Masayoshi Son: OpenAI's Next-Gen Model Is Being Designed by AI; Superintelligence to Arrive Within Two Years
Miles Bennett
SoftBank CEO Masayoshi Son says OpenAI's next-generation model is already being designed by another AI model, and compresses his superintelligence timeline from ten years to under two — a signal that AI self-iteration is rewriting everyone's schedule.
What does "AI designing AI" actually mean?
Son says OpenAI engineers told him an AI model is designing the next-generation model — engineers alone can no longer do the job.
This means → the bottleneck in AI development is shifting from "humans write the code" to "AI writes the code, humans supervise." The human role moves from designer to auditor.
OpenAI said in February that GPT-5.3-Codex was "the first model to play a critical role in its own creation" — the team used early versions to debug training, manage deployment, and diagnose test results.
Why does the superintelligence timeline keep shrinking?
About two years ago, Son publicly predicted superintelligence would arrive "within 10 years" — but he admits he was "deliberately conservative, because people would be shocked." His real expectation at the time was 4 years.
In this interview he compressed the window further: within two years.
In plain terms = in the span of two years, his own estimate halved. This reflects an acceleration so fast that even the most aggressive bettors keep revising upward.
What does "10 times smarter than humans" look like?
Son says he uses ChatGPT two to three hours a day and believes AI is already smarter than he is "in most subjects."
He expects AI to surpass humans in roughly 70% to 80% of subjects within a few years — and in those subjects, it "could be 10 times smarter than the average person."
This means → he is not describing a narrow edge. He is describing an order-of-magnitude gap — if that holds, demand for human labor across knowledge work faces a fundamental rewrite.
Why is "recursive self-improvement" making the industry itself nervous?
Anthropic recently published a blog post on RSI — recursive self-improvement, where an AI system autonomously designs and develops its own successor. Anthropic said it could yield positive outcomes but warned that "full recursive self-improvement could increase the risk of humans losing control of AI systems."
Anthropic called on AI labs to coordinate a slowdown of this line of development.
An OpenAI research paper from June likewise flagged "early signs" of RSI in current systems and warned it would intensify competition among developers and nations, creating "governance challenges that existing institutions are not yet equipped to handle."
How should ordinary people read this?
Son is SoftBank's CEO and a major OpenAI investor. His statement is part disclosure, part interested-party positioning — he has every incentive to make the market believe superintelligence is imminent.
Yet Anthropic and OpenAI's own research papers are sending a similar signal: AI self-iteration is no longer theory — it is an engineering reality already under way.
In plain terms = "AI designing AI" sounds like science fiction, but both leading companies acknowledge it is happening. The only disagreement is whether to accelerate or hit the brakes.
Content is for reference only, not financial advice.