MIT Study: GPT-6 Astra Edits 3.5% of Weights, Lifts Car Task From 0/256 to 1024/1024
DeepLearningAI · x · 2026-11-19
DeepLearning.AI's The Batch reports work by MIT researchers Rikiya Takehi, Jayden Teoh and colleagues showing GPT-6 Astra can interpret, improve and generate neural network weights without explicit retraining, tools or gradient steps.
Experiments
- Identification: four tiny vanilla networks (17,000 parameters each) trained on Breakout, Pong, Pac-Man and Frogger; given only the weights and asked which game, GPT-6 Astra got all four right.
- Weight editing: a 171,000-parameter network learning to drive a car up a ramp in the Kinetix environment had training suspended after failing. Given the problem description, goal, weights and one failed action, GPT-6 Astra edited 3.5% of the weights, taking the car from 0 of 256 successful runs to 1,024 of 1,024.
- Generation: the model was also asked to produce networks that win a game and to improve Qwen3.5-0.8B's math performance.
Takeaway: for networks under 1B parameters, prompting an LLM may be enough to improve weights — a proof of concept.
More from Models
- Gemini 4 Neon lands Day 0 in LangChain's Managed Deep Agents for long-running agent work — hwchase17 · 2026-11-19
- Early user shows off Gemini 4 Neon paintings drawn entirely in code — Swarooprm7 · 2026-11-19
- Google Releases Neon Model, Tops Coding Agent Benchmark Worldwide — oran_ge · 2026-11-19
- Google AI Ultra Plans Now Include YouTube Premium Family for Free — osb103 · 2026-11-19
- Peter Yang: Gemini's 'Neon' Is the Writing Rival to Claude Opus — petergyang · 2026-11-19
- Gemini 4 Neon takes #1 on CursorBench, dominating the Pareto frontier — melvinjohnsonp · 2026-11-19