Thanks for this—helpful and concrete, and did change my mind somewhat. Of course, if it really is just 10x, in terms of orders of magnitude/hyper fast scaling we are pretty close to the wall.
JustisMills
Karma: 758
How Big a Deal are MatMul-Free Transformers?
Week One of Studying Transformers Architecture
Mostly just public text, I think. But I’m not sure how much more you get out of e.g. video transcripts. Maybe a lot! But it wouldn’t surprise me if that was notably worse as a source.
Whoops! Thank you, fixed.
The Data Wall is Important
Maybe worth a slight update on how the AI alignment community would respond? Doesn’t seem like any of the comments on this post are particularly aggressive. I’ve noticed an effect where I worry people will call me dumb when I express imperfect or gestural thoughts, but it usually doesn’t happen. And if anyone’s secretly thinking it, well, that’s their business!
I think self-critique runs into the issues I describe in the post, though without insider information I’m not certain. Naively it seems like existing distortions would become larger with self-critique, though.
For human rating/RL, it seems true that it’s possible to be sample efficient (with human brain behavior as an existence proof), but as far as I know we don’t actually know how to make it sample efficient in that way, and human feedback in the moment is even more finite than human text that’s just out there. So I still see that taking longer than, say, self play.I agree that if outcome-based RL swamps initial training run datasets, then the “playing human roles” section is weaker, but is that the case now? My understanding (could easily be wrong) is that RLHF is a smaller postprocessing layer that only changes models moderately, and nowhere near the bulk of their training.
LLMs seem (relatively) safe
AI Safety Concepts Writeup: WebGPT
Consider Multiclassing
Alignment Risk Doesn’t Require Superintelligence
I journal! It’s a good way to write at least something daily, and often also feels like a good avenue for healthy introspection.
I wrote a reply to this from a more-peripheral-EA perspective on the EA forum here:
https://forum.effectivealtruism.org/posts/YeudcYiArwWrg77Ng/notes-from-a-pledger
Thank you!
My pleasure!
Yeah, that critique is part of why “use more links” is among my least confident advice of the stuff in this post. I like links mostly as an alternative to nothing—if there’s a term of background that ideally your readers should already know, a link is an economical way to give readers below your target audience in background knowledge a leg up. But for really central terms, yeah, better to summarize in your own words.
Yeah, that’s a good pithy summary! I often suggest replacing “this” with “this [x]”.
Yeah, could cut both ways for this I think? On the one hand, if no-MatMul models really are more efficient in the long run, you could probably make custom hardware optimized for the stuff they require (e.g. lots of ternary stuff). But getting there from the ASICs currently in development would be a necessary pivot.
Maybe the race dynamics actually help slow things down here? Since nobody wants to pivot and fall temporarily behind; money might dry up or someone else might get there before the investment pays off and you leapfrog.
But yeah, even in the medium run, as constraints start to flare up, probably ASICs are a factor in changing up architectures.