September 14, 2024, 9:53 PM
September 14, 2024, 9:53 PM
ChatGPT creator OpenAI has launched o1, a new line of artificial intelligence (AI) models generative capable of reasoning and respondingfor complex questions, for example in mathematics.
Unlike their predecessors, these new models have been designed to refine their thought processestry different methods and acknowledge errors before providing a final answer.
OpenAI CEO Sam Altman praised the models as “a new paradigm: an AI that can perform complex general-purpose reasoning.”
However, he warned that the technology “is still flawed, limited and seems more impressive the first time you use it then to spend more time with her.”
Microsoft-backed OpenAI said that in trials, the models performed comparably to students of theDoctorate in difficult physics tasks, chemistry and biology.
They also excelled in math and coding, achieving an 83% success rate in an exam qualification for the International Mathematical Olympiad, compared to 13% for GPT-40, its most advanced general-purpose model.
In a mathematics competition for American high school students, o1 was “among the top 500,” he added..
“Just like a human being who can think for a long time before answering a difficult question, o1 uses a chain of thinking (…) Learn to recognize and correct its mistakes. Learn to break down the trickier stages into simpler ones. Learn to try a different approach when the current one doesn’t work,” OpenAI explained.
The company said the enhanced reasoning capabilities could be used for healthcare researchers to annotate Cell sequencing data, which physicists can generate complex formulas and computer developers are able to build and execute multi-step designs.
He also indicated that the new models are able to better resist attempts to bypass security mechanisms.
OpenAI highlighted that its enhanced security measures included recent agreements with the AI Safety Institutes of The United States and the United Kingdom, which are granted early access to the models for evaluation.
But OpenAI is just one of the key players in the technology industry that has accelerated the development of this type of AI models. Google, Anthropic, Meta, among others, are also working on similar projects, with the aim of creating tools that act as ‘copilots’ in various tasks, increasing their usefulness and effectiveness in professional and personal applications.
Melanie Mitchell, professor at the Santa Fe Institute (New Mexico), author of the book Artificial Intelligence: A Guide for Thinking Humans, suggests that these models do not actually do deep reasoning, but rather they extract reasoning patterns from similar problems in the data with which they were trained, that is, they simulate intelligence.
