AI predictions come without numbers. Six deployments come with them

Over seven days we measured six systems that are already doing work, and listened to the people who build them describe what comes next. The deployments arrive with denominators. The forecasts do not.

What the machines did, with the numbers attached:

• Weather: a cyclone track error of 230 km at five days against 370 km for the European ensemble, worth about 30 hours of extra warning.
• Driving: 220.6 million miles with no human at the wheel, and 82% fewer injury crashes than the human benchmark, which is 707 crashes that did not happen.
• Supply chains: a 30-billion-parameter model deciding Nvidia's weekly material allocation at 86.7% accuracy, against 55.5% for the company's own 550-billion flagship.
• Mathematics: 106 of 226 open problems solved on one benchmark, 43 of them completely, and 120 still open.
• Serving: forty concurrent users on four accelerators after tuning, against fourteen before it.
• Targeting: 13,000 targets processed in 38 days at the Pentagon, one every four minutes and twelve seconds.

Every one of those numbers has an edge. In each case a neural network is doing a narrow job with a known answer, and the answer is scored. A track error is measured against another model's track error. A crash rate is measured against a human benchmark on the same streets, and the operator publishes both. An allocation decision is either the one a planner would have made or it is not. You can lose an argument about any of them, which is what makes them useful.

The counter-example matters as much. In drug discovery the AI advantage shows up in early screening and disappears at phase 2, where the molecule meets a human body. A field that publishes its failures produces claims you can price. A field that publishes only its wins produces slogans.

The lab that chose not to

Now the future, in the words of the people best placed to know it. OpenAI's chief scientist Jakub Pachocki explained in an essay, quoted by The Decoder, why his lab left mathematics on the table, and the sentence is a budget decision rather than a prophecy.

“We believe we could make the models better at specifically mathematics research with additional focus, but we do not prioritize this direction because of the urgency we feel about RSI and automated alignment research.”

— Jakub Pachocki, OpenAI, essay An Alien Mind, September 2026

Jakub Pachocki, OpenAI, in the essay An Alien Mind, September 2026

Read it as an admission about limits. The strongest maths model in the world is a side effect of work aimed somewhere else, because capability grows where a lab points its compute and nowhere else at the same time. That is the shape of the next few years as the people spending the money describe it.

Jensen Huang described his own vantage point at a Goldman Sachs conference on Thursday: Nvidia tracks every gigawatt of land, power and data centre shell on the planet, and every cloud, builder and lab reports back. He forecasts 70% revenue growth from that order book, which is a forecast with a denominator behind it even when the headline number sounds unhinged.

The claims with nothing under them

The risk claims run the other way. An Anthropic researcher put the odds of AI destroying humanity above 10% this decade, a number with no test attached and no way to be wrong before the fact. Mathematician Terence Tao made the more useful version of the worry at this year's International Congress of Mathematicians: if machines produce proofs faster than people can check them, the scarce resource stops being proof and becomes judgement about which results matter.

The sceptics are worth reading for the part data cannot settle. Guardian columnist Adrian Chiles wrote this week that there is a difference between labour-saving and lobotomising, and he is right that no crash statistic answers what happens to people who stop practising a difficult craft. His empirical claim lost to the Waymo numbers. His cultural one is still standing.

A rule for reading the next twelve months

So here is a reading rule for the next twelve months of announcements about the future of AI. Ask what the number is divided by. Two hundred and thirty kilometres against three hundred and seventy has a denominator. Eighty-six point seven per cent against fifty-five point five has a denominator. Superintelligence by a given year has none, and neither does a percentage chance of the end of the world.

The systems in the list above will keep improving in narrow, checkable steps, and the speeches will keep running ahead of them. Both things are true at once, and only one of them ships.

This article is for informational purposes only and does not constitute investment advice.