Ethan Mollick called a claim about editing model weights a big deal, while saying it still needs replication. The claim, posted by Rikiya Takehi, says GPT-6 Astra directly changed 0.016% of Qwen3.5-0.8B’s weights without gradients or training. The post says the edits fixed reasoning loops, improved math performance by up to 6x, and increased instruction following by 3x.
Takehi calls the approach “Weight-Space Programming” and says frontier language models can read and write neural network weights. These results and broader claims are reported by Takehi; Mollick has not described them as independently confirmed.
AI 根据报道生成 · 6 小时前更新
