@KhoriatyMatthew Hey, would be really helpful if you breakdown a lil abt what kind of papers and projects work well for confrences like ICML ( ones you'd consider from the first category among the two you described )
@k1rallik I do beleive he's a nutcase, but at the same time you can't deny the biological research opportunities this guy brought on to the table by making himself a test subject. It may fail or succeed, the experimentation stays valuable.
@kmeanskaran I believe with every day new frontier tech inventions. Knowing and understanding the deeper fundamentals has become even more important.
Especially the Research vs Applied AI run, the volatility replicates Stock Market vs Crypto if anything.
@VukRosic99 The effects would be so darned good in leading humanity to new levels and so darned bad for mass radicalization , that the my brain concludes the better answer to be "kaboom".
@zmkzmkz You think the presentation would be decent enough of a source material to read and learn??? Like given someone has not watched the lecture T_T
Will never understand how people finish writing papers on a weekend,
It literally took 13 hours for me to finally form one elegant figure for my paper. From planning to building to all the iterations and refinement, all just for ONE DAMNED FIGURE T_T.
@prajdabre It's funny how papers like this keep aging well. We started with "most heads don't matter" and now we're asking which ones actually carry computation. Every now and then I revisit this paper and appreciate it a little more. It really shaped mech interp.
@VukRosic99 That's quite real, The more I study attention heads, the more suspicious I become of people casually saying "Layer X does Y". Reality seems a lot messier... and honestly much more interesting.
Spent the day finishing the first draft of a paper I am working on.
Building the work undeneath took weeks.
Writing the 15 pages explaining why it behaves the way it does in those grand academic proses somehow felt harder.
Writing can be outsourced, understanding remains manual
It is humbling to consider that if we harness just 1 millionth of the Sunโs power for AI, that will be much more than a million times the intelligence of all of humanity
Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.
@karpathy Last night I was going thru a video from Makemore series you have on youtube, taking help from claude to help me understand things, it repeatedly explained like "karpathy will make it clear now" , and i suddenly thought "has he ever gotten a call by anthropic?"
Well now we know.