by MIT
MIT research showing a frontier LLM can read and write another neural network's raw weights without gradient descent.
Updated EU status checked Last shown
Sources: creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Weight Space Programming is an MIT research project in which a general-purpose frontier language model inspects and edits the raw weights of another neural network directly, without training or gradient descent. According to creator coverage, the editing model is described as GPT-6 Astra, and it operates on weights as a kind of programmable medium.
Reported experiments include identifying which game a network was trained on, predicting diffusion model outputs, identifying languages, diagnosing and fixing a failing network, writing weights from scratch, and improving a small model's math performance. One cited result raised Qwen 3.5 0.8B on MATH-500 from 7.9% to about 48%.
One reviewer noted that the gain largely came from fixing repetitive reasoning loops, that the editing model used tools, and that the approach did not improve Gemma. These details come from YouTube summaries, not from an official page. No official paper, code or release was available to verify them, so figures should be treated as unconfirmed. This is a research result, not a product or service.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
MIT research where GPT-6 Astra inspects and edits another model's weights without training, e.g. raising Qwen 3.5 0.8B's MATH-500 from 7.9% to 48%. Creator notes the gain mostly fixed repetitive reasoning loops, used tools, and failed to improve Gemma.
A recent paper in which frontier AI models create new models by writing weights directly. Mentioned briefly as a sign AI may become hard to understand, with a pointer to a separate video.
Research paper showing a general-purpose frontier LLM can read and write another network's raw weights directly, without gradient descent. Experiments cover identifying which game a network was trained on, predicting diffusion outputs, identifying languages, fixing a failing network, writing weights from scratch, and improving a small model's math scores.
Closest neighbours by meaning, not just by category.
Model · OpenAI
OpenAI's GPT-6 Astra, a frontier language model used for reasoning, agentic work and weight interpretation research.
Model · Thinking Machines Lab
Open-weight 952B-parameter language model from Thinking Machines Lab, its first public open-weight release.
Model · DeepSeek
Chinese AI lab behind the open-weight DeepSeek-V4 and DeepSeek-R1 models, with a chat app and low-cost API.
Model · OpenAI
OpenAI frontier language model used for coding, reasoning and agentic tasks, available via ChatGPT and the API.
Model · OpenAI
OpenAI's open-weight reasoning models, gpt-oss-120b and gpt-oss-20b, released under Apache 2.0.
Model · OpenAI
OpenAI's mid-tier GPT-6 model, positioned below GPT-6 Astra and above Luna, priced at $2/$10 per million tokens.