass budget, so whether a repeat 50-pass run reaches 48.1% again is not shown, and the 25-pass run scored lower on IFEval than the 10-pass run. Scores were measured with thinking on, a 32,768-token cap and a presence penalty of 1.5; the authors say fixing loops lets the model finish answers it could already work out, which they distinguish from making it better at math.
The documents we read give no dollar or total compute cost; the logs record token usage per turn and say it accumulates over a conversation. The post’s figure shows an interim unedited GPQA Diamond score of 24.5; the logs give the final 22.7. We read no comment from OpenAI or Alibaba’s Qwen team, and we did not contact the authors. Nothing here shows a model editing its own weights.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
MIT CSAIL researchers report that GPT-6 Astra, editing 0.016% of Qwen3.5-0.8B’s weights without training, raised MATH-500 from 7.9% to 48.1%, 6.1 times the base score by our arithmetic. The gain appeared only with 50 forward passes, came from fixing a looping failure, and did not repeat on a model that does not loop.
A note on sourcing. The results and quotations come from the MIT CSAIL authors’ blog post of 2 November and the README files of their released logs on GitHub; we did not rerun any experiment. Ratios and point gains are our arithmetic. We did not contact the authors, OpenAI or Alibaba.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
