Ask HN: How do you keep up with AI advances, esp. Stable Diffusion and GPT-3? | Hacker News Reader

Ask HN: How do you keep up with AI advances, esp. Stable Diffusion and GPT-3?

27 pointsbehnamoh14 comments

Every week there's a new progress made in the fields of language models and diffusion techniques. I think it's almost "too much" for me to handle, and I often feel like I'm left behind.

My gut feeling is that some (most?) these advances are not actually robust or that useful. But still, there's so much to keep in mind. I really like working in this field, but how do I keep myself up-to-date with all these AI achievements and filter noise from signal?

View on HN

Two minute papers [1] is very good YouTube channel. Follows mostly papers related to video and graphics.

[1] https://www.youtube.com/c/K%C3%A1rolyZsolnai

I found that there isn't that much to keep up with anymore. There have been lots of amazing results, but the techniques have changed little over the past 2-3 years. Usually it's a combination of existing models glued together, huge datasets, smart data processing/augmentation and input transformations, etc. It feels like we're now in the "apply and scale what we know works" phase, which is great. I predict this will continue for a while as people figure out how far we can push the current approaches.

This feels different from 3+ years ago. That's when we regularly saw radically different models working better. But there hasn't been a "Transformer moment" or "ResNet moment" or "GAN moment" for quite a while now. In 2014-2018 we had these wow moments every year. Now the wow is not about new developments but more like "wow, I didn't expect this to work so well with more data"

This year it definitely feels like a “diffusion moment”.

Check paperswithcode's State of the Art leaderboards for the tasks I care about (sound classification, style transfer, etc)

https://arxiv.org/list/cs.AI/recent

If it weren't for Hacker News, I wouldn't have a chance to keep up with the new progress.

i dont, i just accept my fate and offer my full submission to our future ai overlords.

seriously i cant even grok the math behind those things, those are way beyond me.

but to be fair, hn posts and comments are great so far in coming up with possible use cases for these things.

Subscribe to this guy's channel: https://www.youtube.com/c/YannicKilcher

created few bookmark to search latest content on HN, youtube, reddit, github, huggingface, twitter.

also writing code and prompts as side gig. accumulated over 40gb of trained models and datasets from various sources.

Can you please share those bookmarks? I can search hn.algolia.com but don't know much about who to follow on other platforms.

Don’t worry about it unless you are joining one of the hundreds of hyped AI startups that will go broke the moment they run out of VC money.

Don't worry about it. Anything that you need to know won't be in the news after a decade or so - it'll be in textbooks.

Why do you need to keep up with it? Asking as someone who’s paid to keep up with it.

Mostly HN and some newsletters:

- https://aiweekly.co/

- https://tldr.tech/

Reply on news.ycombinator.com