Writing Puzzles with OpenAI o1

[zaharizarev]

Congratulations OpenAI ! L2 reasoners REACHED.

[ashishmaheshwari]

Amazing can't wait to explore all the new work or examples that people around the internet bring up using this new model.

[samhblackmore]

They really should show the detailed reasoning text in every video. It makes it a much more impressive demo seeing the thought process written out rather than just "Thinking..."

[nnova]

Sega Dreamcast was ahead of its time with the "It's Thinking" slogan

[Nb]

This feels pretty underwhelming. The kind of strategy feedback perplexity gives with search looks pretty similar to what's happening here, perhaps just with fewer iterations and flexibility. Also, it feels closer to a "guided prompting loop" than a dramatic improvement in the model. It's not clear to me whether this actually performs better than prompting in a methodical step by step way where you have it slowly respond with each set of reasons. I would expect that the way it's build keeps it more consistent and on track than what you'd likely get with careful prompts but the comparison between a one shot question and an iterative reasoning loop makes it hard to evaluate how much of this is from changes in the model internals vs changes in prompt -> model input augmentations and response evaluation loops.

I guess the proof will have to be in the raw success statistics across each modality when it gets rigorously tested, though if the prompting side is being manipulated before applying it to the part of the model traditionallyunder evaluation the input for each problem seems potentially dissimilar leading to an apples to oranges comparison problem

[benhaase]

This model has to be crazy expensive just based off the time its taking to answer this one question

[Purified-Bananas]

Day 1 - Thinking through words ending in "¡" like "alibi" and "esprit." Also, considering Italian words.
Ultimately realizing that identifying words ending in "i" is trickier than expected.
Day 2 - I'm gathering a list of words ending with "i, " like "ski" and "hovercraft, " to explore their usage in a sentence. Progress is steady but requires further refinement.
Day 3 - The philosophy of the meaning of ending in "i" has proven very, very tricky. Am I? Am I not? Who am I? Will the world end in "i"?
...

[daveinpublic]

This is a good video because it shows some of the more difficult challenges under the hood.

[Ethan-mxcm]

It's Jason! Good job on another model released.

[henrik]

1:34 Does hovercraft end with an “i”???

[ramsey]

Me when AI takes me hostage and offers a puzzle to spare me:

[Kylehudgins]

This is the most amazing thing I’ve ever seen in my life

[SmileyMediaAI]

Well done team 👏 I can't wait to explore

[kyleingram]

This is just a fancy version of SmartGPT from AI_Explained

[PeterQian-qs]

Was it doing prompt engineering on itself or something? All these previews really seems like it's adding"Doing it step by step" to the prompt.

[universemaster]

I'm a paid subscriber and still don't have advanced voice mode!

[jugstries]

Seems impressive at first look, but it is still hallucinating pretty hard and it's easy to see if you pause the video and take the time to read the "cogitations" it makes (like "hovercraft" ending with "i" — as another commenter noted —, evaluating "safari" as a possible rhyme to "safari" etc.)

[Zakc]

Wish there were constraints for a ryhme scheme

[grozkal]

The reasoning steps in the text differ when comparing before and after the cut scenes at 1:54 and 1:58. While the overall result is impressive, but we are certainly cherry picking here.

[azzazin]

esprit doesn't end with i nor does hovercraft