The Day AI Creates AI — GPT-5.3 Codex and the Era of "Exponential Squared," and Our Survival Strategy

Hello, this is Hamamoto from TIMEWELL.

I usually introduce tech services here, but today is different. Rather than a new tool, I want to write about a quiet yet enormous tectonic shift that is overturning the foundations of how we work and how organizations exist.

Do you remember February 2020? A strange virus was reportedly spreading in Wuhan. Many people were still casually dining out at restaurants, glancing at the news sideways. Anyone who talked about stockpiling toilet paper back then was considered eccentric. Yet within just three weeks, offices closed and daily life collapsed with an audible crack.

I'm convinced the exact same thing is happening right now in the world of AI.

Matt Shumer, CEO of OthersideAI, published an essay in February 2026 titled "Something Big Is Happening." It was viewed more than 80 million times on X and sent shockwaves not just through the technology industry but into finance, law, and education. What he points to is a "quiet singularity" that occurred on February 5, 2026. GPT-5.3 Codex, released by OpenAI that day, finally began spinning the loop of "recursive self-improvement" — AI improving and accelerating its own evolution.

The Shock of GPT-5.3 Codex — The Day AI Developed "Taste" and Judgment

Until now, AI has been an assistant. A human gives an instruction, AI returns a draft, the human revises it. Repeat.

GPT-5.3 Codex is different. According to Shumer: you say "build this app, the features should be like this, the look roughly like that," then leave your desk. Four hours later you return, and the finished product — built from tens of thousands of lines of code — is sitting there. The AI has launched the app itself, clicked buttons, tested it as a user. If there's a bug, it autonomously rewrites the code, and hands you the result in a state it has judged as "ready to ship."

Honestly, when I first read this, I got chills.

This isn't about AI being able to write code. The essence is that AI has developed the taste and judgment to recognize "what a good product looks like."

OpenAI's official release notes describe GPT-5.3 Codex as "the most capable agentic coding model to date, integrating code generation, reasoning, and general intelligence into one model."

What's even more spine-chilling is testimony from inside OpenAI, reported by Hyperdimensional:

Even early versions of GPT-5.3-Codex showed remarkable capabilities, and our team was able to work with those early versions to improve training and support the rollout of later versions.

In other words, GPT-5.3 Codex debugged its own training process, built performance evaluation tools, and accelerated next-generation model development. It is the first model to have meaningfully participated in the "AI creating AI" loop. The fact that many OpenAI researchers and engineers report "doing fundamentally different work than two months ago" speaks to the seriousness of this shift.

For those unfamiliar with "recursive self-improvement" — in plain terms: AI looks at its own report card, analyzes its weak points, changes its own study methods, and scores even higher on the next test. And each time the cycle runs, the speed of improving the study method itself increases. No human teacher can keep up anymore.

The Physics of Evolution — A Future Accelerating at "Exponential Squared"

Recursive self-improvement doesn't just accelerate evolution. The speed itself accelerates. I call this "exponential squared" — or double exponential.

Let me explain. Even ordinary exponential growth exceeds human intuition — Moore's Law is a perfect example. But what's happening now is a nested acceleration: AI improves itself, the improved AI builds the next AI faster, and that AI builds the next one faster still. An explosive change that exceeds our imagination entirely — exponential raised to another exponent — has begun.

Nabeel S. Qureshi put it memorably: "The fact that recursive self-improvement is actually happening is itself astonishing. This was an SF concept. Now everyone is debating the pace." Science fiction has become reality. And no one knows how far that reality will accelerate.

There's concrete evidence. According to METR, an independent organization that evaluates AI capabilities, the "duration" over which AI can autonomously complete tasks without human assistance is doubling approximately every seven months. And as Zvi Mowshowitz notes, that doubling time itself is shrinking. By late 2025, AI could autonomously complete work that takes an expert five hours. At that pace alone, within a few years comes a world where AI independently executes months-long projects. If it's accelerating, the timeline is shorter still.

Analyst Dean W. Ball presents two scenarios:

Scenario	What It Means
Bear case	New models arrive every one to two months. A step-change similar to how the interval shortened from six to nine months to three to four months following the reasoning paradigm discovery.
Bull case	A "intelligence explosion" — the emergence of "superintelligence" within months to years.

Ball himself thinks the most likely outcome is somewhere in between, but treats both extreme scenarios as "live possibilities." This is the frightening part. Even the most conservative prediction represents a change drastic enough to upend the foundations of our businesses.

The End of White-Collar Work and the Redefinition of "Job"

Who does this wave hit first? Some people think superficially: "programmers will lose their jobs." But that undersells the problem dramatically.

Anthropic CEO Dario Amodei predicts that "50% of entry-level white-collar jobs will be affected within the next one to five years." Law, finance, accounting, consulting, marketing. Looking at a screen, reading information, analyzing it, writing text, making decisions — these "digitally complete" forms of knowledge work are coming within range of AI agent replacement, without needing to wait for advances in physical robotics.

As an aside: I recently handed over an internal business process to AI agents, and work that had previously required three people and half a day was done in 30 minutes. The quality was comparable. That experience was, honestly, a shock.

But this isn't simply a story of existing tasks being automated. The structure of work itself changes. PwC describes this shift as "a return from specialist to generalist." Fragmented and specialized work functions are being integrated into "outcome-focused" roles where one person takes responsibility for broader processes — powered by AI agents.

To take software development as an example: requirements definition, design, implementation, testing, documentation — a workflow that multiple specialists previously divided is now something a single experienced engineer can complete by "directing" a team of AI agents. This engineer is no longer a code writer — they're something we might call an AI orchestrator, maximizing business value.

The Organizational Transformation Required — From "Pyramid" to "Hourglass"

If the way we work changes this fundamentally, the shape of organizations must change too. The traditional "pyramid" structure — a few leaders, many middle managers, and even more entry-level employees — is showing its limits.

PwC proposes two future organizational models:

One is the "diamond." As AI replaces routine work, the entry-level layer shrinks and the middle layer managing AI agents thickens. Operations become more agile, but there's a risk of losing the developmental opportunities that cultivate future leaders.

The other is the "hourglass." Entry-level staff with high AI literacy and leaders handling strategic decisions both form thick layers, while the middle connecting them is streamlined. PwC itself is targeting this form, seeing it as the key to maintaining future competitiveness for knowledge-intensive companies.

Personally, I think the hourglass is more realistic for most Japanese companies. The reason is simple: an organization that cuts off its young talent has no future. Hire large numbers of AI-literate young people and give them real experience early. Boldly streamline middle management and, in its place, establish a new role: the "agent manager."

The "agent manager" concept proposed by Harvard Business Review in February 2026 leads hybrid teams of humans and AI agents, taking responsibility for maximizing performance. What's required is less technical AI knowledge than deep understanding of business processes, systems thinking, and the ability to design workflows combining humans and AI to generate business results.

BCG declares flatly that "AI transformation is workforce transformation" and notes that 70% of AI's value comes from "people." Investing in technology alone is meaningless. Education that enables all employees to use AI in daily work, and the organizational culture to support it, are indispensable.

TIMEWELL's WARP program addresses this challenge head-on. The three-month intensive WARP NEXT systematically covers AI fundamentals, prompt engineering, and workflow automation, bringing participants to the level where they can drive AI projects within their own organizations after completion. WARP BASIC, the annual e-learning program, lets you build AI literacy at your own pace. Developing people who can command AI is no longer "nice to have" — it's a condition for organizational survival.

A Prescription for Adapting to Change

AI is accelerating its own evolution. We have entered an era unlike anything in human history. This is simultaneously a threat and an immeasurable opportunity.

Three things individuals should do: First, stop using AI as a "search engine." Start by handing the most difficult and time-consuming tasks at the core of your work to AI. Second, build the "muscle of adaptation." Getting accustomed to change itself matters far more than mastering any specific tool. Try new AI every day, probe its limits, and repeat experiments that rebuild your workflow. Third, redefine what only you can provide. Building physical trust relationships, making judgments that carry legal responsibility, high-level ethical decision-making. While becoming someone who commands AI, continuously ask: what value can only I offer?

For organizations, three things are urgent: redesigning organizational structure, cultivating agent managers, and investing in company-wide upskilling. Clinging to the pyramid structure is a path of gradual death. Don't leave this to the IT department — find people with deep understanding of real business contexts, and launch a systematic program right now to develop AI orchestration capabilities.

Shumer closes with this: "The future is already here. It just hasn't knocked on your door yet."

Scrambling after the knock comes is already too late. Whether you start preparing from this moment will determine the futures of individuals and companies alike.

Ryuta Hamamoto, TIMEWELL

---

References

• XenoSpectrum. (2026, February 12). AI company CEO warns "something big is happening."
• OpenAI. (2026, February 5). Model Release Notes.
• Ball, D. W. (2026, February 12). On Recursive Self-Improvement (Part II). Hyperdimensional.
• Mowshowitz, Z. (2026, February 12). AI #155: Welcome to Recursive Self-Improvement. Don't Worry About the Vase.
• METR. (2026). Research.
• The Economic Times. (2026, February 5). Anthropic CEO warns: 50% of entry-level white-collar jobs could vanish in 5 years.
• PwC. (2026, January 29). No more pyramids: Rethinking your workforce for the agentic AI era.
• Srinivasan, S., & Wei, V. (2026, February 12). To Thrive in the AI Era, Companies Need Agent Managers. Harvard Business Review.
• Bedard, J., & Beauchene, V. (2026, February 4). AI Transformation Is a Workforce Transformation. Boston Consulting Group.