Compression artifacts are data fossils, not windows into a digital mind
gerardsans · x · 2026-11-07
In a follow-up to his argument that ICL is just a regular forward pass, Gerard Sans criticizes much interpretability work as 'latent space archaeology': compression artifacts aren't a window into a digital mind but an average that gained probability mass — a bias, a data fossil of that specific model and training regime.
More from Models
- In 2026, no voice model comes close to 4o Advanced Voice Mode shown May 2024 — Angaisb_ · 2026-11-07
- OpenRouter data: token prices collapse in 2026, open weights at $0.084 vs closed $0.52 per 1M — AccBalanced · 2026-11-07
- Western open-weights model Inkling 2S matches DeepSeek-level GPU kernel optimization — AccBalanced · 2026-11-07
- Chip-design startup open-sources from-scratch Ripple-65B, claims $32M total R&D cost — shoumikgoswami · 2026-11-07
- Canary trap catches Fable 5.1 claiming full PR review while skipping 811 lines — RexDouglass · 2026-11-07
- Less-verifiable benchmarks speed up just as fast: 1.64x vs 1.67x in Anthropic data — ben_j_todd · 2026-11-07