i find it refreshing compared to standard comments from established researchers proclaiming "don't work on LLMs" or "don't publish many papers" without stating the complexities and incentives of the real world in which all phd students live.

thanks very much for this thoughtful post @kyunghyuncho.bsky.social. i especially liked the part that your take is very nuanced and acknowledges the opportunity cost students feel when working on a research topic in modern AI research landscape.

great thoughtful post @jfrankle.com. respectfully, don't agree that "bitter lesson is over". sepp hochreiter, no doubt a giant of the field, missed it. below is what another giant of the field (michael i jordan) said about deep learning for NLP in his reddit AMA back in 2015 and here we are ...

(2) much less convinced in the case of NLP than, say, vision, the way to go is to couple huge amounts of data with black-box learning architectures." </quote end>

<quote start> "Although current deep learning research tends to claim to encompass NLP, I’m (1) much less convinced about the strength of the results, compared to the results in, say, vision;

great thoughtful post @jfrankle.com. respectfully, don't agree that "bitter lesson is over". sepp hochreiter, no doubt a giant of the field, missed it. below is what another giant of the field (michael i jordan) said about deep learning for NLP in his reddit AMA back in 2015 and here we are ...

thanks!

this is wonderful. @lampinen.bsky.social do you think that pre-training of language models is equivalent to meta learning in some reformulation?

p.s. love your causality work as well.

e.g., i like ludwig schmidt's work which shows better data is the only thing that substantiatlly moves the needle on adversarial robustness. slideslive.com/39015739/a-d...

love this thread @yisongyue.bsky.social ! curious if you also believe that thinking about data as a first class citizen is another research direction that is very rich and wasn't considered important earlier?

totally agree! but do you believe that systems that can train better ML models for us will be able to better science than us as well? i understand this could be a far-stretched conclusion ...

great work! genuinely excited about how this could help neural nets the way your previous work on exploiting linear algebra helped gps.

michael nielsen could be our generation's carl sagan - he just needs to be invited to more podcasts and tv shows.

loved his recent post so had to say this thttps://michaelnotebook.com/optimism/index.html

great slides! can't help but wonder - are we bayesian retrofitting by thinking about in-context learning this way? (see cool post
probapproxincorrect.substack.com/p/some-thoug... from @desirivanova.bsky.social )