Post #418821

2026-02-25 16:22 UTC

this is one of those &quot;must reads&quot; if I were more excitable: https://karpathy.github.io/2026/02/12/microgpt/ No matter how you feel about the use of LLMs et al, it&#39;s important to know what these tools are, how they work, and to _not_ believe they are some magic black box. No they are not.

Replies (2)

@KayOhtie@blimps.xyz 2026-02-25 17:19

@funkatron@indieweb.social The technology behind them has always been enormously fascinating to me even if I don't understand it at a very low level. And understanding even that is what lets people see value in specific, non-generalized use cases, like how modern speech-to-text uses that same probabilistic scoring to match utterances to token sequences to be pretty dang accurate (most of the time) without any per-speaker training. Also helps see why it's so garbage for code/general shit/generative everything.

Open ##2620386
@drikanis@mstdn.ca 2026-02-25 18:46

@funkatron@indieweb.social ugh andrej karpathy.... this writeup seems pretty grounded and informative, but that guy has said a lot of complete nonsense in his talks that i'm still trying to walk back with my leadership.

Open ##2620388