Thoughts, and discussion around gpt o1
"We've developed a new series of AI models designed to spend more time thinking before they respond."
- OpenAI
Thank you @young cloud for opening the thread
#OpenAI o1 🍓
shadow kite
spring zenith
Thoughts, and discussion around gpt o1
> "We've developed a new series of AI mo...
-
Happy birhday!
-
I am thinking about doing bug bounty for o1
young cloud
I'm hard pressed to find specific things people are having trouble getting o1 to do, honestly, it doesn't feel any better/more advanced than gpt4 to me.
Then again, I essentially use a series of questions very similar to strawberry in the process of prompting
wicked acorn
First off, I do appreciate all the hard work done by the amazing and smart people at OpenAI, and I understand that such performance (as shown by by o1 in some areas) is not an easy thing, but..
I don't really appreciate the showbizzy nature of their releases. I mean, all these high production demo videos, and 3 blog posts plus a system card (no technical reports anymore?!), and you can't find any piece of real information in all of these!!
tawny vessel
I still think its just a multi stage concept with several calls. You can get the o1 results with a reply prompt of "Are you sure?" and it should be noted that if you are a foreigner with foreign language that the thoughts are in your language while the query answer is exactly the language that you want, which is weird, cause the normal model is not even really capable to keep the right language often. So what is happening here? It would be easily explainable if there is a prompt just for the thought. That is btw also why they dont offer it to the API, cause that might make it possible to crash the situation of the fragile setup. But thats just a guess.
wicked acorn
Interesting discussion from MLST team https://www.youtube.com/watch?v=nO6sDk6vO0g
tawny vessel
Ok OpenAI is really master of marketing
The complete AI scene was already doing before reasoning prompts with big results. We LITERALLY knew that just adding up "Think about it" is already giving better results, and that asking a model to give its thought it made the model better.
But it seems that all people who talk about the OpenAI new model totally forgot that
And that Lex Friedman approved that youtube video is not like adding value to it
like LITERALLY the opposite 😄
dusty sigil
The complete AI scene was already doing before reasoning prompts with big result...
That's literally the point behind training it on the chain of thought. Which I understand is what they did.
If you weren't working with agents already this is very impresive though. Because you don't need to create the feedback loop, just a single prompt returns a much better answer.
So behind the scenes it should be doing something like checking different chains of thought and returning to you the best result. That makes sense if you see that the APi cost is significantly higher than 4o.
tawny vessel
yeah but the question is really
and that is what we cant see:
is it one model
or not
we dont know
thats what make me fuzzy about it
i heard the SAME rumours back in the day about GPT-4
that people said: GPT-4 is just 7 GPT-3.5 that think about it, and an 8th one that decides what to answer
you know what i mean?
Its like you can TRICK a lot here
and no one will know
dusty sigil
I want to believe it is true, because that could mean the result is now a function of compute time and not a function of more compute resources