RikiyaT/lm_understand-logs

This is being blown way out of proportion. They told Astra to use old school mech interp tools (see the prompt ) which were surely in Astra training set. However, these techniques are not used that…

让 GPT-6 Astra 盯着一个小模型的权重看，它找到了让这个模型「说车轱辘话」的那几个数，改掉以后，数学题正确率从 7.9% 涨到 48.1%。没训练，一步梯度都没跑。 MIT CSAIL 给它 Qwen3.5-0.8B 的权重，只准在练习题上试跑 50 次。它追到是 2 个层写入的信号让模型一直重复，直到撞上长度上限。减掉这个信号，只改了 0.016% 的权重。 GSM8K 也从…

GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights directly, no gradient descent. MATH-500 went from 7.9% to 48.1%. On Gemma 3 270M it moved nothing: edits unblock skills, they don't add them. Which…

The from-scratch results here stood out to me. With no data or gradients, Astra wrote working MinAtar policy weights one-shot, and set a 9x9 Go net's 1.48M weights starting from zero. Together with…

No training run. An AI edited 0.016% of a small open model's weights and its math score went from 7.9% to 48.1%. • GPT-6 Astra got Qwen3.5-0.8B's raw weights as a prompt, plus a Python sandbox and 50…