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

Notes on OpenAI's new o1 chain-of-thought models

simonwillison.net

Post reply on HN
Print view View original

51–60 of 659 posts

Re: Notes on OpenAI's new o1 chain-of-thought models

#51

Post by shmatt »

layer8 wrote: post #18

The o1-preview model still hallucinates non-existing libraries and functions for me, and is quickly wrong about facts that aren't well-represented on the web. It's the usual string of "You're absolutely correct, and I apologize for the oversight in my previous response. [Let me make another guess.]" While the reasoning may have been improved, this doesn't solve the problem of the model having no way to assess if what…

I honestly can’t believe this is the hyped up “strawberry” everyone was claiming is pretty much AGI. Senior employees leaving due to its powers being so extreme

I’m in the “probabilistic token generators aren’t intelligence” camp so I don’t actually believe in AGI, but I’ll be honest the never ending rumors / chatter almost got to me

Remember, this is the model some media outlet reported recently that is so powerful OAI is considering charging $2k/month for

Re: Notes on OpenAI's new o1 chain-of-thought models

#52

Post by deepsquirrelnet »

CuriouslyC wrote: post #42
deepsquirrelnet wrote: post #27

It kind of seems like they just wrote a generalized DSPy program. Can anyone confirm? This has been a very incremental year for OpenAI. If this is what it seems like, then I’ve got to believe they’re stalling for time.

DSPy doesn't do that, you could describe it as a langchain style agent that evaluates its own output though it's better/faster than that. OpenAI is definitely trying to run a hype game to keep the ball rolling. They're burning cash too quickly given their monetization path though, so I think they're going to end up completely in Microsoft's pocket.

It seems pretty close to the multihop QA example in their documentation[1]. I’d imagine you could adapt this to do something similar with more generic constructs.

[1] https://dspy-docs.vercel.app/docs/tutorials/simplified-balee...

Re: Notes on OpenAI's new o1 chain-of-thought models

#53

Post by throwup238 »

tptacek wrote: post #46
layer8 wrote: post #23

Earlier quoted context omitted.

My experience is likely colored by the fact that I tend to turn to LLMs for problems I have trouble solving by myself. I typically don't use them for the low-hanging fruits. That's the frustrating thing. LLMs don't materially reduce the set of problems where I'm running against a wall or have trouble finding information.

I use LLMs for three things: * To catch passive voice and nominalizations in my writing. * To convert Linux kernel subsystems into Python so I can quickly understand them (I'm a C programmer but everyone reads Python faster). * To write dumb programs using languages and libraries I haven't used much before; for instance, I'm an ActiveRecord person and needed to do some SQLAlchemy stuff today, and GPT 4o (and o1) kept…

> * To convert Linux kernel subsystems into Python so I can quickly understand them (I'm a C programmer but everyone reads Python faster).

Do you have an example chat of this output? Sounds interesting. Do you just dump the C source code into the prompt and ask it to convert to Python?

Re: Notes on OpenAI's new o1 chain-of-thought models

#54

Post by COAGULOPATH »

layer8 wrote: post #18

The o1-preview model still hallucinates non-existing libraries and functions for me, and is quickly wrong about facts that aren't well-represented on the web. It's the usual string of "You're absolutely correct, and I apologize for the oversight in my previous response. [Let me make another guess.]" While the reasoning may have been improved, this doesn't solve the problem of the model having no way to assess if what…

Yes, this only helps multi-step reasoning. The model still has problems with general knowledge and deep facts.

There's no way you can "reason" a correct answer to "list the tracklisting of some obscure 1991 demo by a band not on Wikipedia." You either know or you don't.

I usually test new models with questions like "what are the levels in [semi-famous PC game from the 90s]?" The release version of GPT-4 could get about 75% correct. o1-preview gets about half correct. o1-mini gets 0% correct.

Fair enough. The GPT-4 line aren't meant to be search engines or encyclopedias. This is still a useful update though.

Re: Notes on OpenAI's new o1 chain-of-thought models

#55

Post by firejake308 »

shmatt wrote: post #51
layer8 wrote: post #18

The o1-preview model still hallucinates non-existing libraries and functions for me, and is quickly wrong about facts that aren't well-represented on the web. It's the usual string of "You're absolutely correct, and I apologize for the oversight in my previous response. [Let me make another guess.]" While the reasoning may have been improved, this doesn't solve the problem of the model having no way to assess if what…

I honestly can’t believe this is the hyped up “strawberry” everyone was claiming is pretty much AGI. Senior employees leaving due to its powers being so extreme I’m in the “probabilistic token generators aren’t intelligence” camp so I don’t actually believe in AGI, but I’ll be honest the never ending rumors / chatter almost got to me Remember, this is the model some media outlet reported recently that is so powerful…

I mean, considering how many tokens their example prompt consumed, I wouldn't be surprised if it costs ~$2k/month/user to run

Re: Notes on OpenAI's new o1 chain-of-thought models

#56

Post by mediaman »

faangguyindia wrote: post #50
layer8 wrote: post #18

The o1-preview model still hallucinates non-existing libraries and functions for me, and is quickly wrong about facts that aren't well-represented on the web. It's the usual string of "You're absolutely correct, and I apologize for the oversight in my previous response. [Let me make another guess.]" While the reasoning may have been improved, this doesn't solve the problem of the model having no way to assess if what…

>The o1-preview model still hallucinates non-existing libraries and functions for me, and is quickly wrong about facts that aren't well-represented on the web. It's the usual string of "You're absolutely correct, and I apologize for the oversight in my previous response. [Let me make another guess.]" After that you switch to Claude Soñnet and after sometime it also gets stuck. Problem with LLM is that they are not aw…

One trick that people are using, when using Cursor and specifically Cursor's compose function, is to dump library docs into a text file in your repo, and then @ that doc file when you're asking it to do something involving that library.

That seems to eliminate a lot of the issues, though it's not a seamless experience, and it adds another step of having to put the library docs in a text file.

Alternatively, cursor can fetch a web page, so if there's a good page of docs you can bring that in by @ the web page.

Eventually, I could imagine LLMs automatically creating library text doc files to include when the LLM is using them to avoid some of these problems.

It could also solve some of the issues of their shaky understanding of newer frameworks like SvelteKit.

Re: Notes on OpenAI's new o1 chain-of-thought models

#57

Post by tptacek »

throwup238 wrote: post #53
tptacek wrote: post #46

Earlier quoted context omitted.

I use LLMs for three things: * To catch passive voice and nominalizations in my writing. * To convert Linux kernel subsystems into Python so I can quickly understand them (I'm a C programmer but everyone reads Python faster). * To write dumb programs using languages and libraries I haven't used much before; for instance, I'm an ActiveRecord person and needed to do some SQLAlchemy stuff today, and GPT 4o (and o1) kept…

> * To convert Linux kernel subsystems into Python so I can quickly understand them (I'm a C programmer but everyone reads Python faster). Do you have an example chat of this output? Sounds interesting. Do you just dump the C source code into the prompt and ask it to convert to Python?

No, ChatGPT is way cooler than that. It's already read every line of kernel code ever written. I start with a subsystem: the device mapper is a good recent example. I ask things like "explain the linux device mapper. if it was a class in an object-oriented language, what would its interface look like?" and "give me dm_target as a python class". I get stuff like:

    def linear_ctr(target, argc, argv):
        print("Constructor called with args:", argc, argv)
        # Initialize target-specific data here
        return 0
     
    def linear_dtr(target):
        print("Destructor called")
        # Clean up target-specific data here
     
    def linear_map(target, bio):
        print("Mapping I/O request")
        # Perform mapping here
        return 0
     
    linear_target = DmTarget(name="linear", version=(1, 0, 0), module="dm_mod")
    linear_target.set_ctr(linear_ctr)
    linear_target.set_dtr(linear_dtr)
    linear_target.set_map(linear_map)
     
    info = linear_target.get_info()
    print(info)
(A bunch of stuff elided). I don't care at all about the correctness of this code, because I'm just using it as a roadmap for the real Linux kernel code. The example use case code is an example of something GPT 4o provides that I didn't even know I wanted.

Re: Notes on OpenAI's new o1 chain-of-thought models

#58

Post by deisteve »

lukev wrote: post #43

It's interesting to note that there's really two things going on here: 1. A LLM (probably a finetuned GPT-4o) trained specifically to read and emit good chain-of-thought prompts. 2. Runtime code that iteratively re-prompts the model with the chain of thought so far. This sounds like it includes loops, branches and backtracking. This is not "the model", it's regular code invoking the model. Interesting that OpenAI is…

this is why i became skeptical of openai's claims

if they shared the COT the grift wont work

its just RL

Re: Notes on OpenAI's new o1 chain-of-thought models

#59

Post by salomonk_mur »

lukev wrote: post #43

It's interesting to note that there's really two things going on here: 1. A LLM (probably a finetuned GPT-4o) trained specifically to read and emit good chain-of-thought prompts. 2. Runtime code that iteratively re-prompts the model with the chain of thought so far. This sounds like it includes loops, branches and backtracking. This is not "the model", it's regular code invoking the model. Interesting that OpenAI is…

The innovation lies in making the whole loop available to an end user immediately, without them being a programmer. My grandma can build games using ChatGPT now.

Re: Notes on OpenAI's new o1 chain-of-thought models

#60

Post by airstrike »

I've just wasted a few rounds of my weekly o1 ammo by feeding it hard problems I have been working on over the last couple days and for which GPT-4o had failed spectacularly.

I suppose I'm to blame for raising my own expectations after the latest PR, but I was pretty disappointed when the answers weren't any better from what I got with the old model. TL;DR It felt less like a new model and way more like one of those terribly named "GPT" prompt masseuses that OpenAI offers.

Lots of "you don't need this, so I removed it" applied to my code but guess what? I did need the bits you deleted, bro.

It felt as unhelpful and bad at instructions as GPT-4o. "I'm sorry, you're absolutely right". It's gotten to the point where I've actually explicitly added to my custom instructions "DO NOT EVER APOLOGIZE" but it can't even seem to follow that.

Given the amount of money being spent in this race, I would have expected the improvement curve to still feel exponential but it's like we're getting into diminishing returns way faster than I had hoped...

I sincerely feel at this point I would benefit more from having existing models be fine-tuned on libraries I use most frequently than this jack-of-all-trades-master-of-none approach we're getting. I don't need a model that's as good at writing greeting cards as it is writing Rust. Just give me one of the two.

Post reply on HN