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

801–810 of 1001 posts

Re: Learning to Reason with LLMs

#801

Post by coder543 »

fshbbdssbbgdd wrote: post #544
jazzyjackson wrote: post #448

Earlier quoted context omitted.

Isn't there a big "Share" button at the top right of the chatgpt interface? Or are you using another front end?

I often click on those links and get an error that they are unavailable. I’m not sure if it’s openAI trying to prevent people from sharing evidence of the model behaving badly, or an innocuous explanation like the links are temporary.

The link also breaks if the original user deletes the chat that was being linked to, whether on purpose or without realizing it would also break the link.

Re: Learning to Reason with LLMs

#802

Post by kgeist »

Asked it to write PyTorch code which trains an LLM and it produced 23 steps in 62 seconds.

With gpt4-o it immediately failed with random errors like mismatched tensor shapes and stuff like that.

The code produced by gpt-o1 seemed to work for some time but after some training time it produced mismatched batch sizes. Also, gpt-o1 enabled cuda by itself while for gpt-4o, I had to specifically spell it out (it always used cpu). However, showing gpt-o1 the error output resulted in broken code again.

I noticed that back-and-forth iteration when it makes mistakes has worse experience because now there's always 30-60 sec time delays. I had to have 5 back-and-forths before it produced something which does not crash (just like gpt-4o). I also suspect too many tokens inside the CoT context can make it accidentally forget some stuff.

So there's some improvement, but we're still not there...

Re: Learning to Reason with LLMs

#803

Post by koreth1 »

The performance on programming tasks is impressive, but I think the limited context window is still a big problem.

Very few of my day-to-day coding tasks are, "Implement a completely new program that does XYZ," but more like, "Modify a sizable existing code base to do XYZ in a way that's consistent with its existing data model and architecture." And the only way to do those kinds of tasks is to have enough context about the existing code base to know where everything should go and what existing patterns to follow.

But regardless, this does look like a significant step forward.

Re: Learning to Reason with LLMs

#804

Post by obmelvin »

losvedir wrote: post #383
bartman wrote: post #367

This is incredible. In April I used the standard GPT-4 model via ChatGPT to help me reverse engineer the binary bluetooth protocol used by my kitchen fan to integrate it into Home Assistant. It was helpful in a rubber duck way, but could not determine the pattern used to transmit the remaining runtime of the fan in a certain mode. Initial prompt here [0] I pasted the same prompt into o1-preview and o1-mini and both c…

Wow, that is impressive! How were you able to use o1-preview? I pay for ChatGPT, but on chatgpt.com in the model selector I only see 4o, 4o-mini, and 4. Is o1 in that list for you, or is it somewhere else?

The linked release mentions trusted users and links to the usage tier limits. Looking at the pricing, o1-preview only appears for tier 5 - requiring 1k+ spend and initial spend 30+ days ago

edit: sorry - this is for API :)

Re: Learning to Reason with LLMs

#806

Post by hobo_in_library »

Buttons840 wrote: post #752
kmeisthax wrote: post #622

>We believe that a hidden chain of thought presents a unique opportunity for monitoring models. Assuming it is faithful and legible, the hidden chain of thought allows us to "read the mind" of the model and understand its thought process. For example, in the future we may wish to monitor the chain of thought for signs of manipulating the user. However, for this to work the model must have freedom to express its thoug…

I always laughed at the idea of a LLM Skynet "secretly" plotting to nuke humanity, while a bunch of humans watch it unfold before their eyes in plaintext. Now that seems less likely. At least OpenAI can see what it's thinking. A next step might be allowing the LLM to include non-text-based vectors in its internal thoughts, and then do all internal reasoning with raw vectors. Then the LLMs will have truly private thou…

"...either way we're going to need more GPUs." posted the LLM, rubbing it's virtual hands, cackling with delight as it prodded the humans to give it MOAR BRAINS

Re: Learning to Reason with LLMs

#807

Post by closeparen »

spaceman_2020 wrote: post #479
canjobear wrote: post #427

First shot, I gave it a medium-difficulty math problem, something I actually wanted the answer to (derive the KL divergence between two Laplace distributions). It thought for a long time, and still got it wrong, producing a plausible but wrong answer. After some prodding, it revised itself and then got it wrong again. I still feel that I can't rely on these systems.

Look where you were 3 years ago, and where you are now. And then imagine where you will be in 5 more years. If it can almost get a complex problem right now, I'm dead sure it will get it correct within 5 years

It is not at all clear that "produce correct answer" is the natural endpoint of "produce plausible on-topic utterances that look like they could be answers." To do the former you need to know something about the underlying structure of reality (or have seen the answer before), to do the latter you only need to be good at pattern-matching and language.

Re: Learning to Reason with LLMs

#808

Post by lolinder »

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?

Also, now we're paying for output tokens that aren't even output, with no good explanation for why these tokens should be hidden from the person who paid for them.

Re: Learning to Reason with LLMs

#809

Post by JieJie »

OutOfHere wrote: post #781
jazzyjackson wrote: post #448

Earlier quoted context omitted.

Isn't there a big "Share" button at the top right of the chatgpt interface? Or are you using another front end?

Even for regular users, the Share button is not always available or functional. It works sometimes, and other times it disappears. For example, since today, I have no Share button at all for chats.

My share chat link moved into the sidebar in the … menu to the right of each chat title (MacOS Safari).

Re: Learning to Reason with LLMs

#810

Post by navigate8310 »

Buttons840 wrote: post #752
kmeisthax wrote: post #622

>We believe that a hidden chain of thought presents a unique opportunity for monitoring models. Assuming it is faithful and legible, the hidden chain of thought allows us to "read the mind" of the model and understand its thought process. For example, in the future we may wish to monitor the chain of thought for signs of manipulating the user. However, for this to work the model must have freedom to express its thoug…

I always laughed at the idea of a LLM Skynet "secretly" plotting to nuke humanity, while a bunch of humans watch it unfold before their eyes in plaintext. Now that seems less likely. At least OpenAI can see what it's thinking. A next step might be allowing the LLM to include non-text-based vectors in its internal thoughts, and then do all internal reasoning with raw vectors. Then the LLMs will have truly private thou…

At this point the G in GPU must be completely dropped
Post reply on HN