TLDR
- OpenAI launches new ’01’ model family, outperforming ChatGPT-4o
- New models use “chain-of-thought” reasoning for complex tasks
- Even smallest 01 model surpasses GPT-4o in several key areas
- Models emphasize “deliberative reasoning” for more thoughtful answers
- New model will be available on ChatGPT Plus, with limited free access
OpenAI, a leading artificial intelligence company, has introduced a new family of language models called ’01’. Released on Thursday, September 13, 2024, these models are now available to ChatGPT Plus subscribers.
The company claims that the new series offers significant improvements in performance and reasoning capabilities compared to its previous models.
The ’01’ series marks a departure from OpenAI’s usual naming convention. Instead of following the ChatGPT-3, 3.5, and 4o series, the company has decided to “reset the counter” due to what they describe as a major leap forward in AI capability.
This change in naming reflects the company’s belief that the new models represent a significant advancement in the field.
A key feature of the ’01’ models is their use of “chain-of-thought” reasoning. This approach allows the AI to take more time to process information before providing a response.
OpenAI says this results in more thoughtful and coherent answers, especially for complex tasks that require in-depth reasoning.
According to OpenAI’s internal testing, even the smallest model in the ’01’ lineup outperforms the top-tier GPT-4o in several key areas.
We're releasing a preview of OpenAI o1—a new series of AI models designed to spend more time thinking before they respond.
These models can reason through complex tasks and solve harder problems than previous models in science, coding, and math. https://t.co/peKzzKX1bu
— OpenAI (@OpenAI) September 12, 2024
The company reports particularly impressive results in challenges considered to have PhD-level complexity. However, it’s worth noting that the improvements were less dramatic in creative tasks such as creative writing.
The new models emphasize what OpenAI calls “deliberative reasoning.” This process involves the AI system taking additional time to work through its responses internally.
The goal is to produce more carefully considered answers, especially for tasks that require extensive reasoning.
OpenAI has published internal testing results showing improvements over GPT-4o in areas such as coding, calculus, and data analysis. The company also states that human evaluators rated the overall performance of the ’01’ models favorably.
The implementation of chain-of-thought processing during inference is a notable aspect of the new model’s capabilities. This means the model uses a step-by-step approach to reason through a problem before providing a final result to the user.
OpenAI believes this method has the potential to unlock substantial benefits, while also acknowledging that it may increase potential risks associated with heightened intelligence.
OpenAI claims that embedding more guidelines into the chain-of-thought process not only improves accuracy but also makes the model less susceptible to jailbreaking techniques. This is because the model has more time and steps to identify potentially harmful outputs.
Despite these measures, the jailbreaking community reportedly found ways to bypass the AI safety controls within minutes of the model’s release.
The company plans to expand the models’ capabilities in the future, including adding web search functionality and improving multimodal interactions. They also intend to refine the model over time to meet OpenAI’s minimum standards for safety, jailbreak prevention, and autonomy.
