When I listen to conversations about AI, they still tend to revolve around emails, summaries and the promise of taking a few tedious tasks off our hands. When I look at my working day, I see a technology that has already fundamentally changed how we develop software.
The gap between those two experiences has become enormous.
For many people, their impression of AI took shape in the early days of chatbots and has barely changed since: it makes things up, understands little and struggles with simple questions. Hallucinations still happen. But if that is the basis of your overall assessment, you are judging today's systems by the weaknesses of their predecessors.
AI now builds entire software projects, develops mathematical proofs and takes on work that previously required teams. People who experience this every day ask different questions from those who occasionally use it to draft an email. Yet much of what we hear about work, business and education still assumes that the next few years will look much like the last few.
I think that is a profound misjudgment.
01Why we keep talking past each other
Andrej Karpathy is one of the best-known figures in AI research. He was a founding member of OpenAI and led Tesla's AI work on autonomous driving. On October 11, he offered a rough estimate of how differently people now experience this technology.
By his reckoning, around six billion people have never used a language model. One to two billion occasionally use free services. Roughly twenty million work extensively with the best available models, mostly in software development and fields such as mathematics. Then there are about five thousand people in research labs who already have access to unreleased systems.
Four experiences of the same technology
Who uses language models, roughly · one dot = 10 million people
The numbers are estimates, but the distinction explains a lot.
If you tried a free version a few months ago and got wrong answers, the enthusiasm probably seems overblown. If you hand large parts of your daily work to an AI agent, you are more likely to wonder why the subject is not being taken much more seriously.
Both views are rooted in experience. But they no longer describe the same capabilities.
02AI has turned software development upside down at our company
On January 20, 2026, Anthropic CEO Dario Amodei told the World Economic Forum in Davos that AI could take on most, perhaps all, of a software engineer's work within six to twelve months. Even then, he described engineers at Anthropic who let the model write the code while they edited it and handled the surrounding work.
Less than nine months later, I recognize much of that in our own company.
At 506.ai, where I work as a Product Owner, we now deliver several times what we used to, in less time. Our developers are still with us, but their work has fundamentally changed. AI handles more and more of the implementation, which makes defining tasks precisely, reviewing results and deciding how the pieces fit together all the more important.
Our experience does not tell us how far the entire industry has come. But it shows me how much of what sounded like a radical prediction a few months ago is already part of everyday work.
That changes the economics, too. A small team can suddenly take on projects that once required a much larger company. That has consequences for pricing, the scale of projects and competition.
I suspect we will see similar changes across much of the economy over the next few years, particularly wherever work is predominantly digital. Software development already gives us a sense of what that looks like.
03What mathematics tells us about creativity
The boundaries are shifting beyond software development, too.
On October 6, OpenAI published 722 mathematical papers presenting results from an internal model. Many of the proofs were also formalized in Lean, allowing computers to check their formal correctness. The papers have yet to be fully checked, and some have already been withdrawn.
The reactions nevertheless reveal what is at stake. Terence Tao pointed to clever new ideas, while also criticizing the fact that the people involved in producing the proofs were not available to answer questions, give talks or explain the results. Fields medalist Hugo Duminil-Copin described himself as “paralysed.”
A common objection to language models is that they can only reproduce and recombine existing knowledge. They lack the creativity to produce genuinely new insights.
The mathematical results challenge that assumption. If a model discovers a previously unknown, correct proof, new knowledge has been created. Saying that it combined things already known may explain how it got there. It does not diminish the result.
How much of human creativity consists of connecting familiar ideas, trying things out and occasionally stumbling upon a solution? Perhaps the distinction between human creativity and a machine combining what it already knows is less clear-cut than we have made it out to be.
04What can be checked can be improved faster
AI is advancing particularly quickly where results can be checked automatically. Whether software passes a test, a website loads faster or a formalized proof is correct can be established by checking the outcome.
That feedback allows systems to try something, evaluate the result and keep working. With suitable tasks, many such attempts can run simultaneously.
Karpathy sees this as one reason perceptions differ so widely: capabilities are furthest ahead where work is digital, verifiable and economically valuable. Outside those fields, much of the progress can go unnoticed.
Anthropic described what this looks like in practice in September. The team wanted to make claude.ai faster and handed much of the work to Claude. More than three thousand changes were implemented within two weeks. At times, over 150 Claude instances were working simultaneously. According to the team, core workflows on the website and in the desktop app became roughly three times as fast. Humans set the goals and reviewed the changes.
Once Claude could measure something, it could work on improving it. So the team kept looking for more things to measure and improve.
Usually, a team has to decide which problems deserve its limited time. Systems like these make it possible to explore far more options, find more bugs and implement more improvements.
Passing a test does not make a product good. But wherever useful checks exist, development can move much faster. Companies still planning around familiar costs and timelines will face competitors who already work very differently.
05What used to cost too much becomes possible
The implications go well beyond saving time. When certain kinds of intellectual work become substantially cheaper, things that have been left undone suddenly become worth doing.
Every business knows the problem: software has known flaws, processes are needlessly cumbersome, good ideas never make it beyond a meeting. The work would take too long, the budget is too small or something else is more urgent.
As the cost of implementation falls, the calculation changes. Improvements become feasible that nobody would have paid for before. Products for small audiences can become viable. Companies can build their own applications where they previously lacked the budget.
My own side projects, a website about a Magic: The Gathering format and a tool for tracking the housing market, would have been difficult for me to build without these tools.
This could allow far more to be created than before. At the same time, existing business models come under pressure. When a service can be delivered much more cheaply, customers will eventually expect to pay less. When a small team can do the work of a much larger one, staffing needs change.
No profession has to disappear entirely for this to matter. Vacancies can go unfilled, fewer contracts can be awarded and fewer people can be hired into entry-level roles. The effects may already be felt in the job market while we are still debating whether AI can replace people at all.
06We are still thinking too small
What troubles me is how little of this seems to register outside the AI bubble. There is plenty of discussion about individual products and their flaws. There is surprisingly little about what it means for society when more and more intellectual work can be done cheaply and at enormous scale.
Careers, business models and the skills people will need all depend on the answer. We are making plans for education and business based on conditions that are changing before our eyes.
I believe the next few years will bring far more upheaval than most people expect. In digital industries especially, assumptions that have held for years could cease to hold very quickly. For some, this will create enormous opportunities. For others, it will be brutal. We cannot know exactly how fast it will happen or who will be affected first. That uncertainty makes it more urgent to take these questions seriously.
William Gibson found the words for it:
The future is already here. It's just not evenly distributed.
A small group is already working under conditions that still sound like the future to most of the world. Is the world ready for this? I don't think most of it is.