Jump to content

Talk:OpenAI o1

Page contents not supported in other languages.
From Wikipedia, the free encyclopedia

Requested move 14 September 2024

[edit]
The following is a closed discussion of a requested move. Please do not modify it. Subsequent comments should be made in a new section on the talk page. Editors desiring to contest the closing decision should consider a move review after discussing it on the closer's talk page. No further edits should be made to this discussion.

The result of the move request was: Page moved (everyone support). (non-admin closure) 1250metersdeep (talk) 17:20, 20 September 2024 (UTC)[reply]

O1 (generative pre-trained transformer) → OpenAI o1 – The official name, as announced by OpenAI on their website, is "OpenAI o1", and it's also shorter than the current title Alenoach (talk) 23:19, 14 September 2024 (UTC)[reply]

It's like they actively tried to come up with the worst name possible.
People will get it confused with OpenAI the company so I think OpenAI o1 (GPT) maybe? Or Orion-1? The Mining Pickaxe (talk) 16:02, 15 September 2024 (UTC)[reply]
The name "Orion" isn't widely used, so I would avoid that. The title "OpenAI o1 (GPT)" is a bit more explicit, but it doesn't look very clean. By the way, I guess we can continue to use "o1" instead of "OpenAI o1" in the rest of the article, newspapers seem to do that also. Alenoach (talk) 23:17, 15 September 2024 (UTC)[reply]
You right. As much as it pisses me off to say this, OpenAI o1 is really the only reasonable name for this article. :argh: The Mining Pickaxe (talk) 01:03, 16 September 2024 (UTC)[reply]
Support per above. PuppyMonkey (talk) 18:44, 16 September 2024 (UTC)[reply]
Support, this is the name everyone calls it, and it would be the best name for this article. 1250metersdeep (talk) 01:17, 18 September 2024 (UTC)[reply]
The discussion above is closed. Please do not modify it. Subsequent comments should be made on the appropriate discussion page. No further edits should be made to this discussion.

Isn't Strawberry the name of the training algorithm?

[edit]

Thought Strawberry is the name of the new training algorithm that is precisely not normal reinforcement learning, because RL tends to make the model answers converge on certain path, while Strawberry generates a model with a property they call diverse in the sense, that it generates a hole spectrum of path instead of converging onto one only. This spectrum, multi-angle thinking then makes models trained with Strawberry much better suitable for Chain of Thought / Autonomous agent kinds of prompting techniques.

Of course all of this is just rumors, bits&pieces because we are at a point where market dominance and monopoly desires lead to a situation where the research itself is not published to benefit humanity, like in basically all science before. A tragedy 2A00:20:6045:EA22:1068:4F19:6E53:4C33 (talk) 21:02, 26 September 2024 (UTC)[reply]

Hi, I think that strawberry is the name of the model o1 itself rather than the training algorithm, as said for example here.[1] OpenAI gives limited details about how it works. We know that reinforcement learning is being used (possibly among other techniques), OpenAI said it itself.[2]
If I understand correctly: It was notably trained on many problems that have a known correct answer. It generates long chain of thoughts with a high "temperature" (i.e. creativity), and the chain-of-thoughts that get to the good result (and that are estimated by another model to follow good reasoning steps) are rewarded. But if no reliable source precisely explains how it works, I would avoid speculating inside the article. Alenoach (talk) 22:20, 26 September 2024 (UTC)[reply]
Retrieved from "https://en.wikipedia.org/w/index.php?title=Talk:OpenAI_o1&oldid=1247955667"