OpenAI released o1 for complex tasks

OpenAI released o1 for complex tasks: — The model became smarter because it thinks before it answers. We don’t see this process, only a concise description of the result. It uses Chain-of-Thought and, apparently, some criticism and aggregation of results, multiplied by fine-tuning for all of this. — o1 is great at math, physics, and coding, follows instructions well, but knows less about the world around it. — The model solves complex problems in math and programming at the level of international Olympiad medalists. In physics, it’s at the level of a PhD student (approx. 75-80% correct answers). — The new o1 doesn’t need special prompting. It does everything under the hood. — Available for ChatGPT Plus subscribers with limits: 30 messages/week for preview and 50 for mini. Prices will be high because it’s resource-intensive both per token and in the number of reasoning tokens.

Conclusions:

• Models might split into “managed” and “all-in-one” for different task classes.
• Currently, I see no application for myself besides ChatGPT due to cost.
• Need to return to experiments with agent groups.