Recursion

I started a commercial apprenticeship at Hewlett-Packard in Widen on 13 August 1990, aged sixteen. More than 30 years in technology took me through HP, Fujitsu and Cisco, then into entrepreneurship and the work of building humAIne.

I watched the internet, mobile, cloud and big data reshape business. Each wave attracted the same claim: "this changes everything." The claim was broadly right, though rarely in the expected way. AI adds a loop those earlier waves did not have: it can help improve the chips, code, architectures and training methods used to build the next generation.

That loop is recursive self-improvement, and it deserves a clear-eyed look.

A machine that rewrites its own blueprint

Recursive self-improvement is the idea that an AI system can help improve its own code, architecture or reasoning.

An AI system helps improve its code, architecture or reasoning. The improved version then contributes to the next round. Each cycle can make the following cycle faster or more capable, with less human work inside the iteration.

Every other technology in human history has been limited by the speed at which humans can iterate. You design a chip, you test it, you find the flaws, you redesign. That cycle takes months. Years. Careers. But when the thing doing the designing is also the thing being designed, and it's getting better at designing with every cycle, the constraint disappears.

Pieces of it are already happening.

Google's DeepMind used AI to design the next generation of its own chips, and those chips were better than what their best human engineers produced. AlphaCode writes software that competes with the top tier of human programmers. AI systems are now training other AI systems, optimising architectures, pruning inefficiencies, and discovering approaches that no human would have thought of.

These systems are still early.

The curve that breaks your intuition

The acceleration trips people up.

We're wired to think linearly. If something improves by 10% this year, we expect roughly 10% next year. That's how salaries work. That's how most things in our daily lives behave.

Recursive improvement can behave differently.

When one improvement makes the next improvement faster, the curve can turn exponential. Human intuition tends to extend the latest rate in a straight line. That encourages us to expect gradual change and assume we will adjust in time.

That assumption deserves testing.

Ray Kurzweil has been talking about this for decades, the singularity, the intelligence explosion, the point at which artificial intelligence surpasses human intelligence and starts accelerating away from us at a pace we can't follow. For years, people called him a dreamer. Now the biggest companies on the planet are racing toward exactly what he described.

The gap between GPT-3 and GPT-4 was about two years. The capabilities jump was staggering. Now imagine that same magnitude of jump happening every six months. Then every month. Then every week. That's what recursive self-improvement looks like when it hits its stride.

This need not happen tomorrow for the curve to matter. We are already on it, and readiness has no effect on its direction.

What's already real

The evidence matters more than the storyline.

AI-assisted chip design at Google reduced design cycles from months to hours. NVIDIA is using AI to design the next generation of GPUs, the very hardware that runs AI. That's recursion in the physical world.

Meta's AI research lab built systems that can write, test, and debug code. GitHub Copilot already writes a significant chunk of production software at companies worldwide. Developers who use it are becoming dramatically more productive.

In drug discovery, AI models are identifying molecular structures in days that would have taken medicinal chemists years. AlphaFold cracked the protein folding problem that had stumped biology for fifty years.

And in AI research itself, this is the critical piece, AI is now being used to discover new training methods, new architectures, new optimisation techniques. The machine is literally improving the process of building machines.

Each of these alone would be significant. Together, they form a pattern that's hard to ignore.

What this means for everyone else

The consequences reach beyond AI researchers at DeepMind to workers, businesses and society.

Work changes now. Every knowledge worker is likely to have an AI co-pilot within the next few years. The practical divide will run between people who learn where the tool helps and people competing with those who did.

The work around humAIne uses AI for research, analysis, writing, coding and strategic planning. It helps me compress some tasks, but I do not yet have a controlled comparison that would support a headline productivity claim.

Creativity gets redefined. AI can remove some of the mechanical work around creativity. When the mechanical parts of creative work, the drafting, the formatting, the technical execution, are handled by machines, humans are freed to do what we do best: have original ideas, make unexpected connections, feel things deeply enough to create something meaningful.

The best musicians don't fear new instruments. They learn to play them.

Education has to change. Our school systems were designed to produce factory workers: sit down, memorise, repeat, test. That model was already failing. Recursive AI can give any student a personal tutor that is available 24/7 and never loses patience.

The classroom should concentrate on human development, social skills, emotional intelligence, critical thinking and collaboration. Standardised tests have pushed those subjects aside for too long.

Power concentrates, unless we actively prevent it. This is the risk I worry about most. If recursive self-improvement means that whoever has the best AI gets exponentially better AI, you get a winner-take-all dynamic. The gap between the leaders and everyone else can explode. Countries, companies, individuals.

That's not a future I want to live in.

The philosophical earthquake

The philosophical question is uncomfortable.

For all of human history, we've been the smartest thing on the planet. Everything we've built, art, science, philosophy, civilisation, rests on the assumption that human intelligence is the pinnacle. The ceiling.

What happens when the ceiling disappears?

When a machine can reason better than you, create more beautifully than you, solve problems faster than you, what is your purpose? What makes you you?

This can be liberating if we approach it with our eyes open.

I believe human value rests in consciousness and experience as much as intelligence. We feel things. We love, grieve and laugh at absurd jokes. A sunset matters because we are alive and know we will not be forever, regardless of its computational interest.

No machine has that. No machine, no matter how recursively improved, will understand what it feels like to hold your child for the first time.

I know. I have two sons. And the moments that defined me as a person had nothing to do with how smart I was. They had to do with how present I was.

Why I am building humAIne

I founded the pre-incorporation venture humAIne for this question. The name combines human and AI, with "AI" capitalised inside the word because it is becoming part of how we work and live.

My driving belief, the thing I wrote on a piece of paper years ago in a personal workbook that I still have, is this: Eine Welt, wo Menschen im digitalen Zeitalter wieder miteinander Mitgefühl geben. A world where people in the digital age show each other compassion again.

That is a pro-human position.

The risk I worry about most is human passivity: outsourcing our thinking, creativity, decisions and eventually our agency to systems that optimise for efficiency rather than meaning.

I've spent my career, from HP to Cisco to building businesses with my wife Sigrun, learning that technology is a tool for better lives, more connection and less fear.

We have to help shape how AI changes the world by making choices, setting boundaries and building systems we want to live with.

Where do we go from here

Four things need to happen.

We need AI literacy to become as fundamental as reading. It should teach what AI can and cannot do, how it makes decisions and where it fails, without demanding that everyone learn to code. That understanding is becoming essential.

We need regulation that's as fast and adaptive as the technology itself. The current approach, committee meetings about AI policy while the technology doubles in capability every few months, is like trying to regulate the internet with fax machines.

Productivity gains from recursive AI need to flow beyond shareholders and Silicon Valley. This requires political will, new economic models, and a level of cooperation between nations that we haven't demonstrated yet.

And we need to invest in what makes us human. In relationships, in community, in experiences that no algorithm can replicate. Because the more capable the machines become, the more valuable those distinctly human qualities will be.

The technology will advance whether we are ready or not. Our job is to protect creativity, compassion and human agency as it does.

Martin Uetz is the sole founder of pre-incorporation humAIne, focused on the intersection of people, technology and business. He writes from Switzerland and Iceland, usually with too much coffee and not enough sleep.