OpenAI has released latest advanced reasoning models called the “o1” series.
Currently available in two versions – o1-preview and o1-mini – o1 is designed to perform complex reasoning tasks, marking what OpenAI calls “a brand new paradigm” in AI development.
“This is what we consider to be the brand new paradigm of those models,” explained Mira Murati, Chief Technology Officer of OpenAI, in a Statement on Wired“It is a lot better at handling very complex pondering tasks.”
Unlike previous versions, which were characterised primarily by their scalability, for instance by solving an issue using computing power, O1 goals to copy the human thought technique of “pondering through” problems.
Instead of generating a single answer, the model works step-by-step, considering multiple approaches and revising itself when essential. This method is known as “thought chain” stimulation.
This enables it to resolve complex problems in mathematics, coding, and other areas with a level of precision that existing models, including GPT-4o, struggle to attain.
We publish a preview of OpenAI o1 – a brand new series of AI models designed to spend more time pondering before reacting.
These models can handle complex tasks and solve tougher problems than previous models in science, coding and arithmetic. https://t.co/peKzzKX1bu
— OpenAI (@OpenAI) 12 September 2024
Mark Chen, OpenAI's vice chairman of research, explained how o1 excels at improving the educational process. “The model sharpens its pondering and refines the strategies it uses to reach at the reply,” Chen said.
He demonstrated the model using several mathematical puzzles and advanced chemistry questions that had previously puzzled GPT-4o.
A riddle that puzzled previous models was: “A princess is as old because the prince will likely be if the princess is twice as old because the prince was when the princess's age was half the sum of their present ages. How old are the prince and the princess?”
Model o1 determined the right answer: The prince is 30 and the princess is 40.
How to access o1
ChatGPT Plus users can already access o1 from ChatGPT.
This is a surprise, as GPT-4o's voice feature is being introduced months after the demo version. I don't think many individuals would have thought that o1 would take off so quickly.
o1 appears to be connected to OpenAI's Strawberry project. Now, strangely enough, it's necessary to notice that almost all AI models don't know the way many Rs are in Strawberry. This puts their pondering skills to the test.
I tested this in o1 and it worked rather well. Obviously, o1's approach helps to resolve such questions efficiently.
Sam Altman's recent spate of strawberry-related social media conversations is likely to be related to this famous strawberry AI problem and o1's codename “Project Strawberry.” If not, that's a wierd coincidence.
A fundamental change in problem solving
The O1 model's ability to resolve problems “logically” represents an advance in artificial intelligence – something that might prove quite groundbreaking if its performance is proven in practice.
The latest models have already shown strong performance in tests akin to the American Invitational Mathematics Examination (AIME).
According to OpenAI, the brand new model solved 83% of the issues encountered in AIME, in comparison with only 12% for GPT-4o.
The strengths of o1 are obvious, but there are also compromises.
Due to the more sophisticated methodology, the model takes longer to generate responses. We don't yet know the way pronounced this will likely be and what impact it should have on the user experience.
The strange origins of O1
o1 comes following discussions about an OpenAI project codenamed “Strawberry” that was created in late 2023.
Rumor had it that it was an AI model able to autonomous web exploration, designed to conduct “in-depth research” reasonably than simply retrieving information.
The rumors about Strawberry recently gained momentum when The Information leaked some details about OpenAI's internal projects. Apparently, OpenAI is developing two versions of Strawberry.
- One of them is a smaller, simplified version intended for integration with ChatGPT. It goals to enhance reasoning skills in scenarios where users need more thoughtful, detailed answers reasonably than quick reactions. That feels like o1.
- Another is a bigger, more powerful version that will likely be used to generate high-quality “synthetic” training data for OpenAI’s next flagship language model, codenamed “Orion.” This may or will not be linked to o1.
OpenAI has not provided a direct explanation of what Strawberry really is.
A complement, not a alternative
Murati emphasized that o1 shouldn’t be intended as a alternative for GPT-4o, but as a complement.
“There are two paradigms,” she said. “The scaling paradigm and this latest paradigm. We expect to merge them.”
As OpenAI continues to develop GPT-5, which can likely be even larger and more powerful than GPT-4o, future models may incorporate the reasoning capabilities of o1.
This fusion could address the persistent limitations of huge language models (LLMs), akin to their difficulty in solving seemingly easy problems that require logical reasoning.
Anthropic and Google are said to be racing to integrate similar features into their models. Google's AlphaProof project, for instance, also combines language models with reinforcement learning to resolve difficult mathematical problems.
However, Chen believes OpenAI has the sting. “I believe we've made some breakthroughs here,” he said. “I believe that's a part of our edge. It's actually pretty good at drawing conclusions across the board.”
Yoshua Bengio, a number one AI researcher and winner of the distinguished Turing Award, praised the progress but urged caution.
“If AI systems were to show real pondering, this is able to enable the consistency of the facts, arguments and conclusions drawn by the AI,” he told the FT.
Safety and ethical considerations
As a part of its commitment to responsible AI, OpenAI has enhanced the safety features of the O1 series, including content security tools enabled by default.
These tools help prevent the model from producing harmful or unsafe outputs.
“We are pleased to announce that Prompt Shields and Protected Materials for Text are actually generally available within the Azure OpenAI Service,” OpenAI said in a Microsoft blog post.
The o1 series is obtainable for early access in Microsoft Azure AI Studio and GitHub Models, with a broader release planned soon.
OpenAI hopes that o1 will enable developers and corporations to innovate more cost-effectively, in keeping with its broader mission to make AI more accessible to enterprise users.
“We imagine this can allow us to deliver information more cheaply,” Chen concluded. “And I believe that's really the core mission of our company.”
All in all, an exciting publication. It will likely be very interesting to see what questions, problems and tasks o1 is coping with.
