Not sure why Silicon Valley is so fixated on marketing AI to scientists as a tool for hypothesis generation? We have no problems generating hypotheses. We can crank out half-baked ideas just fine. You can help by pipetting, debugging unmaintained code, and troubleshooting weird unreported edge cases - ThreadSky

ddelalamo.bsky.social • 4 days ago

Not sure why Silicon Valley is so fixated on marketing AI to scientists as a tool for hypothesis generation? We have no problems generating hypotheses. We can crank out half-baked ideas just fine. You can help by pipetting, debugging unmaintained code, and troubleshooting weird unreported edge cases

Comments

franknoe.bsky.social•4 days ago

Awesome 😂

pedrobeltrao.bsky.social•4 days ago

As someone that reuses public data to address questions, I think there is a very large potential for a model that can process across the whole of the literature to connect distant observations and concepts. The literature is so vast and adding more so quickly that these tools will be important.

ddelalamo.bsky.social•3 days ago

I’ve actually spent some time thinking today about this since you posted it this morning - a major source of skepticism for me is that good scientific dialogs will sometimes yield answers to questions I didn’t think to ask, and SV hasn’t moved away from the chatbot interface, which imo can’t do that

pedrobeltrao.bsky.social•3 days ago

I don't think the methods are there yet but they keep getting closer. As an example, I recently tried to get suggestions for projects that fit a grant call and my group's expertise. The directions for the questions were too obvious but all very valid. A discussion for longer than a short txt.

ddelalamo.bsky.social•3 days ago

Were you using the $200/mo deep research tool from OpenAI? I’ve heard things about that model that are nothing short of extraordinary (but that it is still flawed)

pedrobeltrao.bsky.social•2 days ago

I use the free OpenAI and Gemini models. Even the free models have now these "reasoning" models but usage is capped. I think these models will soon be able to connect findings from buried literature or poorly connected fields. This will be useful to trigger new questions.

pedrobeltrao.bsky.social•2 days ago

Being able to take such questions and very large public datasets to address such questions is still not possible but you could see how it would work. You would need to have agents that can make analysis on demand on such datasets for example.

hans2sachs.bsky.social•4 days ago

Something to help understand the "entirety" of literature would be amazing. However, it can put us in danger to produce even higher quantities of literature instead of focusing on quality. End-to-end paper generation comes to mind...

viralpersistence.bsky.social•4 days ago

I want to help with the code so bad. used to have a job doing this but I’d do it for free if there was some sort of volunteer program.

ddelalamo.bsky.social•4 days ago

There are tons of academic codebases with outstanding issues on github. Also lots of old scientific software written in stuff like Fortran that needs to be ported to something more modern

viralpersistence.bsky.social•4 days ago

interesting. I typically assume academics don’t want random people submitting pull requests but maybe I’m wrong. I see you study proteins, which I’ve been interested in writing code for—is there any specific software you’d recommend trying to work on?

ddelalamo.bsky.social•4 days ago

Depends on how much work you'd like to do. If you hate yourself and are prepared to really jump in the deep end, PyRosetta always needs work. Otherwise there are plenty of random bugs in BioPython that need to be fixed. I can try to think of a few others

viralpersistence.bsky.social•4 days ago

I hate myself! I worked with biopython at my previous job also

jpglmrodrigues.bsky.social•3 days ago

We happily welcome contributions! Check our open issues for ideas on where to start and if you have any questions about how to go about it, etc, just ask.

https://github.com/biopython/biopython/issues

btnaughton.bsky.social•4 days ago

I think trawling old papers and patents for ideas that were e.g., ahead of their time could be very useful! Nobody is going to go back and read all the literature from the 1970s.

ddelalamo.bsky.social•4 days ago

That's actually an excellent use of this technology that I never thought of

ciannwn.bsky.social•4 days ago

I dont think they can really help with those things that need correctness either. The AI, like them, is better at being the 'ideas guy'.

davsal1.bsky.social•4 days ago

Yeah, may be for the same reason why code is not maintained once the paper is published. No reward anymore. Maybe at some point somebody will train a LLM that takes a whole software and returns a debugged/optimized version. That something deserving a Nobel.

garantes.bsky.social•4 days ago

I guess it is pure marketing, maybe inspired by AF2 success: If AI can "teach" scientists, it can "teach" anyone...

daweonline.bsky.social•4 days ago

I have the same feelings. It would be more useful if Google gave me money for my research 😅

ddelalamo.bsky.social•4 days ago

Or they made sheets competitive with excel by, for example, not converting random PDB codes to dates

daweonline.bsky.social•4 days ago

😂 true!

ddelalamo.bsky.social•4 days ago

I try not to be a crank, this is genuinely incredible tech, but a rule of thumb for succeeding in science is to Always Be Closing™ and LLMs do not close, they do between 5% to 80% of the work before you, the closer, needs to come in and decide whether their output needs a polish or a hard reset

lilopozzo.bsky.social•4 days ago

This hype and emphasis is likely disincentivizing generations of future scientists that could think there is no future for them and for creative thinking. This could not be more ‘wrong’.

g--0.bsky.social•4 days ago

none of those latter things are things that AI is capable of though

ddelalamo.bsky.social•4 days ago

Yes, that's the point! Nobody's working on these impactful use cases and it's annoying

(BTW, I recognize LLMs are not the best type of ML for things like pipetting. I'm just being facetious)

g--0.bsky.social•3 days ago

some of those things are addressed through normal computing

Comments

Posting Rules

Reply