'GPT-6 Astra raised a small Qwen model's math score from 7.9% to 48.1% by editing 0.016% of its weights,' MIT says; critics say the prompt held hints
First seen on X 5 hours ago@kimmonismus ♥ 3,4809 views
MIT researchers said OpenAI's GPT-6 Astra raised a small Qwen model's math score from 7.9% to 48.1% by directly editing 0.016% of its weights, with no gradient-based training.
The post summarizing it (kimmonismus, 3,500 likes) called it a possible step toward self-improving AI. Replies gave details: it subtracted one direction in the MLPs of layers 15 and 19 (131,072 weights), IFEval jumped from 21% to 62.3%, and it needed 50 forward passes; 10 or 25 gave little gain.
Pushback followed. Adam Karvonen wrote that Astra did not natively read the weights: Qwen got stuck in a loop and Astra used code to run standard interpretability techniques, found the looping direction and ablated it (840 likes). Another user confirmed the prompt was very long and hinted at finding neurons and directions, with a reply that stating it plainly would not have spread outside non-specialist circles. Some said the more surprising result was reading raw weights of game-playing networks without hints.
“Absolutely incredible: MIT researchers say GPT-6 Astra improved a small Qwen model’s math score from 7.9% to 48.1% by directly editing just 0.016% of its weights, withou…”
