Wednesday 12 August 2026
From Prediction to Production
We are getting better at predicting the future. But what happens when prediction becomes an instruction to the present?
Tomorrow, the Moon’s shadow will cross Europe, reports just about every news outlet in the Western hemisphere. Mashable noted that taking a photo of the eclipse might damage your phone.1 Reykjavík hotels have doubled the prices for their rooms, »a 98% jump compared with the same day last year«, writes The Guardian.2 Millions of people already know where they will be standing when the event takes place.
»The future is no longer merely something we have forecast. It is something we have chosen«
That’s a remarkable thing to know. It’s also an old one. The Antikythera mechanism, built more than 2,000 years ago, could calculate the cycles of the heavens and predict eclipses.3 Human beings have spent a very long time trying to pull the future into the present. What is new is what we can do with it once we get it here.
A prediction used to tell us what was likely to happen. Increasingly, it tells us what to do before it happens. If the forecast says rain, we take an umbrella. If a flood model says a river will overflow, we build a barrier. If a disease model predicts an outbreak, we vaccinate. If a financial model predicts inflation, a central bank changes interest rates. The predicted future enters the present as a cause.
This is the basic logic of control: measure the world, model where it’s going, compare that future with a desired state, and intervene before the difference becomes reality. The better the model, the earlier the intervention can occur.4
The strange consequence is that a sufficiently powerful prediction can make itself look wrong. A successful flood barrier means the predicted flood damage does not happen. A successful police operation means the predicted crime does not occur. A successful epidemic intervention means the epidemic never becomes the event the model predicted.5
Prediction has become pre-emption.
And now we’re beginning to build machines that can do something beyond prediction. LLMs can generate possible texts; generative design systems can generate possible buildings; scientific models can generate possible molecules, experiments, and materials. Computational systems can increasingly produce not one forecast of the future but thousands of candidate futures, evaluate them, and select among them.6
That changes the question. We have spent centuries asking machines to tell us what will happen. The more interesting question is the new one: which of the futures available to us should we make real?
This is already how much of advanced design works. In architecture, for example, the computer can generate thousands of possible configurations and evaluate them against constraints such as energy, daylight, cost, structure, and material use. The architect is no longer simply drawing a future building. The machine is searching a space of possible buildings before any of them exists.7
The same logic is spreading into drug discovery, engineering, logistics, materials science, and autonomous systems. A model predicts, a generator proposes alternatives, an evaluator analyses them, a physical or economic process tests the winner, the result becomes new data – and the loop starts over.
And so prediction is beginning to disappear as a separate activity. Instead, it's becoming one stage in a larger process: predict → generate → evaluate → select → produce. And at the end of that process, the future is no longer merely something we have forecast. It is something we have chosen.
There’s a political question hiding inside this technical one. If machines can search enormous spaces of possible futures, who decides what counts as a good future? What is optimised, what is discarded, and whose preferences are encoded in the objective function? A system capable of finding the most efficient future is not thereby capable of deciding whether that future is desirable.
That may become one of the central political problems of the near future. We’re acquiring extraordinary powers to search the future before we build it. The difficult part will not be finding possible worlds. It will be deciding which ones deserve to exist.
Tomorrow, the eclipse will happen exactly as predicted. We cannot negotiate with it; we can only watch. Much of the future we’re about to build will be different. Increasingly, we’ll be able to see possible versions of it before they exist, compare them, and choose between them. The future may be arriving in advance not as a prediction, but as an option. And eventually, perhaps, the most important question will no longer be »what happens next«. It will be which future we decide to make real.
References
1 Kimberly Gedeon (2026) »The solar eclipse is this week – but don't take photos of it with your phone«. Mashable, 11 August, 2026. https://mashable.com/tech/solar-eclipse-2026-dont-take-photos-phone
2 Kalyeena Makortoff (2026) Hotel prices in Europe jump before total solar eclipse. The Guardian, 11 August 2026.
3 Tony Freeth et al. (2006), »Decoding the ancient Greek astronomical calculator known as the Antikythera Mechanism«, Nature, 444, pp. 587–591. The paper established that the mechanism modelled complex astronomical cycles, including the Saros cycle used to predict eclipses.
4 Camacho, E.F. and Bordons, C. (2013), Model Predictive Control, 2nd edn. London: Springer. Model-predictive control explicitly uses models of future system behaviour to select present control actions.
5 National Institute of Justice (2009), »Predictive Policing: The Role of Crime Forecasting in Law Enforcement Operations«, U.S. Department of Justice. The report describes predictive policing as using forecasts of crime to guide preventive police interventions. https://www.ojp.gov/library/publications/predictive-policing-role-crime-forecasting-law-enforcement-operations
6 Lu, C. et al. (2026), »Towards end-to-end automation of AI research«, Nature, 651, pp. 914–919. The paper describes The AI Scientist, an autonomous system capable of generating research ideas, implementing and running experiments, analysing results, writing manuscripts and performing automated peer review.
7 Cagan, J. et al. (2013), »Generative design: A paradigm for design optimization«, Journal of Mechanical Design, 135(11). The paper describes computational systems that generate and evaluate alternative designs against specified objectives and constraints.