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

281–290 of 659 posts

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

#281

Post by bambax »

Near the end, the quote from OpenAI researcher Jason Wei seems damning to me:

> Results on AIME and GPQA are really strong, but that doesn’t necessarily translate to something that a user can feel. Even as someone working in science, it’s not easy to find the slice of prompts where GPT-4o fails, o1 does well, and I can grade the answer. But when you do find such prompts, o1 feels totally magical. We all need to find harder prompts.

Results are "strong" but can't be felt by the user? What does that even mean?

But the last sentence is the worst: "we all need to find harder prompts". If I understand it correctly, it means we should go looking for new problems / craft specific questions that would let these new models shine.

"This hammer hammers better, but in most cases it's not obvious how better it is. But when you stumble upon a very specific kind of nail, man does it feel magical! We need to craft more of those weird nails to help the world understand the value of this hammer."

But why? Why would we do that? Wouldn't our time be better spent trying to solve our actual, current problems, using any tool available?

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

#282

Post by multjoy »

ramraj07 wrote: post #112
pedrosorio wrote: post #110

Earlier quoted context omitted.

> It knows english at or above a level equal to most fluent speakers, and it also can produce output that is not just a likely output, but is a logical output This is not an apt description of the system that insists the doctor is the mother of the boy involved in a car accident when elementary understanding of English and very little logic show that answer to be obviously wrong. https://x.com/colin_fraser/status/183…

Many of my PhD and post doc colleagues who emigrated from Korea, China and India who didn’t have English as the medium of instruction would struggle with this question. They only recover when you give them a hint. They’re some of the smartest people in general. If you try to stop stumping these models with trick questions and ask it straightforward reasoning systems it is extremely performant (O1 is definitely a step…

“Don’t be mean to LLMs, it isn’t their fault that they’re not actually intelligent”

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

#283

Post by ssl-3 »

fragmede wrote: post #240

I posted this on the other thread, but the two tests I had, it passed when ChatGPT-4 failed. https://chatgpt.com/share/66e35c37-60c4-8009-8cf9-8fe61f57d3... https://chatgpt.com/share/66e35f0e-6c98-8009-a128-e9ac677480...

The farmer riddle isn't quite right as you presented it. One of the parts that makes it interesting is that the boat can't carry everything at one time[1]. It can't happen in one trip; something must be left behind.

It solved the correct version fine: https://chatgpt.com/share/66e3f9bb-632c-8005-9c95-142424e396...

1: https://en.wikipedia.org/wiki/Wolf,_goat_and_cabbage_problem

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

#284

Post by carlmr »

arthurcolle wrote: post #213
aktuel wrote: post #195

Earlier quoted context omitted.

I have to say, having to tell it to ask me clarifying questions DOES make it really look smart!

imagine if you make it keep going without having to reprompt it

Isn't that the exact point of o1, that it has time to think for itself without reprompting?

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

#285

Post by thrdbndndn »

bambax wrote: post #281

Near the end, the quote from OpenAI researcher Jason Wei seems damning to me: > Results on AIME and GPQA are really strong, but that doesn’t necessarily translate to something that a user can feel. Even as someone working in science, it’s not easy to find the slice of prompts where GPT-4o fails, o1 does well, and I can grade the answer. But when you do find such prompts, o1 feels totally magical. We all need to find…

> What does that even mean?

It explicitly says "Results on AIME and GPQA are really strong". So I would assume it means it can get (statistically significantly, I assume) better score in AIME and GPQA benchmarks compared to 4o.

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

#286

Post by johndough »

gloosx wrote: post #244
c-fe wrote: post #208

Earlier quoted context omitted.

With multiplication the question makes sense due to the commutative property but division does not have that so the question becomes ambiguous... And now I see that the model even points this out.

There is no ambiguity, the problem is that three numbers, divided together, without the order specified, must be equal to their sum. You can find solutions for a / b / c, or b / c / a, or c / a / b, any combination of them and the solution will be correct according to the problem description. Besides, what's does it even has to do with it concluding with confidence: "The fundamental issue is that division tends to ma…

> There is no ambiguity

Yet you give three different interpretations:

> You can find solutions for a / b / c, or b / c / a, or c / a / b

This is a clear case of ambiguity.

Even the classic question is ambiguous: "Which 3 numbers give the same result when added or multiplied together?"

Lets say the three numbers are x, y and z and the result is r. A valid interpretation would be to multiply/add every pair of numbers:

    x * y = r
    y * z = r
    x * z = r
    x + y = r
    y + z = r
    x + z = r
However, I do not think that this ambiguity is the reason why OpenAI o1 fails here. It simply started with an untractable approach to solve this problem (plugging in random numbers) and did not attempt a more promising approach because it was not trained to do so.

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

#287

Post by benterix »

I remember Murati's interview where she said about this PhD level reasoning and so on, so I was excited to see what they come up with - and it looks like they just used a bunch of models (like 4o's) and linked them in a chain of thought - which is exactly what we have been doing ourselves for a long time to get better results. So you have the usual disadvantages (time and money) and lose the only advantage you had when doing it yourself, i.e. inspecting the immediate steps to understand the moment where it goes wrong so that you can correct it in the right place.

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

#288

Post by throwaway40602 »

oersted wrote: post #265
freediver wrote: post #141

Not seeing major advance in quality with o1, but seeing major negative impact on cost and latency. Kagi LLM benchmarking project: https://help.kagi.com/kagi/ai/llm-benchmark.html

Kagi is most likely evaluating it mainly on deriving an answer for the user from search result snippets. Indeed, GPT-4o is plenty good at this already, and o1 would only perform better on particular types of hard requests, while being so much slower. If you look at Appendix A in the o1 post [1], this becomes quite clear. There's a huge jump in performance in "puzzle" tasks like competitive maths or programming. But t…

hey buddy, you're talking to owner of kagi, and the kagi benchmark is a traditional one

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

#289

Post by icrbow »

gmerc wrote: post #258
MPSimmons wrote: post #106

Earlier quoted context omitted.

The failure is in how you're using it. I don't mean this as a personal attack, but more to shed light on what's happening. A lot of people use LLMs as a search engine. It makes sense - it's basically a lossy compressed database of everything its ever read, and it generates output that is statistically likely - varying degrees of likeliness depending on the temperature, as well as how many times the particular weights…

Interns are cheaper than o1-preview

Not for long.

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

#290

Post by benterix »

bambax wrote: post #281

Near the end, the quote from OpenAI researcher Jason Wei seems damning to me: > Results on AIME and GPQA are really strong, but that doesn’t necessarily translate to something that a user can feel. Even as someone working in science, it’s not easy to find the slice of prompts where GPT-4o fails, o1 does well, and I can grade the answer. But when you do find such prompts, o1 feels totally magical. We all need to find…

> But why? Why would we do that?

Because OpenAI needs a steady influx of money, big money. In order to do so, they have to convince the people who are giving them money that they are the best. An objective way to achieve this is by benchmarking. But once you enter this game, you start optimizing for benchmarks.

At the same time, in the real world, Anthropic is following them in huge leaps and for many users Claude 3.5 is already the default tool for daily work.

Post reply on HN