OpenAI launches o1, an artificial intelligence model capable of “reasoning”

OpenAI ensures that this new AI can solve more complex problems than previous models.

OpenAI today announced a new family of artificial intelligencebaptized o1capable of “human-like reasoning” and tackling complex problems in mathematics, programming and science. The announcement marks another step in OpenAI’s goal of creating general artificial intelligence.

“We are at the beginning of a new paradigm: AI capable of complex general reasoning,” OpenAI CEO Sam Altman wrote on the social network X.

The model o1 by OpenAI learned to think criticallyto consider different approaches and to recognize one's own mistakes in order to learn. This deep reasoning ability allows o1 to respond complex problems more efficient than previous models, according to the company. Because AI spends more time thinking and generating a more thoughtful response, rather than responding immediately like GPT models, it will be possible improve the quality of responses.

“o1 think for a few secondsBut our goal is for future versions to think for hours, days, even weeks. The cost of the answer will be higher, but how much would you pay for a new cancer drug? For innovative batteries? For a test of the Riemann hypothesis? “AI can be more than chatbots,” OpenAI researcher Noam Brown said in a post on X.

The company that created ChatGPT assures in a press release that this new model has shown in the tests carried out a performance comparable to that of doctoral students in fields such as physics, chemistry and biology. Likewise, he points out that they have achieved good results in mathematics and programming. According to OpenAI, in the International Mathematical Olympiad exam, o1 achieved a score of 83%, which is a significant improvement over its predecessors.

The AI ​​lab says the launch of o1, known during its development under the code name of Strawberriesdoes not mean that the company is abandoning the development of new versions of its GPT models.

Boundaries

In fact, OpenAI points out that These new models still have limitations. As it is a nascent model, it does not yet have many functions that make ChatGPT useful, such as searching for information on the Internet and downloading files and images. In addition, the response may take longer. In this sense, the to start up He notes that in many use cases, his GPT-4o model “will be more effective in the near term.”

The model, in English, can be testing in a preliminary version via ChatGPTby selecting it in the template selector for paying users of its Plus and Team versions. Business and Education plan subscribers will be able to access it next week.

The company also launched a smaller, faster model, called o1 minispecially designed for program code. In both cases, there is a weekly usage limitationmore precisely 50 questions for o1 and 30 for the mini version.

The company led by Sam Altman assures that its goal in the future is to integrate it into ChatGPT, so that the chatbot can use the most appropriate model (o1 or GPT) depending on the request made by the user. Likewise, he emphasizes that In future versions, m o1 will be able to search for information onlineas well as understanding files and images.

Training

OpenAI’s GPT models were trained to reproduce text patterns. However, the o1 model’s approach is different: instead of just imitating patterns, it learns to solve problems on its own using reinforcement learning, a technique that teaches the system through rewards and penalties.

Furthermore, o1 uses a human-like reasoning processin which you analyze problems step by step to arrive at a solution, rather than responding instantly. In this way, AI is supposed to have fewer “hallucinations,” a term used when chatbots invent information.

The announcement represents a new step in the way AI models process informationsince they are no longer limited to prior training, like GPT. Now, performance can be improved by providing more computing resources during the process in which the model “thinks” to generate the answer, so that it can perform more complex or deeper reasoning and, therefore, improve the quality of its answers.

These types of models aren’t always better, according to Brown. “Many tasks don’t require reasoning, and sometimes it’s not worth waiting for an o1 answer instead of a fast GPT-4o answer. One of the motivations for releasing a preview is to see which use cases are most popular, so improve the models,” he explains.

OpenAI seeks to create a general artificial intelligence, that is, one that seeks to replicate human cognitive ability. This intelligence, if achieved, could be capable of understanding, learning, and applying knowledge to any intellectual task.

This technological career requires considerable financial resources. The company negotiates a funding round of up to $6.5 billion for a valuation of $150 billion.



Source link

Leave a Comment