Photo Courtesy of EXANTE
In 2000 I was at Deutsche Bank, building a new trading platform, watching friends leave one by one for startups that were going to make them rich. They had stock options. They had retirement mapped out at forty. I stayed at the bank and felt the envy of a man doing sensible work while everyone around him seemed to be getting away with something.
Then the crash came, and the something turned out to be nothing. Pets.com went from a Super Bowl advert in January to liquidation in November. The Nasdaq lost most of its value. Thousands of companies died; the web did not. It carried on and built Google. I was still at the bank, and I have rarely been so grateful for a boring decision.
I mention this because I have seen the current mood before, and I recognise the smell.
The oldest trick in the industry
“Artificial intelligence” is a marketing term. It was coined in the 1950s by scientists who wanted money from the American military, and it worked—which is the only reason we still say it. Every decade or so the phrase gets dusted off, oversold, and followed by an “AI winter” when the funding leaves. We are on the fourth or fifth turn of that wheel, depending on how you count.
The technology underneath is real. Deep learning did something startling at the ImageNet competition in 2012, and within three years your phone could identify what was in a photograph. But real capability and honest expectation are different things, and the gap between them is where the money is made.
The pattern never changes. A genuine breakthrough arrives, and capital floods in behind it. The promises inflate to justify the valuations until someone is on a stage telling you half your engineers will be redundant by Tuesday. Then reality turns up, the money leaves, and the people still standing go back to work.
The radiologists who never disappeared
In 2016, Geoffrey Hinton—one of the people who built this technology—told a room that we should stop training radiologists because deep learning would beat them inside five years. He was mistaken. Cheaper, faster imaging meant doctors ordered far more of it, and demand for radiologists climbed. When a task gets cheap and quick, we tend to want more of it, not less.
I think about that every time someone tells me my profession is finished. The tools have landed for software this time. In 2025 the agentic coding assistants stopped being a party trick and started closing real work—taking an issue, writing the code, and opening a pull request you can then argue with. My engineers use them. They ship faster for it.
What has not happened is anyone being replaced. The tool still needs an adult in the room. Someone who can tell a plausible answer from the right one and who is on call when it falls over at 9:31 on a Monday. That adult is an engineer. There is simply more demand now for the problems we can afford to solve.
The point we keep forgetting
DeepMind did not write conventional software to crack protein folding, a problem that had defeated biologists for fifty years. It built a model that solved it, then published the structures of 200 million proteins for free. New antibiotics and malaria vaccine research have already come out of that one act. Nobody asked how many lines of code it took.
That is the lesson under the noise. We were never hired to write code—the code was always the means. The job was to solve someone’s problem, and it still is.
I have spent twenty-odd years mistaking the one for the other, and the machines have done me the favour of pointing it out.
Richard Forss is CTO of
This story was distributed as a release by Sanya Kapoor under