Charles Dai

[Charles Dai @longstar42]

@timhuyn Hmm, I think something interesting would be to see emerging trends? So domains that I don't know xD
You could probably ask an LLM to create a taxonomy of AI into domains / subdomains / microdomains, and then track the growth rate in terms of publications and citations.

[Charles Dai @longstar42]

I've been reading AI papers recently, and I'm left wondering: what is driving or will drive the progress in LLMs over the next few years? I feel that scaling pretraining is already near the current limit, though there is still room for improvement. Let me know your thoughts :)

[Charles Dai @longstar42]

@gaganghotra_ @ns123abc Their main current problem is that they have the money but due to US restrictions, they can't buy enough cards..

[Charles Dai @longstar42]

@timhuyn The trending section of your website looks like an interesting functionality :) it would be even more interesting if there was a trend for each subdomain of AI, which would answer my questions on the short-term

[Charles Dai @longstar42]

@timhuyn Maybe in parallel of scaling RL data, we should also scale entropy, for example by increasing the size of the vocabulary? Also I'm wondering if we should gather high quality data for RL (for example, in order to develop "taste"), or simply make LLM tame lots of real-world pbs

[Charles Dai @longstar42]

@timhuyn It seems that without a proper mechanism to avoid entropy collapse, the entropy at the start caps the capabilities you can get through RL. The authors propose some fix tho, but it's hard to see how much we can push against this limitation.

[Charles Dai @longstar42]

@GZhan57 @_Jason_Dean_ That's until they turn you into an anime girl out of benevolence

[Charles Dai @longstar42]

Some people are talking about recursive self-improvement as the next step, but how much is it really a thing? I get that AI is insanely good at math and code now, but it still feels like it's lacking research taste for the hardest real-life research problems.

[Charles Dai @longstar42]

Also, I feel like the big progress recently were due to new axis of scaling : CoT and agents thanks to RL,.. but RL feels very hard, since high quality data seems intrinsincally rare, and also training good reward models is hard for subtle tasks bcs of spurious correlations etc