On September 12, OpenAI released a new generation of O1 series models, claiming that it can reason about complex tasks and solve more difficult problems than previous scientific, coding, and mathematical models. These advances are also considered key breakthroughs towards artificial general intelligence (AGI).

Open a new chapter in complex reasoning

OpenAI said that the new model demonstrated new heights of artificial intelligence in handling complex inference tasks, so it was decided to name it a completely new identifier instead of continuing the naming sequence of "GPT-4". This marks a major step forward in the era of AI – ushering in large models capable of general, complex inference.

It should be noted that the chat function of the new model O1 is relatively basic. In contrast to GPT-4 before it, O1 is unable to browse the web or analyze files. Although it has the ability to analyze images, this feature is temporarily turned off pending further testing. In addition, O1 has a limit on the number of messages - 30 messages per week for the O1-Preview version and 50 messages per week for the O1-Mini version.

From now on, the o1-preview and o1-mini versions will already be available in ChatGPT Plus/Team and the API interface, and enterprise and educational institution users will start getting priority access after next Monday.

Sam, CEO of OpenAI "This is our most robust and consistent model series to date, o1, and our best inference model to date," Altman said. While O1 still has some shortcomings and limitations, it's still impressive in practice. ”

Solve complex problems

OpenAI's latest model, O1, is capable of solving more complex scientific, coding, and mathematical puzzles than previous GPTs.

According to Jerry Tworek, head of research at OpenAI, O1's training method is fundamentally different from its predecessor. In the past, GPT models mainly mimicked patterns in training data, while o1 was trained to solve problems independently through reinforcement learning, using reward and punishment mechanisms to teach AI to use "thinking chains" to analyze problems step by step, similar to the way humans think.

This means that ChatGPT now thinks deeply before providing an answer, rather than just giving an immediate response. This improvement has allowed ChatGPT to move from relying solely on intuitive and quick reactions (System 1) to being able to respond thoughtfully (System 2) to more complex problems.

Inference large models are characterized by AI spending more time thinking before providing an answer, rather than simply predicting word order to generate an answer. In some cases, users can see the AI show thought processes like "I'm thinking about whether this is feasible" or "Time is running out, I need to give an answer as soon as possible." However, OpenAI points out that these are not original thought chains, but "summaries of model generation."

The test showed that in the International Mathematical Olympiad Qualifying Tournament, GPT-4o solved only 13% of the problems, while o1 solved 83%. In the Codeforces test of programming ability, o1 reached the 89th percentile, while GPT-4o only reached the 11th percentile.

OpenAI has found that O1's performance continues to improve as more reinforcement learning and more time to think are allowed. In addition, O1 outperformed even human experts in some tests, approaching the level of a science scientist, becoming the first model to achieve this feat in this test.

At the same time, OpenAI also released the o1-mini model, which is faster and cheaper than the o1-preview, and the price is reduced by 80%, which is suitable for those application scenarios that require inference but do not require extensive background knowledge.

OpenAI is in the process of raising a new round of funding

Although the new OpenAI o1 model does not yet have comprehensive problem-solving capabilities, its significantly enhanced inference capabilities make it more valuable for applications in professional fields such as science, programming, and mathematics. In addition, O1 boosts AI The standard of agent technology has greatly enhanced the ability of scientific research and production, but it is not of great significance to the consumer market.

Jim Fan, chief scientist at NVIDIA, said that the new O1 requires more computing power and data support, and can form a data flywheel effect. The correct answer and its thought process can be used as high-quality training data to help continuously improve its inference core, a process similar to AlphaGo's value network that generates more refined data through Monte Carlo Tree Search (MCTS) to improve its capabilities.

OpenAI's o1 series models not only greatly enhance inference capabilities, but also introduce a new Scaling paradigm: unlocking inference time computation (Test) through reinforcement learning (RL). time compute）。

Separately, OpenAI is reportedly in the process of raising a new round of funding at a valuation of about $150 billion and is expected to raise about $6.5 billion from investors including Apple, Nvidia, and Microsoft. At the same time, OpenAI is also negotiating a $5 billion revolving credit facility with banks.