OpenAI o1 : When artificial intelligence learns to think 🧠

Cliquez ici pour lire en français

OpenAI, the company at the forefront of artificial intelligence, has just unveiled a new series of AI models called OpenAI o1. This announcement marks a significant milestone in the field of generative AI, promising reasoning capabilities far superior to those of its predecessors.

OpenAI o1-preview: an AI that thinks before responding 🧠

Unlike previous models, OpenAI o1 is designed to « think » before responding. Models like GPT-4, for example, are primarily based on statistical predictions. This innovative approach allows the model to refine its thinking process, try different strategies, and even recognize its own errors.

To illustrate this advancement, OpenAI tested its models with the International Mathematics Olympiad. While GPT-4 solved only 13% of the problems, o1 achieved an impressive score of 83%. In coding, it ranked in the 89th percentile in Codeforces competitions, far surpassing the performance of previous models.

Applications in research and development 🔬

OpenAI o1 is not only excelling in mathematics and programming. OpenAI has showcased several real-world use cases where this AI can transform how researchers and developers work. For example, healthcare researchers can use o1 to annotate cell sequencing data, while physicists can leverage it to generate complex mathematical formulas necessary for fields like quantum optics.

Moreover, OpenAI o1-mini, a lightweight version of this model, is specially optimized for code generation and debugging. Developers thus benefit from an efficient and economical solution. This mini model is 80% less costly to use, making it accessible to a wider range of users without sacrificing reasoning capabilities.

Source : OpenAI

A model still in development 🛠️

It’s important to note that OpenAI o1 is still in its early stages and doesn’t yet have some key features that make GPT-4 so popular, such as web browsing and handling user-shared files. However, OpenAI promises continuous improvements, and for complex reasoning tasks, o1 already represents a considerable advancement over previous models.

Safety and ethics at the core of development 🛡️

OpenAI emphasizes having focused on safety and ethics during the development of these new models. Rigorous tests have been conducted to ensure that o1 respects safety guidelines, even in the face of attempts to circumvent them. In one of their most difficult security tests, o1-preview scored 84. For comparison, GPT-4 had only scored 22, demonstrating a clear improvement in terms of security.

The company has also strengthened its collaboration with the U.S. and U.K. governments, notably through partnerships with their AI safety institutes. These agreements allow these institutes early access to a research version of the new models to evaluate their safety before they are made publicly available.

How to Try OpenAI o1 ? 💻

ChatGPT Plus and Team users can now access OpenAI o1-preview and o1-mini via ChatGPT. However, due to its early launch phase, access is currently limited to 30 messages per week for o1-preview and 50 for o1-mini. These restrictions are expected to be gradually lifted with future updates.

With the launch of OpenAI o1, the company is taking a new step in the development of generative AI. This advancement promises revolutionary applications across many fields while raising new questions about the impact of these technologies on our society.

Do you use or have you ever used ChatGPT? What do you think about generative AI in general? Come discuss it in the comments.