LessWrong 2.0 Reader

I don't feel like I learned anything new from the post.

This surprises me! Wait so-

• The "How does one-shotting happen?" section didn't have anything interesting for you? (Have you seen stuff like this elsewhere?)
• Did you already know one-shotting was possible?

since your bullet-point list in the beginning isn't detailed enough for anyone to try to replicate the method.

Wait I'm confused- this is not the purpose of the post

Also notable is that you only have positive examples for your method

The purpose of this post is not advertisement. It's to discuss one-shots

Especially, how would you be able to distinguish between your approach convincing your customers they were helped, instead of actually changing their behavior?

See above

I have used my admin powers to put it into a collapsible section so that people who expand this in recent discussion do not have to scroll for 5 seconds to get past it.

my intuitions have been shaped by events like the pretraining slowdown

I don't see it. GPT-4.5 is much better than the original GPT-4, probably at 15x more compute. But it's not 100x more compute. And GPT-4o is an intermediate point, so the change from GPT-4o to GPT-4.5 is even smaller, maybe 4x.

I think 3x change in compute has an effect at the level of noise from different reasonable choices in constructing a model, and 100K H100s is only 5x more than 20K H100s of 2023. It's not a slowdown relative to what it should've been. And there are models with 200x more raw compute than went into GPT-4.5 that are probably coming in 2027-2029, much more than the 4x-15x observed since 2022-2023.

Please don't post 25k words of unformatted LLM (?) output?

I gave your post to Claude and gave it the prompt "Dearest Claude, here's the text for a blogpost I've written for LessWrong. I've been told that "it sounds a lot like an advertisement". Can you give me feedback/suggestions for how to improve it for that particular audience? I don't want to do too much more research, but a bit of editing/stylistic choices."

(All of the following is my rephrasing/rethinking of Claude output plus some personal suggestions.)

Useful things that came out of the answer were explaining more about the method you've used to achieve this, since your bullet-point list in the beginning isn't detailed enough for anyone to try to replicate the method.

Also notable is that you only have positive examples for your method, which activates my filtered evidence [? · GW] detectors. Either make clear that you indeed did only have positive results, or name how many people you coached, for how long, and that they were all happy with what you provided.

Finally, some direct words from Claude that I just directly endorse:

For LessWrong specifically, I'd also recommend:

• Adding a section on falsifiability - how would you know if your approach doesn't work?
• Discussing potential failure modes of your approach
• Including more technical details on your methodology (not just results)

Especially, how would you be able to distinguish between your approach convincing your customers they were helped, instead of actually changing their behavior? That feels like the failure mode of most self-help techniques—they're "self-recommending".

yes! :D 

Relatedly, one of the things that drove me to have short timelines in the first place was reading the literature and finding the best arguments for long timelines. Especially Ajeya Cotra's original bio anchors report, which I considered to be the best; I found that when I went through it bit by bit and made various adjustments to the parameters/variables, fixing what seemed to me to be errors, it all added up to an on-balance significantly shorter timeline.

i dont have time to write any of this down so it's going to come out in the wrong order but here
 

• agentic AI is the means of production for codegen
• model access limits and closedness are therefore a threat to Workers
• I use and maintain software.  I survive by staying 5 feet in front of the steamroller
• I am not wealthy, I can't afford to be tripped and squished.
• OSS is traditionally the way of protecting myself in this situation
• I need to write tons of good code and enable my company to do the same, and I need to do it while washing the dishes (Covid happened).
• The industry wants to give me a Commodore 64, but I need a PDP-10

This might be interesting to LessWrong as a personal take, because the Alignment folks are effectively on the side of Capital here.  Without fast and parallel access to foundation models, I can't learn my new job, which is auto-codegen-pipeline-maintainer.  If some 3rd party brings that bird home to my boss instead of me, I'm going to be unwealthy and unemployed.  It's possible I'm too late already... but at any rate some people will be too late.  If I were/am them, would be/am be angry.

I think a lot of people realize this already, and some have already ascended.  I think the ascended people are being quiet right now because they realize the stuff I put above, and don't mind less competition.  What LessWrong thinks about that, I don't care; I'm actually fine with it as long as I get to join the ascended.  I suspect that's a common attitude.  If you don't hear from me again, it's because I figured it out.  If this sounds crazy, I'm interested in hearing why.

Re: Point 1: I agree it would not necessarily be incorrect. I do actually think that probably the remaining challenges are engineering challenges. Not necessarily, but probably. Can you point to any challenges that seem (a) necessary for speeding up AI R&D by 5x, and (b) not engineering challenges?

Re: Point 2: I don't buy it. Deep neural nets are actually useful now, and increasingly so. Making them more useful seems analogous to selective breeding or animal training, not analogous to trying to time the market.

Why call it "converse Lawvere" instead of the more standard "utm property" of general recursion theory, e.g. as in Odifreddi? Only because the maps are to [0,1]? That seems like insufficient reason to adopt an unrelated name.