Learning to Reason with LLMs • Hacker News Board

Re: Learning to Reason with LLMs

#61

Post by ffsm8 » Thu, Sep 12, 2024, 5:30 PM UTC

modeless wrote: post #33

> We have found that the performance of o1 consistently improves with more reinforcement learning (train-time compute) and with more time spent thinking (test-time compute). Wow. So we can expect scaling to continue after all. Hyperscalers feeling pretty good about their big bets right now. Jensen is smiling. This is the most important thing. Performance today matters less than the scaling laws. I think everyone has…

It'd be interesting for sure if true. Gotta remember that this is a marketing post though, let's wait a few months and see if its actually true. Things are definitely interesting, wherever these techniques will get us AGI or not

Re: Learning to Reason with LLMs

#62

Post by andrewla » Thu, Sep 12, 2024, 5:30 PM UTC

This is something that people have toyed with to improve the quality of LLM responses. Often instructing the LLM to "think about" a problem before giving the answer will greatly improve the quality of response. For example, if you ask it how many letters are in the correctly spelled version of a misspelled word, it will first give the correct spelling, and then the number (which is often correct). But if you instruct it to only give the number the accuracy is greatly reduced.

I like the idea too that they turbocharged it by taking the limits off during the "thinking" state -- so if an LLM wants to think about horrible racist things or how to build bombs or other things that RLHF filters out that's fine so long as it isn't reflected in the final answer.

Re: Learning to Reason with LLMs

#63

Post by jazzyjackson » Thu, Sep 12, 2024, 5:30 PM UTC

Dang, I just payed out for Kagi Assistant.

Using Claude 3 Opus I noticed it performs and while browsing the web for me. I don't guess that's a change in the model for doing reasoning.

Re: Learning to Reason with LLMs

#64

Post by afruitpie » Thu, Sep 12, 2024, 5:30 PM UTC

crakenzak wrote: post #28

> we are releasing an early version of this model, OpenAI o1-preview, for immediate use in ChatGPT Awesome!

Rate limited to 30 messages per week for ChatGPT Plus subscribers at launch: https://openai.com/index/introducing-openai-o1-preview/

Re: Learning to Reason with LLMs

#65

Post by levzzz » Thu, Sep 12, 2024, 5:30 PM UTC

[dead]

Re: Learning to Reason with LLMs

#66

Post by joshribakoff » Thu, Sep 12, 2024, 5:31 PM UTC

awesomeMilou wrote: post #57

rfw300 wrote: post #39

A lot of skepticism here, but these are astonishing results! People should realize we’re reaching the point where LLMs are surpassing humans in any task limited in scope enough to be a “benchmark”. And as anyone who’s spent time using Claude 3.5 Sonnet / GPT-4o can attest, these things really are useful and smart! (And, if these results hold up, O1 is much, much smarter.) This is a nerve-wracking time to be a knowled…

Is it? They talk about 10k attempts to reach gold medal status in the mathematics olympiad, but zero shot performance doesn't even place it in the upper 50th percentile. Maybe I'm confused but 10k attempts on the same problem set would make anyone an expert in that topic? It's also weird that zero shot performance is so bad, but over a lot of attempts it seems to get correct answers? Or is it learning from previous a…

The correct metaphor is that 10,000 attempts would allow anyone to cherry pick a successful attempt. You’re conflating cherry picking with online learning. This is like if an entire school of students randomized their answers on a multiple choice test, and then you point to someone who scored 100% and claim it is proof of the school’s expertise.

Re: Learning to Reason with LLMs

#67

Post by achrono » Thu, Sep 12, 2024, 5:31 PM UTC

sk11001 wrote: post #34

gliiics wrote: post #20

Congrats to OpenAI for yet another product that has nothing to do with the word "open"

And Apple's product line this year? Phones. Nothing to do with fruit. Almost 50 years of lying to people. Names should mean something!

Did Apple start their company by saying they will be selling apples?

Re: Learning to Reason with LLMs

#68

Post by deisteve » Thu, Sep 12, 2024, 5:31 PM UTC

bigstrat2003 wrote: post #45

rfw300 wrote: post #39

A lot of skepticism here, but these are astonishing results! People should realize we’re reaching the point where LLMs are surpassing humans in any task limited in scope enough to be a “benchmark”. And as anyone who’s spent time using Claude 3.5 Sonnet / GPT-4o can attest, these things really are useful and smart! (And, if these results hold up, O1 is much, much smarter.) This is a nerve-wracking time to be a knowled…

I cannot, in fact, attest that they are useful and smart. LLMs remain a fun toy for me, not something that actually produces useful results.

'Not useful' is a pretty low bar to clear, especially when you consider the state of the art just 5 years ago. LLMs may not be solving world hunger, but they're already being used in production for coding

If you're not seeing value in them, maybe it's because you're not looking at the right problems. Or maybe you're just not using them correctly. Either way, dismissing an entire field of research because it doesn't fit your narrow use case is pretty short-sighted.

FWIW, I've been using LLMs to generate production code and it's saved me weeks if not months. YMMV, I guess

Re: Learning to Reason with LLMs

#69

Post by orbital-decay » Thu, Sep 12, 2024, 5:31 PM UTC

Wait, are they comparing 4o without CoT and o1 with built-in CoT?

Re: Learning to Reason with LLMs

#70

Post by kickofline » Thu, Sep 12, 2024, 5:31 PM UTC

LLM performance, recently, seemingly hit the top of the S-curve. It remains to be seen if this is the next leap forward or just the rest of that curve.