Notes on OpenAI's new o1 chain-of-thought models (simonwillison.net)

Article summary

OpenAI has released two new models, o1-preview and o1-mini, which are designed to improve reasoning capabilities through a chain-of-thought approach. These models use reinforcement learning to refine their strategies and can handle more complicated prompts. The models introduce new trade-offs in terms of cost and performance, with longer response times and higher token usage. The o1 models are currently only available to tier 5 accounts and have limitations such as no system prompt support and no streaming support.

Main themes

• Chain-of-thought models
• Reasoning capabilities
• Reinforcement learning
• Model trade-offs
• API limitations
• Token usage

What commenters say

• The new models' ability to generate synthetic data could be used to train future models, creating a potential level-up process for AI development.
• Hiding the reasoning tokens from users is seen as a step backwards in terms of transparency and debugability, and may be motivated by anti-competitive behavior.
• Some users do not care about the details of the reasoning process and are only interested in the end result, while others believe that understanding the process is crucial for effective use and improvement.
• The pricing model for the new models, which charges for invisible reasoning tokens, is seen as unfair and potentially exploitable by some, while others argue that it is a necessary trade-off for the improved capabilities.
• The use of chain-of-thought models could potentially revolutionize software development by allowing for rapid generation of high-quality code, but others argue that this will not scale and will require significant human oversight.
• The abstraction of complex processes and the use of AI as a tool for speeding up development is seen as a double-edged sword, offering benefits but also potentially leading to a loss of understanding and control.