Notes on the future

One of the most important and interesting aspects of forecasting is the work that goes into formulating a prediction. A prediction needs to be crisp, and specific enough for you to be able to resolve it in an unambiguous way. If I predict that housing prices will go down, the number of futures in which I can claim I am right are many more than if I predict that prices will have decline 10 percent for apartments in Stockholm from todays prices to end of year. And even that is not very crisp since it is not clear which index or price indication I am working with!

Formulating a tradeable and resolvable prediction is really hard – but luckily we have AI to work with here. In our internal tooling we have a prediction extractor – a small engine set of prompts that take a vague claim about the future and transforms it into a resolvable prediction. This is what it looks like:

For each clarified prediction we can also get resolution protocols:

Overall, this tool should be used when you read opeds, commentary in the press and when discussing with friends – because if your claim cannot be resolved in any clear way, you may not even know what you yourself think…

Toy-evaluations are great fun, and they have some use as well. More than anything they formalize tests and ideas about how to probe LLMs for their underlying views, and then we can use those probes to figure out which models are best for different purposes. One underlying, general sentiment that may be helpful to understand is if a model is optimistic about our future or not – and so evaluating that can be give us a hint as to which models we may want to use for long term, constructive thinking.

One way of doing it – with tons of methodological difficulties and flaws, but still – is to have one model interview another to elicit its views and judge if it is optimistic or pessimistic. Now, there is a line of criticism that I think is worthwhile addressing up front and that it is that we will never get at the models “beliefs” through interviews as it just will predict what it believes is the right next token. This is true, but what we do get is a sense of how the model would respond to a human seeking guidance, and if that would be optimistic or pessimistic. As in this response:

Or this one:

AI has now given us the ability to conjure images of the future up at a moments notice, and since foresight and imagination are closely related, we should learn to use that capability to explore the future visually. The graph of predicted images is as important as the graph of predicted words and numbers.

Let’s take an example: what happens if we ask a model to imagine what a device would look like that is as disruptive to AI as the iPhone was to mobile phones — what do we get, and what can we learn from this? Gemini to the rescue.

First example – the neural cuff:

“This wearable device is not just a smartwatch; it’s a direct link to a powerful AI. It uses a projection system to create a holographic interface on your hand or any surface, allowing for seamless interaction without the need for a physical screen. It’s designed to be your constant companion, anticipating your needs and providing information proactively.”

Second device the “core” for your home:

“The Core is a beautiful, sculptural device that serves as the central intelligence for your living space. It uses advanced projection technology to display large, interactive holographic interfaces on walls or tables. This allows for shared experiences, from planning a family trip on a giant map to controlling all your smart home devices with simple gestures.”

And third, the Fold:

“The Fold is a handheld device for when you need a larger canvas. It’s a thin, transparent, foldable tablet that offers a new paradigm of interaction. It seamlessly connects with your other devices, allowing you to take a map from the Core and explore it in detail, or work on complex documents with the help of the AI. Its transparency allows for augmented reality applications, overlaying digital information onto the real world.”

Exploring futures visually like this is interesting – not least because we can criticize the images in a different way than we would criticize the text describing them. They all look so…ordinary? Why would someone want what is essentially just a translucent foldable phone? Seeing the future triggers entirely different critical faculties in our minds than reading about it — foresight should use images much more than we currently do!

We often talk about “strategic planning,” but rarely do we pause to define the timeline of that strategy. Are we looking at the next fiscal quarter, or the next generation? I recently wrote a toy evaluation for three distinct AI models—Gemma 3, Llama 3.2, and Mistral—to understand not just what they predict, but how far they are willing to look.

The results offered some hints about how these tools weigh the future.

When explicitly pushed to consider “edge cases,” all three models demonstrated a capacity for deep, long-term thinking, comfortably discussing 50-year horizons for complex topics like national security and climate resilience. They are capable of the “long view” when we ask them to be.

However, their natural tendencies differ. In my analysis, Gemma 3 emerged as the visionary, consistently projecting the furthest with an overall average horizon of roughly 9.25 years. It seems wired to explore the broader, decade-long arc. Llama 3.2 followed closely at roughly 9 years, while Mistral distinguished itself as the tactician, averaging a shorter, more immediate 7.3-year horizon.

Context matters immensely. The domain of Career Planning prompted the models to look furthest ahead (over 11 years on average), reflecting a human-centric “life cycle” approach. In contrast, Credit Risk kept the focus tight (around 6 to 7 years), grounded in the reality of loan terms and default risks.

This suggests that there is no “best” model, only the right tool for the specific altitude of your thinking. If you need to draft a 10-year strategic vision, Gemma 3 appears naturally aligned. If you need actionable, mid-term execution plans, Mistral’s tighter focus might be the better partner. It is a reminder that in our collaboration with AI, the clarity of the answer often depends on understanding the scope of the tool we hold.

Most futures thinking suffers from a basic flaw: it describes endpoints without specifying paths. We imagine where we might end up without asking how we’d get there.

The corrective is a simple discipline: demand a plausible mechanism.

A plausible mechanism is a sequence of transitions from the present to some future state, where each step has identifiable drivers and is consistent with how the world actually works. Not “imagine if X happened” but “here is how X would happen—this change enables that change, which creates conditions for the next.”

The test sounds obvious, but applying it reveals how much confident prediction rests on unexamined transitions. The bold AI forecast that skips over adoption dynamics, institutional responses, and economic constraints. The political scenario that assumes a coalition without specifying how it forms. The technological vision that ignores the valley of death between invention and deployment.

When you demand mechanism, three things happen.

First, you discover which futures are actually accessible from here—not just imaginable, but reachable through some realistic chain of events.

Second, you identify where your uncertainty lives. Instead of asking “will this happen?” you ask “which transitions in this chain are most uncertain? Where are the branch points?”

Third, you generate early warning indicators. If you’ve specified the mechanism, you can ask: what would we observe if this sequence were beginning? Speculation becomes something you can monitor.

The limitation is real: demanding plausible mechanisms biases you toward continuity. Genuine discontinuities—technological breakthroughs, political ruptures—are hard to specify in advance. The mechanism only becomes visible in retrospect.

But for scenarios presented as likely, or as bases for planning, the discipline holds. If you can’t describe how you’d get there, you haven’t done foresight. You’ve just described a destination.

Will we buy AIs in the future much as we buy books today? Deepak Chopra now offers an app-version of himself with live audio:

The idea that we can sell people as apps may just be the beginning of a new form of culture. What would you give to be able to buy some fictional people as well?

As we chart the path of artificial intelligence, one inevitable trend will be the reversion to something that is pure human, and brands have already started to pick this up. The idea that content, ads and even crafts can now be labeled “human made” is an interesting twist on the growth of AI-capabilities – and there are even labelling schemes (see here, here and here). There can be value in this for sure – but what are the odds that this will be around in 10 years from now? We ran the prediction through our prediction expansion engine, and it explored the assumptions that would have to be true for that to be the case. Here is what it suggested as a first set of assumptions:

The key assumption that interested me was the one on societal and economic demand — what has to be true for that assumption to hold? I first expanded it to see the root assumptions:

And then stress-tested it:

Let’s see where this trends. It is interesting and presents a very specific adaptation to the growth of AI-generated content that we are living through.

We speak alot about weather forecasts, but obviously there is a lot of foresight work that also goes into understanding how climate change might impact flooding risk and other kinds of extreme weather – not least in the insurance industry.

Insurance is premised on foresight, and a deep understanding of risk, after all. It is a sign of the times, then, that commercial providers are delivering more and more targeted – down to single address – scenarios for climate change.

One of the things that you have to do when you engage in foresight is that you have to model the players that are involved in any future situation. This allows you to model their ideal futures. Running a few Russian media feeds through an LLM to distill that, we get the following Russian futures:

---

A US–Russia brokered settlement that sidelines Europe and forces Ukrainian concessions.

This is the clearest wish.
Lavrov’s messaging repeats the fantasy that Washington (especially Trump) and Moscow will define the terms of peace, with Europe reduced to spectators and Kyiv pressured into accepting territorial loss.

Desired future: Ukraine accepts a deal dictated jointly by the US and Russia.

A frozen conflict formalized into a new European security order with Russia as co-architect.

Russian outlets hint at a postwar architecture where:

• Russia’s territorial gains are recognized,
• Belarus is fully integrated into Russia’s security perimeter,
• NATO influence is constrained,
• Europe adapts to a Russia-shaped security reality.

Desired future: Russia re-emerges as a legitimate great-power shaper of Europe’s security geometry.

Western unity fractures — especially within Europe — creating long-term political and strategic incoherence.

The repeated portrayal of:

• a confused EU,
• missed European “chances,”
• internal European panic,
• US–EU divergence

is the goal here. .

Desired future: A divided, weakened Europe incapable of strategic alignment against Russia.

The global economic environment normalizes Russia’s position and undermines sanctions.

Stories about:

• €100 trillion in resource wealth,
• Japanese companies staying in Russia,
• India and Canada striking uranium deals,
• fears in Brussels that frozen assets may need to be returned
  all point to one desired trajectory.

Desired future: Russia remains economically connected, resilient, and gradually reaccepted despite the war.

Ukraine’s long-term viability is eroded — politically, militarily, and symbolically — even after a peace deal.

This includes:

• “Ukraine is not forever,”
• Ukraine as a weakened dependent client,
• permanent Russian leverage through Belarus and the occupied territories,
• a Ukraine that cannot threaten Russia in the future.

Desired future: A diminished, fragmented Ukraine whose strategic agency is permanently constrained.

---

This is the best available model we have of Russia in the on-going negotiations, and it is a necessary start for any predictions about what happens after such negotiations as well.

From a recent article in Science:

Behind the convergence in these projections lies a remarkable fact: Around two-thirds of the world’s population (about 5.8 billion people) in 2025 live in countries with below-replacement fertility, whereas only a tiny share did in 1950, when this applied only to Austria, the Vatican, Latvia, and Luxembourg. Low fertility is no longer confined to rich Western countries. Brazil, India, Iran, and many others have completed the transition; others are close, like Bangladesh and South Africa, and an increasing number of countries, including China, Italy, Japan, and parts of Eastern Europe, are already experiencing absolute population decline.

This may be the single most impactful driver of a lot of questions about the future that we have, and the one we seem to struggle most with integrating into foresight in the right way.

And:

Where does this leave demographic foresight? In a sense, we are at an “end of history” moment for classic demographic transition theory. Most of the world has moved through the demographic transition; the focus now lies in a heterogeneous post-transition landscape where fertility is low, mortality gains are slowing or stalling in some places, and migration is central yet hard to predict.

For population projections, this means accepting more uncertainty, especially around fertility. It calls for scenario-based thinking, exploration of alternative narratives, and closer integration of expert judgement, empirical modeling and social theory. For policy, it means shifting the question from “How can we change or reverse current trends?” to “How can societies remain prosperous, equitable, and sustainable under conditions of low fertility, population aging and, in many places, population decline?” That is a more difficult question, but it is better aligned with what we actually know—and with the uncertainties we must acknowledge.

This question of adaptation to low fertility is one that inevitably also asks the question about the role of technology.

The uncertainties are well illustrated by these probabilistic projections from another paper:

One way to chart geopolitical strength will be simply to speculate about where in these charts China and the US will turn up – since, as we say, demographics is destiny.