New weight-space benchmark ranks GPT-6 Astra first at 96.3 for writing neural net weights from scratch
BorisMPower · x · 2026-11-05
Developer calebfoundry released an open benchmark testing whether LLMs can write neural network weights from scratch: given MinAtar game rules (Breakout, Seaquest, Asterix) and a zero-initialized 600→32→32→6 policy net, models must output every weight. GPT-6 Astra leads at 96.3 against a 500M PPO baseline, followed by GPT-6.1 Sol and Claude 5.5. A related track tests direct weight editing to improve Qwen3.5-0.8B on MATH-500/GSM8K. The author remains skeptical but wants the ecosystem to probe the idea's limits; BorisMPower speculates the same capability could one day edit DNA.
More from Models
- Users report models feeling smarter ahead of expected OpenAI release tonight — xiaohu · 2026-11-05
- ByteDance open-sources OpenViking: agent memory that cuts input tokens up to 91% — Promptmethus · 2026-11-05
- Qwen3.5-0.8B 'weight-space programming' finetune trends on Hugging Face — RikiyaT · 2026-11-05
- Hugging Face launches ML Intern, an open-source agent that does ML for you — Gradio · 2026-11-05
- ChatGPT, Claude and Gemini all meter usage now — none tells you what one message costs — Ok_Negotiation_2587 · 2026-11-05
- Jot: Google engineer's free open-source Mac dictation app powered by Gemini 3.5 Transcribe — bigaiguy · 2026-11-05