[Exclusive] OpenAI pioneers science AI

OpenAI Chief Executive Officer Sam Altman. [Source: OpenAI]

OpenAI, the creator behind ChatGPT, an artificial intelligence (AI) chatbot, is expected to create an AI agent specialized in science.

According to multiple industry sources on February 22, OpenAI has been employing physics professors as data labelers to build a scientific AI model.

Data labelers collect and process training data to enhance model performance. AI agents typically perform better when fed with high-quality clean data, free of errors, which OpenAI has reportedly entrusted to physics professors.

“If AI consistently solves physics questions against top scholars, it will eventually enter a stage of solving difficult problems on its own,” said an industry insider.

Previously, DeepMind Technologies‘ AlphaGo used a similar approach.

Before its launch in 2015, DeepMind trained the Go AI through matches against Go champions, eventually defeating top Go player Lee Se-dol in the following year.

It showcased AI’s capabilities to independently master game rules and strategies, surpassing human capabilities.

“Physics researchers can provide AI with principles to understand the world,” said another industry insider. “OpenAI’s engagement of physics professors as data labelers is highly expected to significantly enhance AI quality and efficiency.”

OpenAI has not publicly commented on its work in scientific AI development. However, OpenAI Chief Executive Officer Sam Altman has commented at an industry event that he finds AI applications in scientific discovery most intriguing, expressing his belief that AI will advance science.

Some of OpenAI‘s alleged efforts in physics are reflected in Sora, the company’s generative AI model that produces videos from prompts in texts.

[Source: OpenAI]

Sora has impressed viewers with vivid videos up to one minute long, about four times longer than videos generated by competitors’ models, which typically run for 15 seconds.

Particularly, the videos generated by Sora resemble real videos without common distortions found in AI-generated videos.

“We are teaching AI how to understand the physical world,” OpenAI said.

Typically, large-scale language models generate sentences based on patterns but do not comprehend physical rules.

In contrast, humans naturally understand, based on day-to-day observations, that a ball thrown up into the sky falls back down onto the ground.

When scientific AI reaches a certain level, it is expected to solve physical puzzles on its own, such as in the areas of quantum gravity theories.

As OpenAI ventures into the scientific domain, multi-modal AI that incorporates various AI functions is also expected to accelerate.

OpenAI has previously created Gym, a toolkit for developing and comparing reinforcement learning (RL) algorithms, and Debate Game, an AI that can engage in political debates.