MIT had GPT-6 Astra read Qwen's raw weights and fix a looping bug — no gradients, no training
ccerrato147 · x · 2026-11-04
An AI opened another AI's weights, found the bug, and fixed it with no training run. The MIT CSAIL team gave GPT-6 'Astra' the raw weights of a small Qwen model, a Python tool, some practice problems and 50 test runs; Astra located a known looping habit in two layers and subtracted it, and the model got better at math. The author compiles ten verified facts from MIT CSAIL's post, the model card and four independent reruns.
More from Models
- ChatGPT 老订阅者怒斥 OpenAI：改 token 计费却不公布额度上限 — LowRepublic4680 · 2026-11-04
- Tiny LoRA adapter beats DNA model Carbon-X on RareBench, exposing generalization limits — danielmckinn0n · 2026-11-04
- LangChain benchmarks 7 browser toolsets: Stagehand wins with 90% pass rate at $1.98 vs $83.62 — LangChain · 2026-11-04
- DeepSeek quietly launches "DeepSeek Harness" YouTube channel, hinting at plugin-based agent framework — gnukeith · 2026-11-04
- Reddit user seeks uncensored local model that runs on a weak PC for search and bulk downloads — Emotional_Orange_279 · 2026-11-04
- DeepSeek partial outage fuels speculation compute is shifting to unreleased V4.1 Pro — AccBalanced · 2026-11-04