Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Learning to Reason with LLMs

openai.com

Post reply on HN
Print view View original

941–950 of 1001 posts

Re: Learning to Reason with LLMs

#941

Post by qudat »

mountainriver wrote: post #542
utdiscant wrote: post #326

Feels like a lot of commenters here miss the difference between just doing chain-of-thought prompting, and what is happening here, which is learning a good chain of thought strategy using reinforcement learning. "Through reinforcement learning, o1 learns to hone its chain of thought and refine the strategies it uses." When looking at the chain of thought (COT) in the examples, you can see that the model employs diffe…

It’s basically a scaled Tree of Thoughts

In the primary CoT research paper they discuss figuring out how to train models using formal languages instead of just natural ones. I'm guessing this is one piece to the model learning tree-like reasoning.

Based on the quick searching it seems like they are using RL to provide positive/negative feedback on which "paths" to choose when performing CoT.

Re: Learning to Reason with LLMs

#942

Post by fwip »

rcarmo wrote: post #562
fwip wrote: post #557

Earlier quoted context omitted.

I suspect that actually reading the "chain of thought" would reveal obvious "logic" errors embarrassingly often.

It would still be auditable. In a few industries that is the only blocker for adoption--even if the outputs are incorrect.

Oh, perhaps. I mean that OpenAI won't do it because it would be bad for business and pop the AI bubble early.

Re: Learning to Reason with LLMs

#943

Post by whynotminot »

fsloth wrote: post #446
hakanderyal wrote: post #418

Earlier quoted context omitted.

Software engineering contains a lot more than just writing code. If we somehow get AGI, it'll change everything, not just SWE. If not, my belief is that there will be a lot more demand for good SWEs to harness the power of LLMs, not less. Use them to get better at it faster.

Agree, SWE as a profession is not going anywhere, unless we AGI, and that would mean all the rules change anyway. Actually now is really good time to get to SWE. The craft contains lots of pointless cruft that LLM:s cut through like knife through hot butter. I’m actually enjoying my job now more than ever since I dont’t need to pretend to like the abysmal tools the industry forces on us (like git), and can focus most…

I'd agree with this take. Everyone is so pessimistic about LLMs, but I've really enjoyed this new era.

A lot of the tasks that used to take considerable time are so much faster and less tedious now. It still puts a smile on my face to tell an LLM to write me scripts that do X Y and Z. Or hand it code and ask for unit tests.

And I feel like I'm more likely to reach for work that I might otherwise shrink from / outside my usual comfort zone, because asking questions of an LLM is just so much better than doing trivial beginner tutorials or diving through 15 vaguely related stack overflow questions (I wonder if SO has seen any significant dip in traffic over the last year).

Most people I've seen disappointed with these tools are doing way more advanced work than I appear to be doing in my day to day work. They fail me too here and there, but more often than not I'm able to get at least something helpful or useful out of them.

Re: Learning to Reason with LLMs

#944

Post by shthed »

gradus_ad wrote: post #117

Interesting sequence from the Cipher CoT: Third pair: 'dn' to 'i' 'd'=4, 'n'=14 Sum:4+14=18 Average:18/2=9 9 corresponds to 'i'(9='i') But 'i' is 9, so that seems off by 1. So perhaps we need to think carefully about letters. Wait, 18/2=9, 9 corresponds to 'I' So this works. ----- This looks like recovery from a hallucination. Is it realistic to expect CoT to be able to recover from hallucinations this quickly?

Seems like a huge waste of tokens for it to try to work all this out manually, as soon as it came up with the decipher algorithm it should realise it can write some code to execute.

Re: Learning to Reason with LLMs

#946

Post by WanderPanda »

cs391231 wrote: post #408

Student here. Can someone give me one reason why I should continue in software engineering that isn't denial and hopium?

If the writing & arts vs. doing laundry & cleaning dishes is any indication, it does not look rosy. All the fun and rewarding parts (low hanging fruits / quick wins) of coding might be automated. What remains are probably things like debugging race conditions in legacy systems and system administration etc.

Re: Learning to Reason with LLMs

#947

Post by barrell »

p-e-w wrote: post #935
arnaudsm wrote: post #825

Earlier quoted context omitted.

Some of the queries run for multiple minutes. 40 tokens/sec is too slow for CoT. I hope OpenAI is investing in low-latency like Groq's tech that can reach 1k tokens/sec.

It's slow and expensive if you compare it with other LLMs. It's lightning fast and dirt cheap if you compare it to consulting with a human expert, which it appears to be competitive with.

I would say consulting with a human. Any expert who has a conversation with chatGPT about their field will verify that it is very far from expert

Re: Learning to Reason with LLMs

#949

Post by slashdave »

patapong wrote: post #618
slashdave wrote: post #384

Earlier quoted context omitted.

> My expectations would be that a model that is better at one thing is better across the board. No, it's the opposite. This is simply a function of resources applied during training.

To some extent I agree, but until now all of the big jumps (GPT2 -> GPT3 -> GPT4) have meant significant improvements across all tasks. This does not seem to be the case here, this model seems to be vastly stronger on certain tasks but not much of an improvement on other tasks. Maybe we will have to wait for GPT5 for that :)

> (GPT2 -> GPT3 -> GPT4) have meant significant improvements

These were all train the same way. It's fairly clear that o1 was not.

> Maybe we will have to wait for GPT5 for that :)

There will be no GPT5, for the simple reason that scaling has reached a limit and there is no more text data to train on.

Re: Learning to Reason with LLMs

#950

Post by jstummbillig »

baq wrote: post #662
activatedgeek wrote: post #610

Earlier quoted context omitted.

Reasoning tokens are indeed billed as output tokens. > While reasoning tokens are not visible via the API, they still occupy space in the model's context window and are billed as output tokens. From here: https://platform.openai.com/docs/guides/reasoning

This is concerning - how do you know you aren’t being fleeced out of your money here…? You’ll get your results, but did you really use that much?

I think it's fantastic that now, for very little money, everyone gets to share a narrow but stressful subset of what it feels like to employ other people.

Really, I recommend reading this part of the thread while thinking about the analogy. It's great.

Post reply on HN