it's a tongue-in-cheek comment on the quote from Demis. W/ current methods, data availability is the primary determinant of performance (see plot), and unlike quote, DM hasn't found a magic "innovative algorithm" to get around it. I think both more (diverse) data + better algos will be part of soln

lmao

Note that I believe that still much improvement can & will be made on model side & even more so, leveraging more existing data sources. But it's very clear that only focusing on the model is a losing strategy with no moat.

🫡🚀 lets gooo

Bonus: you can now watch on YouTube the recording of the seminar we gave on Boltz-1 with @mmbronstein.bsky.social and VantAI! www.youtube.com/watch?v=mN5p...

actually agree here that we should provide "sensible defaults". I think we could just provide a few presets ("high quality balanced", "high-data balanced", "unfiltered-unbalanced"). On the exact specifics of those presets would be great to share notes, you have a lot of exp here obviously

but if we can make metrics less strict that would give leeway to optimize relevance so this great. By no means current splits are the final word, it's a framework to research splits easily. 💯 agree that more than 1 split is likely end-state. Would love to work together on the next iteration! (2/2)

This is incredibly (!) useful David, really appreciate you taking the time to dig in. Orthogonal descriptors validation is great to see. We'll check what happened with redundant ligand in tests, intention was to remove that. it's non-trivial to optimize generalizability, relevance and quality (1/2)

Congrats!! 🥳