OpenAI The rumored Strawberry.AI ModelsNow available, officially known as "o1".It is the company's first model with "reasoning" capabilities.
o1 and o1-mini
OpenAI says special training of the model can answer more complex questions faster than humans. It was released alongside o1-mini, a smaller, lower-cost version.
OpenAI says the release of the o1 model is a key step toward its human-like AI ambitions.
The o1 model is currently in the "preview" phase, and officials emphasize that it is still in the early stages of development, is more expensive and slower to use than the GPT-4o model, but is better at writing code and solving multi-step problems.
price
OpenAI says that starting today, ChatGPT Plus and Team users will have access to o1-preview and o1-mini, while Enterprise and Edu users will get access early next week.
OpenAI says it plans to extend access to o1-mini to all ChatGPT free users, but has not yet set a release date.
The cost for developers to gain access to o1 is quite high: if the API is called, the input tokens for o1-preview cost per million. 15 USD (currently about Rs. 107), and the output word meta cost is per million 60 USD (currently about 427 RMB).
In comparison, GPT-4o's input word element cost per million 5 USD (currently about 35.6 RMB), and the output word meta cost is per million 15 USD (currently about 107 RMB).Thus the o1 model increases the input lexical element cost by a factor of two and the output lexical element cost by a factor of three.
Training methods
Jerry Tworek, OpenAI's head of research, said that o1 uses a different training method than current models, without giving details.
He mentioned that o1"It was trained using a new optimization algorithm and a training dataset specifically tailored for it.”
OpenAI trained the o1 model to solve problems on its own using a technique called reinforcement learning, which teaches the system through rewards and penalties, according to the report. Subsequently, o1 uses "thought chains" to process queries, similar to the way humans solve problems through step-by-step reasoning.
As a result of this new training method, OpenAI says the model should be more accurate.
We can't say we've solved the hallucination problem, but at least from the observations," Tworek said.This model produces significantly fewer hallucinations”.
o1 Model strengths
The main difference between the o1 model and GPT-4o is that it is able to handle complex programming and mathematical problems better than its predecessor and can explain its reasoning process, as emphasized by OpenAI.
Bob McGrew, OpenAI's chief research officer, said:
- This model definitely outperformed me in answering AP math exam questions, and I minored in math in college.
He mentioned that OpenAI also had o1 take the qualifying exam for the International Mathematical Olympiad, and that GPT-4o correctly solved only 13%.o1 model is able to address 83%.
The new model reached 89th place among entrants in an online programming competition called the Codeforces contest, and OpenAI claims that the model's next update will perform "on par with PhD students" on challenging benchmark tasks in physics, chemistry, and biology.
At the same time, the o1 is less capable than the GPT-4o in many areas. it does not perform as well as the latter in terms of world factual knowledge. In addition, it does not have the ability to browse the Web or process documents and images. Nevertheless, the company believes it represents a whole new class of capabilities. The name o1 is meant to indicate "resetting the counter back to 1".