GPT-6 Astra edits 0.016% of Qwen3.5-0.8B weights, boosting math up to 6x with no training
_lewtun · x · 2026-11-03
- Weight-Space Programming: A developer asked GPT-6 Astra to directly edit Qwen3.5-0.8B's weights to improve math — no gradients, no training.
- The model changed only 0.016% of weights, fixing the 0.8B model's reasoning loops and improving math up to 6x, with instruction following up 3x.
- The author argues frontier LLMs can now natively read and write neural network weights, dubbing it "Weight-Space Programming."
- Maxime Labonne's RT notes a cheaper alternative: just ask Astra to use Antidoom (FTPO for eliminating Doom Loops), claiming it's 100x cheaper and faster.
More from Models
- paul_cal catalogs the untold failures of LLM tuning — paul_cal · 2026-11-03
- DeepSeek V5 reportedly to be pretrained on Huawei Ascend chips, Ascend 960DT due Q1 2027 — teortaxesTex · 2026-11-03
- Rumors swirl of massive GPT-6.5 pretraining run reportedly using millions of GPUs — VraserX · 2026-11-03
- Tsinghua spinoff Apex Intelligence open-weights Apex-35B, built by automated AI research system — jiqizhixin · 2026-11-03
- Microsoft's MAI-DxO hits 83% diagnostic accuracy on 304 NEJM cases vs 36% for physicians — jiqizhixin · 2026-11-03
- Google releases EmbeddingGemma 2 paper: unified multimodal embeddings beating models 12x larger — osanseviero · 2026-11-03