Claude was trained on protein sequences at scale, Phylo test finds — no other frontier model shows it
AllThingsApx · x · 2026-11-18
A test by Xinming Tu at Phylo concludes the latest Claude models were trained on protein sequences at scale.
Method: next-amino-acid prediction on 668 bacterial/archaeal proteins from 167 unknown-function Pfam families (no solved structures, ≤3 papers, no close UniProt relatives), compared against shuffled copies — the real-minus-shuffled accuracy gap measures sequence knowledge rather than paper recitation.
Results:
- Claude Fable 5.2 scores +45.1 points, Claude Opus 5.5 +27.4; OpenAI GPT-6.1 Sol/Astra, Gemini, and all open-weight models sit within ±3
- Fable 5.2 beats every open protein language model tested (ProGen3-3B +32.5, ProtGPT3-10B +27.7); Opus 5.5 matches ProtGPT3-10B
- Like protein LMs, performance scales with family size, and holds even in families with zero papers — implying training on sequence databases, not literature
Blog and evaluation code are open. The author asks whether the open-model ecosystem will follow suit.
Related event: Phylo Tests Suggest Claude Trained Heavily on Protein Sequences(2 posts)→
More from Models
- Users rip ChatGPT voice mode: random interruptions, accent drift it can't fix — Yamapama · 2026-11-18
- Oru'el claims Kanook beats NVIDIA Flora and OpenAI on RoboRewardBench by 30-50% — AashaySachdeva · 2026-11-18
- Agent-tuned MiMo scores 0/10 in a real loop while base Qwen 3.5 gets 7-8/10 — GapNew4766 · 2026-11-18
- Tencent releases Hy4 flash: matches Claude Haiku 5.5 (high) at 18x lower task cost — ahuja_priyank · 2026-11-18
- Musk amplifies test: Grok bot finished a task in 2 hours, OpenAI's Dot froze — elonmusk · 2026-11-18
- XBow benchmarks GPT-5.6-sol vs GPT-6-sol vs GPT-6.1-sol on vulnerability discovery — moyix · 2026-11-18