CUDA Mode Keynote | Andrej Karpathy | Eureka Labs

[@saturdaysequalsyouth]

This is the same talk that was posted on Latent Space, except the audio quality is much better

[@iony_mikler]

Your simulation seems to run very smooth. What is the computer you are using?

[@dj0ai]

1:24 OMG! Andrej Karpathy lives in 2X. Me? I live in 0,75X

[@ultrasound1459]

3:00 the goat 🐐 🙌

[@eriye92]

The GOAT

[@sizwemamba689]

Excellent

[@elliptictree]

Interesting.

[@kehindeakeredolu3525]

first comment

[@IvanMalison]

i dont get why he's wasting his time doing this. Abstractions are good! Seems like a really strange thing to do.

[@marshallmcluhan33]

Wasting time is what abstraction is all about. Slow tokens

[@philippleser4125]

No. Abstraction is not good per se. Most developers are brain washed that way. One of the reasons why Software has become bloated and sliw.Performance sucks. AK did the right thing.

[@FamilyYoutubeTV-x6d]

I think you have a very valid point, sir. If we want to have "AI computers", we need to have performant and fast model training, execution, inference; otherwise, they are a waste of resources and not worth the hype or money. That's not to say that some "abstraction" is not good. ML and LLMs are not traditional software (in the deterministic sense), and given their goal of wanting to either mimic or emulate intelligence, some abstraction is bound to be necessary. Abstraction is crucial in mathematics and physics, as well as science in general.

[@ultrasound1459]

Abstraction = Slow Ai

[@FamilyYoutubeTV-x6d]

@@ultrasound1459 We are intelligent and have slow areas of intelligence too. Slow AI is not the bottleneck. Planning (and reasoning) is (are) one (two) of them.

[@dontwannabefound]

I’ll save you 30 mins. He basically rewrites pytorch code to train gpt2 in c in 3000 LOC. there.