Frontier LLMs may be post-trained on protein data: SOD1 ranking test matches ESM2
alex_peys · x · 2026-11-15
A user ran an experiment giving LLMs (no tools, no internet) just 453 amino acid sequences of SOD1 (ALS protein) variants plus assay type, asking them to predict withheld dimerization readouts. The resulting agent, Astra, roughly matched ESM2's ranking ability and trailed ESM3.
This raises the question of whether frontier labs are post-training on protein language model outputs — or whether it's an emergent capability from massive public biological corpora (cue the bitter lesson).
alexpeys proposes a test: if labs train on protein datasets, asking each LLM for a 20% sequence edit that still folds, then checking with AlphaFold, should reveal large variance across models exposing who trained on protein data. Still unverified speculation.
More from Models
- 文心 Ernie 5.5 Flash 预览版现身 RuneBench，号称接近 GLM 5.2 水平 — mariofilhoml · 2026-11-15
- GLM-5.3 on Two DGX Stations Peaks at 607.8 tok/s but 32 Concurrency Causes 453s Latency — petrusenko_max · 2026-11-15
- Unverified leak claims Gemini 4 Hydrogen targets fast agentic execution — feblet · 2026-11-15
- Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 Sol — The Decoder · 2026-11-15
- Anthropic says outdated prompting habits are slowing down Claude Opus 5.5 — The Decoder · 2026-11-15
- Anthropic IPO deck chart sparks debate: Mythos could be retrained with a fraction of the compute — Witty_County5128 · 2026-11-15