I know what collapse feels like from the inside. At forty-two, I had built a company from scratch with borrowed money — SMAC Data Systems, founded on $5,000 across four credit cards — watched it grow to over $25 million in revenue, and then lost every cent of it. Not a slow erosion. A full collapse. The kind that strips you down to the essentials and asks you to start over with nothing but whatever you had learned from failing. I tried again. Failed again. Rebuilt again. What I understand now that I did not understand then is this: human excellence is never a ceiling. It is always provisional. What you build today can be taken away tomorrow — by markets, by timing, by forces entirely outside your control.
The way AI is now affecting human professional excellence is not new. It is just faster, more thorough, and without the option of a third attempt.
The most profound shift underway in artificial intelligence is not the rise of larger models or the explosion of compute. It is the emergence of recursive self-improvement — the moment when a system becomes capable of refining its own internal processes and using each improvement as fuel for the next. Human civilisation has never confronted a phenomenon like this. Machines can run thousands of experiments in parallel. When such a system begins to improve its own reasoning, code, and internal architecture, the result is not incremental progress. It is compounding, discontinuous progress. It is unstoppable.
During training, frontier models continue to discover more efficient internal representations without any human guidance. During fine-tuning, they refine their reasoning strategies. During inference — the very moment they are being used — they engage in self-correction, self-reflection, and tool-assisted reasoning. The boundary between "training" and "using" is dissolving. Each improvement in reasoning enables more efficient internal processes, which enables faster learning. The cycle feeds itself.
Policymakers imagine gradual change. Executives plan for disruptions that unfold in years, not days. Universities prepare students for careers that are already being automated. Recursive self-improvement does not operate on human timelines. It operates on machine timelines.
Last week's dispatch drew more responses than I expected. Several readers asked the same question: "Ashok, if you saw this coming in 1987, why does it feel like a surprise?" The honest answer is that I didn't fully see it. I saw a direction. I felt it the way you feel a tide before you see it. The AI winter convinced me — as it convinced everyone — that the direction was wrong. It wasn't wrong. It was early.
That distinction — between wrong and early — is the most important lesson I've carried from forty years in technology. And it is the one that nobody in a position of institutional power seems to have learned.
Recursive Self-Improvement: How It Works
- Training — Models discover efficient internal representations without human guidance
- Fine-tuning — Models refine their own reasoning strategies
- Inference — Real-time self-correction during active use
- Synthetic data — Models generate their own training material on problems humans have never solved
- Agent loops — Sub-agents handle subtasks; outputs feed back into the main model
- Result — Each improvement accelerates the next. The cycle has no natural ceiling.
The Opacity Problem
- Researchers cannot explain why GPT-5.2 reasons so well
- Internal representations cannot be mapped with precision
- Emergent behaviors cannot be predicted in advance
- The more powerful the system, the less interpretable it becomes
- RLHF aligns outputs, not internal goals — a veneer, not a foundation