The full-blooded version of O1 is leaked! The reasoning of the math Olympiad pictures is easy to grasp, and Ultraman is online to reveal the spoilers of O2
Yuyang from Aofei Temple
Quantum Bit | Public Account QbitAI
The full-blood version of o1 has been leaked!
Unexpectedly, OpenAI still said “definitely this year”, and many netizens have already taken the lead in experiencing it (doge).
Here’s what’s going on: Some netizens discovered that OpenAI has deployed the complete version of o1 on the ChatGPT official website, and everyone can use it directly without getting stuck in the bug!
The netizen who first discovered this confirmed through testing that the leaked model has the image reasoning ability mentioned by Ultraman .
△ Image source: ????@legit_rumors
The log also shows that the model is the o1 original model.
So, before OpenAI could fix it, netizens took the opportunity to have a great time playing with it.
How powerful is full-blooded o1?
The most obvious difference between the leaked o1 and the already released o1-preview and o1-mini is that the leaked o1 supports uploading attachments.
This means that we can finally test the multimodal capabilities of o1.
The netizens didn't waste any time and went straight to the point of intensity. They posted a screenshot of the proof question from the Putnam Mathematical Competition and threw it to the full-blooded O1.
For positive integer n, fn(x)=cos(x)cos(2x)cos(3x)…cos(nx). Find the smallest n such that |fn '' (0)|＞2023.
After thinking for 18 seconds, o1 gave the answer: 18.
Correct answer:
Here is a more complicated one. Although I encountered a small problem in image recognition, I still got it right.
With such an amazing performance, it seems that correctly identifying the first black hole photo in human history is nothing (doge).
However, there are also cases where the model fails. For example, ask the large model to count the number of triangles in the picture.
A netizen named anagh made a screen comparison between the leaked model and GPT-4o: