Simon Willison replicates GPT-6 Astra weight-interpretation feat with Qwen 3.8 27B in 45 minutes
cephaloform · x · 2026-11-04
A buzzy new paper claims GPT-6 Astra can directly interpret model weights via prompts full of floating-point numbers. Simon Willison tried it on his laptop with Qwen 3.8 27B: after 45+ minutes, that model achieved the same thing.
This suggests direct weight interpretation may not be unique to Astra — open models can pull it off given enough inference time. Worth following the paper and his replication details.
More from Models
- Gemini 4 Carbon takes #1 on Artificial Analysis Intelligence Index — thursdai_pod · 2026-11-04
- Open decision model REAX 1.3 launches: €0.40 per million input tokens, output free — airesearch12 · 2026-11-04
- REAX 1.3 ships open weights under MIT license with two miner-trained models inside — airesearch12 · 2026-11-04
- Anthropic faces questions over Claude's rules on AI-generated video game clones — emmanuelvivier · 2026-11-04
- Qwen, GPT and Gemini Flash all fail the same way on this service — matdryhurst · 2026-11-04
- Sub-1B open OCR models benchmarked: Surya 2.5 tops OmniParseBench at 85.9 — VikParuchuri · 2026-11-04