Softbank developing autonomous AI agents; an AI model that can predict and capture human cognition

Speaking at a customer event Wednesday in Tokyo, Softbank Chairman and CEO Masayoshi Son said his company is developing “the world’s first” artificial intelligence (AI) agent system that can autonomously perform complex tasks. Human programmers will no longer needed. “The AI agents will think for themselves and improve on their own…so the era of humans doing the programming is coming to an end,”

Softbank estimated it needed to create around 1000 agents per person – a large number because “employees have complex thought processes. The agents will be active 24 hours a day, 365 days a year and will interact with each other.” Son estimates the agents will be at least four times as productive and four times as efficient as humans, and would cost around 40 Japanese yen (US$0.27) per agent per month. At that rate, the billion-agent plan would cost SoftBank $3.2 billion annually.

“For 40 yen per agent per month, the agent will independently memorize, negotiate and conduct learning. So with these actions being taken, it’s incredibly cheap,” Son said. “I’m excited to see how the AI agents will interact with one another and advance given tasks,” Son added that the AI agents, to achieve the goals, will “self-evolve and self-replicate” to execute subtasks.

Unlike generative AI, which needs human commands to carry out tasks, an AI agent performs tasks on its own by designing workflows with data available to it. It is expected to enhance productivity at companies by helping their decision-making and problem-solving.

While the CEO’s intent is clear, details of just how and when SoftBank will build this giant AI workforce are scarce. Son admitted the 1 billion target would be “challenging” and that the company had not yet developed the necessary software to support the huge numbers of agents. He said his team needed to build a toolkit for creating more agents and an operating system to orchestrate and coordinate them. Son, one of the world’s most ardent AI evangelists, is betting the company’s future on the technology.

According to Son, the capabilities of AI agents had already surpassed PhD-holders in advanced fields including physics, mathematics and chemistry. “There are no questions it can’t comprehend. We’re almost at a stage where there are hardly any limitations,” he enthused. Son acknowledged the problem of AI hallucinations, but dismissed it as “a temporary and minor issue.” Son said the development of huge AI data centers, such as the $500 billion Stargate project, would enable exponential growth in computing power and AI capabilities.

Softbank Group Corp. Chairman and CEO Masayoshi Son (L) and OpenAI CEO Sam Altman at an event on July 16, 2025. (Kyodo)

The project comes as SoftBank Group and OpenAI, the developer of chatbot ChatGPT, said in February they had agreed to establish a joint venture to promote AI services for corporations. Wednesday’s event included a cameo appearance from Sam Altman, CEO of SoftBank partner OpenAI, who said he was confident about the future of AI because the scaling law would exist “for a long time” and that cost was continually going down. “I think the first era of AI, the…ChatGPT initial era was about an AI that you could ask anything and it could tell you all these things,” Altman said.

“Now as these (AI) agents roll out, AI can do things for you…You can ask the computer to do something in natural language, a sort of vaguely defined complex task, and it can understand you and execute it for you,” Altman said. “The productivity and potential that it unlocks for the world is quite huge.”

……………………………………………………………………………………………………………………………………………..

According to the NY Times, an international team of scientists believe that A.I. systems can help them understand how the human mind works. They have created a ChatGPT-like system that can play the part of a human in a psychological experiment and behave as if it has a human mind. Details about the system, known as Centaur, were published on Wednesday in the journal Nature. Dr. Marcel Binz, a cognitive scientist at Helmholtz Munich, a German research center, is the author of the new AI study.

The researchers gathered a range of studies to train Meta’s LLaMA AI LLM— some that they had carried out themselves, and others that were conducted by other groups. In one study, human volunteers played a game in which they steered a spaceship in search of treasure. In another, they memorized lists of words. In yet another, they played a pair of slot machines with different payouts and figured out how to win as much money as possible. All told, 160 experiments were chosen for LLaMA to train on, including over 10 million responses from more than 60,000 volunteers.

References:

https://english.kyodonews.net/articles/-/57396#google_vignette

https://www.lightreading.com/ai-machine-learning/softbank-aims-for-1-billion-ai-agents-this-year

https://www.nytimes.com/2025/07/02/science/ai-psychology-mind.html

https://www.nature.com/articles/s41586-025-09215-4