OpenAI Launches New AI Models That Can 'Reason'

According to OpenAI, these models are capable of solving more difficult problems in science, programming, and mathematics than previous models.

OpenAI logo. Photo: REUTERS/Dado Ruvic

The new models, called o1 and o1-mini, were officially available on ChatGPT on Thursday. Noam Brown, a researcher at OpenAI, confirmed that the models are the Strawberry project mentioned earlier.

According to OpenAI, the o1 model scored 83% in the International Mathematical Olympiad qualifying round, far surpassing the 13% of its predecessor, GPT-4o. The model also improved its ability to answer competitive programming questions and achieved PhD-level accuracy on scientific problem-solving scales.

Brown said the model achieved this feat using a technique called “chain-of-thought,” which involves breaking down a complex problem into simpler logical steps. Instead of requiring human input, OpenAI automated this ability so the model could solve the problem on its own.

“We train these models to think carefully before giving an answer, just like a human would,” OpenAI said. “Through training, they learn to refine their thought processes, try different strategies, and recognize errors.”

Cao Phong (according to Reuters)

Source: https://www.congluan.vn/mo-hinh-ai-moi-cua-openai-co-kha-nang-ly-luan-lam-duoc-ca-de-toan-olympic-post312158.html