Hacker News Reader

Comment by mmmore | Hacker News Reader

Does the use of "foundation" and "multi-modal" for describing this model mean anything, or are those just used as buzzwords? Funnily enough, the only place those terms appear in the paper is in the abstract.

Also the paper says they basically copied the methods used for AlphaFold, but then included the ability to input language embeddings, and input some other side constraints that I don't have the biology knowledge to understand. They don't show any data that indicate how much these changes improve performance. They show a very modest improvement over AF3 (small enough that I would think it could be achieve through randomness/small variations in the training parameters). So I don't think this is very revolutionary, but I suppose it replicates AF3.

View on HN

If by "multi-modal", you mean "it takes several different datatypes as input or output", then yes, it's multi-modal. See Figure 1 in the Tech Report.

Foundational maybe isn't the best label for this kind of model. My understanding of foundational models is that they are made to be a baseline which can be further fine tuned for specific downstream tasks. This seems more like an already fine tuned model, but I haven't looked carefully enough at the methodology to say.

Would you then call it a buzzword, or is there some gentler excluded-middle interpretation of that word's application to the project?

It’s about like referring to a famous person’s red carpet attire as “off the shelf [designer name]”. It downplays the effort that went into it more than anything.

I don't think it's a particular buzzword here. They claim it's useful across a range of tasks, and that's the key part imo.

Now, "predictions for parts of drug discovery" isn't the widest range, so perhaps you need to consider "foundation" as somewhat context dependent, but I don't think it's a wild claim. Neither "foundation" nor "fine tuned" are really better than each other, but those are probably the two ends of a spectrum here.

My get-out clause here is that someone with a better understanding of the field may say these are actually extremely narrowly trained things, and the tests are equivalent to multiple different coding problem challenges rather than programming/translation/poetry/etc.

There is a pretty noticeable improvement for antibody-antigen interactions - looks like double-digit percents. Check out figure 4 here: https://chaiassets.com/chai-1/paper/technical_report_v1.pdf

Figure 4 is comparing the model with itself, unless I'm misunderstanding it. The takeaway seems to be the model performs better if you give it extra "constraints", i.e. extra info already known about the protein.

The table with a comparison to alpha fold gives a less than one percentage point improvement.

Reply on news.ycombinator.com