←
Michele11 posts
Michele avatar

Michele

@infinitycrab2

Deep learning scientist at CRG Barcelona, working on protein language models and RL-based alignment.

◷ Joined June 2026▧ 2 Photos and videos
Nick avatar
⟳ Michele retweeted
Nick@nickcammarata·Oct 15

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com/

Michele avatar
Michele@infinitycrab2·Jun 19

6/ Try it yourself — no GPU needed: 🔗 Online API for ProtGPT3-MSA: huggingface.co/spaces/AI4PD/… Just upload a FASTA file and get homolog-conditioned designs. Preprint: doi.org/10.64898/2026.06.04.…

Michele avatar
Michele@infinitycrab2·Jun 19

5/ Beyond family-conditioning, prompts can steer properties too. Prompting ProtGPT3-MSA with more stable variants biases generation toward higher-stability sequences. More broadly, we explore inference-time steering via Feynman–Kac-style sequential Monte Carlo.

Michele avatar
Michele@infinitycrab2·Jun 19

4/ RL alignment matters too. We applied DPO across all single-sequence model sizes to reduce low-complexity (repetitive) sequences: • >20% reduction for 112M and 1.3B • ~50% reduction for 10B — in a single iteration! Sequence diversity is preserved.

Michele avatar
Michele@infinitycrab2·Jun 19

3/ Tested on defluorinases — enzymes that break C–F bonds in PFAS pollutants — using just 7 known sequences. ProtGPT3-MSA beat fine-tuned single-sequence models on computational success, with comparable experimental results: 18/20 cloned designs were soluble and expressed.

Post attachment
Michele avatar
Michele@infinitycrab2·Jun 19

2/ The big finding: a 112M MSA model prompted with as few as 15 homologs outperforms supervised fine-tuning (SFT) of models up to 10B parameters on target family generation — across sequence, structure, AND family-level consistency metrics. No weight updates. Just prompting.

Michele avatar
Michele@infinitycrab2·Jun 19

1/ We're releasing 7 models in total: • ProtGPT3-112M, 1.3B, 10B (single-sequence) + the RL-aligned versions to reduce low-complexity generations while preserving diversity. • ProtGPT3-MSA (112M, homolog-conditioned) Models + code: huggingface.co/collections/A…

Post attachment