"How will AI impact science?": michaelnotebook.com/mc2023/
Transcript of a short survey talk given at the Metascience 2023 Conference. The focus is on extant systems & the near term, not the long-term impact of AI on science (though I hope the talk helps ground longer-term thinking)
Conversation
I also felt 's question strike at the heart! But I wonder if it is more about 'well-definedness of the problem' rather than cleanness of data that makes it unique. Finding a protein structure given an amino sequence seems almost as well defined as winning a chess game.
1
3
8
I don't think so. Current LLMs are at least somewhat useful as substrates for exploring vague ideas and questions. I expect multi-model ones will get gradually better.
Still, I agree with you that it's a lot easier to make progress when you have a well-defined problem
2
2
6
Show replies
I wonder how would you interpret the Eureqa story I've sent via DM in the context of your afterword.
1
great write-up and talk. Regarding alpha zero there was a paper pnas.org/doi/10.1073/pn that tried to inspect the chess 'concepts' learned by the network and how it related to human chess knowledge.
Show additional replies, including those that may contain offensive content
Show
Discover more
Sourced from across Twitter
The scale of modern GPU computation is so incomprehensible that I regularly find that even experts underestimate it. A 4090 can do ~150 THOUSAND fp32 ops per pixel per frame at 4k 60 Hz, and can load kilobytes for every single pixel from VRAM (and more from on-die SRAM).
6
19
156
Show this thread
There is a lot of excitement for AI to be a universal tutor. And it shows real promise, but there are some important problems that need to be solved.
To get a sense of how good it is, try this prompt (in GPT-4): chat.openai.com/share/ec1018ec
More in our paper: papers.ssrn.com/sol3/papers.cf
18
136
733
Show this thread
This is very tangential, but even among scientists, debunking doesn't work.
Let's look at some of the empirical work on the topic of replication failures.
First: papers that don't replicate receive more citations than ones that do.
Quote Tweet
So uhm….A new meta-analysis of >200 effect-sizes (n > 60,000) in @NatureHumBehav found that on average debunking scientific misinformation had no effect
Yikes!
nature.com/articles/s4156
15
103
451
Show this thread
One beam. He nudged one beam.
Quote Tweet
0:38
232.6K views
147
145
1,184









