Nice Christmas read: The NY Times writes about how AI "hallucinations" can actually be a good thing when it comes to designing proteins. www.nytimes.com/2024/12/23/s... - ThreadSky

About ThreadSky

martinvoegele.bsky.social • 64 days ago

Nice Christmas read: The NY Times writes about how AI "hallucinations" can actually be a good thing when it comes to designing proteins.

https://www.nytimes.com/2024/12/23/science/ai-hallucinations-science.html?unlocked_article_code=1.kE4.NWHE.s4V4Jd71gGH1

Comments

josipk.bsky.social•64 days ago

I'd say hallucinations (differences between what's real and what the model predicts) are never desired. The difference between language modelling and designing proteins is that often in language modelling we care about what is factual. Not always, e.g. when we use language models to invent stories.

josipk.bsky.social•64 days ago

In designing proteins we often don't know what makes sense physically (not hallucination) and what not (hallucination), so we try more reliable methods on what generative model says it's more likely.

josipk.bsky.social•64 days ago

I think in both cases we do care about (self-)consistency. While I understand what (logical) consistency means in language modelling, I have trouble understanding what it means in protein design. Probably because I don't know (yet) enough physics and biochemistry.

josipk.bsky.social•64 days ago

Thanks for the article, good food for thought!

martinvoegele.bsky.social•64 days ago

Yeah, I agree with those who wouldn't call a successfully predicted novel protein a "hallucination." Probably that's also why the Times put it in quotation marks.

Posting Rules

Be respectful to others
No spam or self-promotion
Stay on topic
Follow Bluesky's terms of service

Comments

Posting Rules

Reply