A new OpenAI model is being developed by another AI system—superintelligence is already almost here.

How an AI model “writes” the next one: a possible step toward superintelligence

In a recent CNBC interview, SoftBank CEO Masayoshi Son said that OpenAI is developing its future model using an already existing AI system. According to him, this “recursive” method will be applied to all large models: each new system will create the next one, resulting in exponential gains in intelligence.

> *“As soon as that happens, an AI model will generate the next AI model. It will be exponentially smarter than all of us. That’s superintelligence,”* Son said, adding that engineers’ capabilities will no longer keep up with the pace of development.

OpenAI has not commented on the details of its research, but in February announced GPT‑5.3‑Codex as “the first model that played an important role in its own creation.” The Codex team used early versions for debugging training, deployment management, and result evaluation.

How Masayoshi Son’s forecasts have changed
Two years ago, Son spoke of the emergence of artificial superintelligence (ASI) in 10 years and being 10,000 times smarter than a human. This week he revised his estimate: now, according to him, ASI could arise in just two years.

> *“In my view it should have happened in four years, not ten. Now I say it will happen within the next two years,”* he noted.

Son also mentioned that he uses ChatGPT from OpenAI 2–3 hours a day, considering it “smarter” in most areas. According to SoftBank’s forecast, in the next couple of years AI will outperform humans in roughly 70–80 % of domains, and where it surpasses human intelligence it could be up to ten times smarter.

Why this matters
SoftBank has long invested in AI: from owning Arm (a chip manufacturer) and a stake in OpenAI to investments in robotics and autonomous driving. He believes the AI revolution will be 50 times larger than the dot‑com boom of the early 2000s.

However, the risks of “recursive self‑improvement” (RSI)—when a system autonomously designs its successor—have drawn attention. Anthropic published a warning about potential loss of control over such systems but also noted possible benefits from RSI.

OpenAI’s recent work mentioned “early signs” of RSI in modern models and predicts increased competitive pressure among developers and countries, as well as managerial challenges that existing institutions are not ready to handle.