The technology that makes the others

Most technologies make us stronger at something. Fire gave us energy we could control. The wheel moved things. Factories multiplied labour. The internet moved information. Each one amplified a part of what people can do.

AI amplifies something different: reasoning. And reasoning is what produces every other technology. That's why I put it on the same short list as fire, the wheel, industrialisation and the internet, and why I don't think of it as just a better kind of software.

Why intelligence is different

A machine that does physical work raises output. A machine that helps discover better machines raises the rate at which output improves. That difference is the core of how I see AI.

Progress has always depended on people understanding the world, guessing at explanations, designing solutions, testing them and turning what works into tools. All of that runs on human attention, which is limited. Researchers get tired, specialise narrowly and can only try so many ideas. Institutions help by putting many people together, but coordination has costs of its own.

If AI can take on parts of that work, such as exploring a proof, writing and testing software, checking a design or reading through results, the same people and money can attempt far more. Some of what they find will improve computing, algorithms, energy and AI itself.

I don't think that loop runs without friction. Experiments take time, compute and energy are finite, things have to be manufactured and verified, and returns can diminish. But a technology that can help improve the methods that make it better has never existed before, and it's why I take the bigger predictions seriously.

Why I watch mathematics

There's a big difference between a model repeating what people already know and a model helping find something new. Mathematics is where that difference is easiest to see, because a proof can be checked. A system can propose an argument or search for a counterexample, and a person or a formal verifier can confirm whether it holds.

So when AI systems contribute to research mathematics, not textbook problems, I treat that as evidence of a real capability. It doesn't prove an AI can design a medicine or a fusion reactor; each of those has obstacles of its own. But having no proof of a future outcome is not the same as having no reason to expect it.

Most research runs on the same loop as mathematics: state the problem, propose answers, test them, find the errors, refine, repeat. What changes from field to field is how hard the testing is. Code can be run. Proofs can be checked. Designs can be simulated. Biology eventually has to meet a real experiment. The easier a field is to verify, the sooner I expect AI to change it.

From chatbots to agents

I don't think the chat window is where this ends. Systems that use tools, read files, browse and carry out tasks in several steps already exist in paid products. They're limited and not always reliable, but the direction is visible. What I expect is that capable, affordable personal agents become normal: you describe the goal, the agent does the steps, and you check the work.

It may end up looking less like one giant model and more like a general system that hands work to specialised ones, one for mathematics, one for code, one for research. That's a guess about architecture, not a certainty.

Programming shows where this goes. People once wrote assembly and had to think about the machine directly. Languages like Python hid most of that. Now you can describe what you want in plain language and steer the result. You still need to understand the system, its architecture, security, costs and the law around it, but you no longer have to hand-write every layer underneath.

Who gets to build

This is the part that matters most to me. Serious work usually takes expensive equipment, the right institution, a network and money. Plenty of people with good ideas never get to test them.

Cheap, capable intelligence lowers that bar. A student far from a top university can go deep into hard questions. A small team can do what used to take a large one. Someone outside the usual circles can build a prototype and find out whether the idea holds.

It won't remove every inequality. Compute, money, labs and power will still be unevenly shared, and AI systems have biases of their own. But it can make the ability to try depend less on credentials. That cuts both ways: ideas should be judged on their reasoning and evidence. An argument doesn't get better because a famous person makes it, or worse because an eighteen-year-old does.

Prosperity, and its limits

Better farming fed more people. Machines multiplied what factories could make. Electricity changed everything it touched. Computing changed how we design and coordinate. AI can push all of these further while speeding up the research behind the next round. Medicine, energy, engineering and, eventually, space exploration all stand to change.

Intelligence isn't magic, though. A better way to design a medicine doesn't make the medicine safe. A better design doesn't build the factory. Abundant production doesn't guarantee fair distribution, and a system that can recommend a policy can't settle what people value. These are specific problems to work on, not general reasons to dismiss the idea.

When the text runs out

One question I keep coming back to: what happens when AI runs out of training data? The stock of good human writing is finite. People keep producing more, and data isn't used up when a model reads it, but what's needed may grow faster than the supply of genuinely new, high-quality material.

There's a second problem. More and more of what's online is written with AI's help, and a model trained carelessly on models' output compounds their errors and narrows what it knows. Synthetic data isn't bad in itself. A maths problem with a checked solution is useful whoever wrote it. The question is whether the data teaches something true and new.

I think the way through is learning from experience that can be checked, not just from more documents: proofs that verify, code that passes tests, simulations, robots in the world and, eventually, experiments that produce new measurements. Each has limits. A simulation is only as good as its model of the world, and a badly designed reward gets gamed. But running out of one kind of data isn't running out of ways to learn. A million unchecked answers teach nothing; a few checked ones teach a lot.

What I'm sure of, and what I'm not

Some of this can already be seen. AI writes and debugs software, helps with mathematics, uses tools and finishes tasks in several steps, and enormous amounts of money and energy are going into building it.

Some is extrapolation with a mechanism behind it: more of software, analysis and research being automated, agents becoming more capable, and discovery speeding up first in the fields that are easiest to verify.

And some is genuinely open: general intelligence, superintelligence, research run largely by AI, real abundance. I don't know the timelines or the architectures, and I won't pretend to. What I'm confident about is the direction. I think the next ten or twenty years could change the world far more than our instincts from history suggest.

Being excited about that isn't the same as being naive about it. I'd rather examine an ambitious idea than dismiss it for being ambitious.

Why it matters to me

If even part of this happens in the 2030s and 2040s, I'll be in my twenties and thirties. That's why physics, chemistry and mathematics feel like more than exam subjects to me. They're the foundations of the things I want to understand and, one day, help build.

AI isn't a reason to stop learning. It's a reason to learn more, because the people who understand these tools will decide what they're used for. I don't want to watch this period from the outside.

The future isn't settled. Working out what it could be, and helping build some of it, is what excites me most.