AI Companies Search For New Paths Towards Smarter Intelligence, Old Method Experiences Limits

Illustration of AI development that is more similar to how humans think (photo: x @johnidouglas)

Jakarta - Artificial intelligence (AI) companies, including OpenAI, face unexpected challenges in developing more sophisticated big language models. They are now focusing on new training techniques that allow the "thinking" algorithm to be more human-like to increase AI intelligence.

This was disclosed by AI scientists, researchers, and investors who believe that this approach could change the technology race in the AI sector, impacting resource needs, such as energy and computer chips.

Previously, OpenAI succeeded with an approach to increasing data and computing power to improve AI models, as used in ChatGPT. However, Ilya Sutskever, co-founder of OpenAI who is now leading a new AI lab called Safe Super Intelligence (SSI), stated that this approach is starting to show limitations. "The 2010s era is a scalability era, but now we are back to the era of discovery," he said.

The researchers are now facing challenges in the form of a model training process that costs up to tens of millions of dollars, requiring thousands of chips operating simultaneously, and vulnerable to hardware failures. In addition, the problem of running out of easily accessible data and energy limitations further complicates the situation.

As a solution, some researchers developed a technique called "test-time compute," which allows AI models to dedicate more computing power when operated to solve complex problems. OpenAI, for example, has used this technique on their new model, "o1," which allows problem solving in several steps such as human mindset.

This move is also of great concern to investors, including Sequoia and Andreessen Horowitz, who are considering its impact on their investments. This shift could affect Nvidia's demand for an AI chip, which has been a market mainstay.

Nvidia CEO Jensen Huang acknowledged the importance of this technique in increasing demand for chips for AI use while operating, not just for training. "We are now finding a second-scale law in use," Huang said.

Smarter AI development with this new approach can change the industrial landscape of AI and create new challenges and opportunities in the tech market.