Post

Ryan Moulton
@moultano.bsky.social

I didn't have the vision that saturating any objective we put in front of it would naturally result in it doing things way beyond the tasks we were already deploying ML to do. I sort of thought "Existing stuff gets better. We're still mostly doing binary classification." bsky.app/profile/moul...

Ryan Moulton@moultano.bsky.social

Had I been slightly more converted, perhaps watching BERT demolish every language benchmark at the time would have inspired me to buy NVIDIA.

Ryan Moulton
@moultano.bsky.social

But then even after gpt3 had what to me seemed like obviously actual intelligence, I didn't do the extrapolation and really prepare in any way. I think I was scared of thinking about it.

aron
@adler.dev

for me it's only really been the last month or two using Opus 4.5 that I started to realise even the most scifi-esque of predictions about AI were going to come true after all

0 reposts3 likes

aron
@adler.dev

this might make me look ir-Rational, but I've found it a useful heuristic to not extrapolate every line infinitely far into the future. it was less an opinionated stance that progress would halt once AI started approaching human-levels bc of limitations in the training corpus, and more just, like...

aron
@adler.dev

if you're using LLMs to write code or explain things and it always falls short of your own abilities at some point.. it doesn't really matter whether it failed way before your own capabilities or really close to them. the heuristic "this isn't as clever as me" solidifies. it's hard to look past that