LLMs are beginning to outperform human forecasters. If the trajectory holds, AI will rival superforecasters on all but the most complex of questions by some point next year. Even if all progress ceases tomorrow, we would still have in our hands today a set of tools and processes which produce accurate, accessible, abundant information about the near future.
All else being equal, it won’t matter much. Increasing the scale, speed, and accuracy of forecasting will not make forecasts useful to people who make decisions. I want to offer my reflections on why this is the case as someone paid to think about the future by people who make decisions. In doing so, I will propose some routes forward which might make forecasting, especially AI forecasting, more useful.
tl;dr
Judgemental forecasting may be niche, but prediction is ubiquitous. All decisions are predictions: they encode a view of the future and the effect your actions are likely to have. The process which produces those predictions is unconscious and often wrong, but it’s so load-bearing that forecasting barely touches it.
The core of that process is relevance realisation, a never-ending cycle which produces your frame - the boundary between what matters to you right now and what can be ignored. Most forecasting takes place downstream of this process, which means it’s structurally detached from relevance.
Relevance realisation is generative: deciding where to pay attention, what matters to you, what should be inside your hypothesis space. But forecasting is evaluative: it’s remarkably good at answering questions and testing hypotheses. The category error, then, is in expecting an evaluative tool to do generative work.
Projects like AI 2027 show an alternative. By making scenario forecasts habitable and compelling it’s easier for people to climb inside and live through the unfolding drama. ‘Well-calibrated narratives’ are great for communicating the meaning of a forecast widely, and for internalising new, critical questions you hadn’t previously considered.
But even more useful for active decision-making is building the underlying model itself. The foresight discipline has found over many decades that not only are models more useful tools for people making decisions, but true influence comes from getting them to build the model themselves.
For any tool, process, or platform to be truly useful, it needs to do both generative and evaluative work. My suggestion is a ‘living thesis,’ a way to continuously create relevance, find leverage, and test the viability of potential actions. This is constant and cyclical, rather than linear, meaning it works in line with how the brain has evolved to interpret and act within the world.
There are already some signs that parts of this approach are emerging. What may be most difficult is letting go of Tetlock-style forecasts being ends in themselves; instead, I think they must become boring utilities like water or intelligence, powering living theses: Bloomberg terminals for the rest of us.
No one wants my golden goose
We have been able to produce well-calibrated forecasts for decades. People can learn how to make good predictions. We also know that with time, you can improve: judgement is a teachable skill. Surely every government, military, and the entire S&P 500 would be jumping at this by now? Not only has this not happened, but some institutions - Downing Street, for example - have gone as far as creating and then disbanding their forecasting operations.
This is a weird problem, and no one really knows what to do with it. Explaining why forecasting has not been adopted en masse is something of a cottage industry. The main ideas are not hard to find:
Ego: tracking your record is exposing and threatening to powerful people. Forecasting is status-disruptive.
(Mis)trust: your past average performance means little to someone making a single, high-stakes decision.
Format: default-mode forecasting is passive. But strategy is normative: ‘should we do X?’ rather than ‘will Y happen?’
Relevance: forecasting questions are interesting, yet not obviously relevant to the sorts of specific, secret, complex problems people/companies/governments actually deal with.
Utility: forecasts don’t inherently, automatically lead to action.
Beyond gambling, judgemental forecasting remains a niche activity disconnected from most decision-making. The community’s response has been quite rationalist-shaped: double down, make it more complex, try harder. But more of the same won’t bring about product-market fit. A major category error is causing us to solve the wrong problem.
First, all decisions are predictions
One of the obvious places where forecasting should be well-established by now is corporate strategy. In theory, companies create strategies to stay competitive and respond to the external environment. A good strategy requires (1) a useful model of what’s happening, (2) a judgement of what needs to change, and (3) a coherent plan to make it change. By default, lots of organisations are pretty bad at this: much of what passes for ‘strategy’ is, instead, mere assumption and wishful thinking.
What’s striking is that even bad strategy rests on judgements about the future. Whenever you make a decision about what you ‘should’ do, you unconsciously form a view of future events and the effects your actions will have. Those judgements may be partial, poorly-calibrated, or just flat-out wrong, but they’re still there. You may only see the decision, but look between the lines, and you’ll find an unarticulated view of the future which coheres with the actions you’re planning to take. In other words, all decisions encode predictions.
Once you see this, and especially with ‘should’ decisions - those which express some change you want to make, or something you need to do - it’s hard to unsee! This matters, because it means forecasting has been competing with an unconscious, deeply-ingrained process all along.
Judgemental forecasting has failed to take hold, but prediction itself is already ubiquitous. It happens automatically and below the level of awareness. Trying to make decisions - especially normative decisions - without some element of prediction is like trying to outrun your own legs. So why hasn’t forecasting been the obvious upgrade we’d all hoped?
Betting against BlackRock
I think judgemental forecasting is kind of like CBT: surface unconscious beliefs, make predictions explicit, and challenge them in the light of day. Incrementally, and by making this process conscious, you improve; albeit on the narrow dimension of calibrating the forecasts you’ve already decided to track. The end result - the forecast itself - is what matters, but it barely touches the process that determines what questions to ask in the first place. Forecasting hands you a very precise laser, but where you point it is up to you.
Figuring out what’s relevant is a harder problem to solve. It can’t be done computationally: to decide what’s irrelevant, you must first take the whole search space into consideration and assess everything, making the exercise pointless. Neither people, nor machines, can work this way without grinding to an indefinite halt. This is the ‘frame problem’ in cognition: how do you turn the ill-defined ‘large world’ into a tractable model of a ‘small world?’ Which parts of the territory should go on the map? What’s worth your attention?
In Naturalizing Relevance Realization, Johannes Jaeger, Anna Riedl, and their coauthors describe how the mind resolves this tension. Relevance emerges from an ongoing negotiation between opposing cognitive priorities: accuracy vs. flexibility, exploration vs. exploitation, generalisation vs. specialisation. Relevance is essentially your frame; the boundary you draw between what matters and what can be ignored for now.
According to this explanation, there is no final, objectively-correct frame through which to view the world. The mind sees only temporarily-useful frames, and they’re in a constant state of emergence and degradation. You continuously renegotiate from wherever you happen to be, in a never-ending cycle. As your context or your goals change, the frames you were using become less fit for purpose. Realising you need a better frame and updating is a moment of insight which allows for more effective action or a better grip on the situation.
For relevance, there is no ‘view from nowhere.’ Relevance is an inherently embodied process; it is created, not discovered. It’s specific to your context, your goals, and your position in the system. It requires skin in the game. Most forecasting, however, takes place downstream of this process, after the frame has been set. This means forecasting, as practiced, is structurally detached from relevance.
This all sounds a bit theoretical, so let’s look at an example. In A stock price is not a prediction, Matt Levine contrasts the purpose of stock markets with prediction markets, taking the bidding war for Warner as an example. While participants in prediction markets were able to bet on the likely winner of the acquisition, doing so was, in Levine’s view, adjacent to the needs of real economic actors. To a Warner shareholder like BlackRock, the question of who wins such a war is basically besides the point. They’re interested in the consequences of the deal; how it affects the share price, restructures the sector, changes their exposure, and so on. They’re not interested in general questions for general observers. BlackRock’s context and goals are specific; their frame is specific, and so, what’s relevant is specific to them alone.
Chief Investment Strategist Wei Li made this point directly: “point forecasting for strategic horizons is basically pointless, which is why we’re rolling out scenarios to navigate the longer term.” That doesn’t mean forecasts aren’t interesting or accurate. But BlackRock - like any actor with skin in the game - cannottake a view from nowhere, and especially not on events in general. Creating scenarios is a more involved process; they’re built from somewhere; you have to decide which dimensions matter most to you and to your decisions. As Levine says, “people with real exposure to economic activity are not making clean bets on isolated events, disaggregated from their consequences. The consequences matter.”
If you can't change how people think, change what they think with
When constructing relevance, people implicitly bring their experiences, expertise, and worldview to the task of identifying what matters. We could imagine this like weights and biases in an LLM: deep parts of the machinery which determine overall behaviour. Or, they might be like emotional schemas: the unconscious rules which shape your interactions with the world. Either way, relevance doesn’t come from thin air - it’s formed on top of a substrate: your fundamental orientation to the world.
This matters for forecasting, because intervening after relevance has already been locked in is just too little, too late. A higher-leverage, or even just plainly necessary, place to play is deeper in the system, at the level where relevance is created. If you could bring someone’s orientation more in line with reality, then everything downstream might fall more neatly into place.
Squint a little, and the steps we’ve discussed map remarkably well to Col. John Boyd’s OODA loop, another cognitive framework for decision-making:
Observe ↔ Orient ↔ Decide ↔ Act
Boyd was a US Air Force pilot who developed OODA to understand how decision-making works in dogfights. OODA describes a “nonlinear process with constant feedback and feed-forward channels,” a continuous negotiation between an actor (in this case, a pilot), their environment, and the actions they might take. For Boyd, orientation was the most important step: he even called it his ‘Big O.’ He saw that a pilot’s orientation was the product of their culture, beliefs, past experiences, personality, as well as their analysis of new information. As well as providing a complex set of filters producing potential decisions, orientation also fed backwards into observation - how the pilot screened unfolding events to form a working model of ‘what’s going on.’
Boyd's OODA loop
Tetlock-style forecasting today aims to intervene at the decision step. The assumption is that better forecasts naturally trickle through to better decisions. The expectation, I believe, is that better information causes people to update their worldview accordingly, which then leads to more effective decisions and actions. If only we were so rational!
Remember that there’s a two-way flow between orientation and observation, and it’s entirely possible that better observations don’t automatically follow through to better orientation. This is where I think forecasting has got stuck. Instead, forecasting is firmly evaluative work: it depends on the question and resolution criteria already being set, often by someone else. This means, by default, it’s downstream of deciding what matters.
Now, this can still be useful, but only if a) you’ve asked for a forecast on a specific question, b) you’re sure you’re asking all the right questions, and c) you plan to take action on the basis of a forecast. If these conditions aren’t met, it’s much harder to retroactively make the forecast useful. This isn’t what we’d want, but it’s the actually-existing state of play. This is a structural problem. Forecasting as currently practiced is not designed to intervene at the level of relevance or orientation, and so, it fails to influence decisions.
The category error
This takes us to the heart of the issue. Relevance realisation is generative work; it’s the business of determining where to pay attention, or, what should be inside your hypothesis space. It produces the frame. It is intimately connected to your goals, dependent on your worldview, and utterly specific to you. Moreover, it’s non-optional. Relevance realisation happens one way or another. If you’re not directly intervening at this stage then it will still happen automatically.
Judgemental forecasting, by contrast, is evaluative work. It’s remarkably good at testing claims and answering questions. But most often those questions are disconnected from any coherent goal, and they don’t come from you: they float unmoored on prediction markets or in forecasting competitions. Forecasting happens outside of relevance.
The category error, then, is in trying to make an evaluative tool do generative work. Forecasting lacks relevance because it is not involved in the process that creates relevance in the first place.
Well-calibrated narratives
A few weeks after launching AI 2027, Scott Alexander wrote that “almost a million people visited our webpage; 166,000 watched our Dwarkesh interview. People even made memes about us.” By the end of the year, according to Oliver Habryka, “5+ million read [it], 20+ million watched Youtube videos about it, and ~80 million people read about it in articles.”
As ‘infrastructure for communicating forecasts’ goes, AI 2027 is kind of the inverse of a prediction market. This makes for a useful comparison. Where prediction markets offer broad aggregation of (basically) any topic, AI 2027 is a deep investigation of precisely one topic. More importantly, the forecast itself is habitable: through narrative, scenarios, and sheer drama, you can climb inside and live through it, year by year. It’s easier to integrate, and generous, too: beyond headline forecasts, the project offers a rich, explicit model of how the world works today.
As Zvi wrote, “if you want to know what a recursive self-improvement or AI R&D acceleration scenario looks like in a way that helps you picture one, and that lets you dive into details and considerations, this is the best resource available yet and it isn’t close.”
I think habitability is hugely important for making forecasting more useful. Living through the drama of the forecast means you can easily internalise new, critical hinge points: competition with China, regulatory signals, the sights and sounds of scenario pathways. You might have decomposed the project into a standard set of forecasting questions for prediction markets, and got similar results, but absent any narrative meaning they too would have little broad appeal. I’d argue the headline forecast - superhuman AI in 2027 - only holds weight because the rationale is both rigorous and habitable to readers. The same prediction stated solely as a probability would fall flat. People wouldn’t know what to do with it; nor would it mean that much.
Instead, reading AI 2027 is closer to a generative act. It introduces people to new information they can actually parse; and, at a deeper level, offers a more powerful frame through which to understand and engage with the problem.
This does not mean abandoning rigorous forecasting. Quite the opposite: the research is immensely detailed. The team (including Daniel Kokotajlo and Eli Lifland) could hardly be described as amateurs! Narrative does not replace reasoning; it sits atop it. It does more than make the forecast accessible, it compresses its many variables and trade-offs into a format human beings can internalise. How you create the information is one thing; how you make it potent is very much another.
All (world) models are wrong, some (world) models are useful
This is a big step in the right direction. The habitable, narrative format is great for sharing and internalising a forecast, but it still has its limits. It’s someone else’s story, set within someone else’s frame. You become aware of new issues you might not have considered, but the decision about what to include in the first place is out of reach. You, reader, are still downstream.
(I'm not criticising AI 2027 here, quite the opposite: I think the team have landed on well-calibrated narratives as an important, and enduring foresight artefact for a specific purpose; where communication is more important than active decision-making. I’m excited to see what they might look like for the other issues of our time.)
That said, the process of building the underlying model must have been incredibly illuminating for the AI 2027 team. It often is, because you’re grappling with the core process of relevance realisation: deciding what’s worth your attention. This is an active process, and good models are useful tools. Here’s Michael Story reflecting on his time in Downing Street:
“During the pandemic, Dominic Cummings said some of the most useful stuff that he received and circulated in the British government was not forecasting. It was qualitative information explaining the general model of what’s going on, which enabled decision-makers to think more clearly about their options for action and the likely consequences. If you’re worried about a new disease outbreak, you don’t just want a percentage probability estimate about future case numbers, you want an explanation of how the virus is likely to spread, what you can do about it, how you can prevent it.”
A good qualitative model enables people “to think more clearly;” which means they internalise, and develop, a better frame for decision-making. Isn’t that exactly what we’re after? As Kay & King set out in Radical Uncertainty, decision-makers rely on ‘abductive explanations,’ not forecasts. These are handy narratives which explain ‘what’s really going on here,’ stories which point to the material factors swinging an issue one way or another. Good forecasting, incidentally, does the same; it involves building a model of how things work today - which often turns out to be way more useful than a fixed view of ‘where things are going.’ That’s because a conditional model which highlights points of leverage and uncertainty lends itself to action and decision. And pound for pound, discovering a new, important crux is probably more useful than forecasting on the issues you already know about.
We don’t have to start from scratch here. Figuring out how to make this process participatory is something the foresight discipline (forecasting’s older cousin) has emphasised for decades - with some success.
It might sound like ‘forecasting’ and ‘foresight’ are just different words for the same thing. The two actually represent quite distinct approaches to thinking about the future, which are “often seen to be at philosophical loggerheads.” Judgemental forecasting is rigorously quantitative, as we’ve seen, while foresight is much more qualitative and speculative. The foresight discipline has been established for longer and has had - it’s fair to say - more success in getting close to decision-making, especially in government and industry. Views differ as to why. Uncharitably, it might be that the qualitative haziness (or, uncertainty) of traditional foresight questions, which often look to five- or ten-year time horizons, serves as useful cover to people in power - who can just do whatever they were going to do anyway. Charitably, it might be that it gives non-specialists a practical, challenging, horizon-widening way to work through fundamental uncertainty. You can pick your poison. That the two disciplines aim to solve the same problem, and yet regard each other with suspicion, continues to be amusing to me. Call it academic politics, or the narcissism of small differences. But as we shall see, and even in the opinion of one Philip E. Tetlock, we really need both to figure out how to get along.
What the foresight discipline figured out is that if you want to influence people, they must build the model. In a lot of foresight work, the process is the product. The influence comes from a subtle transfer of ownership: as the decision-maker gets more involved in constructing, testing, and playing with a model, they increasingly feel like it’s theirs. They develop their own theory of change and causality; at which point, any decisions which follow, follow naturally.
Remember, of course, that this will happen with or without you. The problem has never been that people in power have no model of the world or a theory of how their actions will affect things. The problem is that they do, it’s just unconscious, untestable, and often wrong.
Handing someone a single point forecast merely explains what happens if they do nothing. But conditional models are a good way to capture uncertainty and to identify points of leverage for people to use. The downside, of course, is that the traditional way of doing this work is pretty costly. It still operates on a consultancy basis, which is hard to scale. Is there a better alternative?
They must build the model
A couple of years ago, the Forecasting Research Institute published a paper on Conditional Trees. The paper describes a method for “generating informative questions about complex topics,” by constructing simple Bayesian networks of highly-informative forecasts. It’s a way to build more illuminating models which reveal crux issues: the uncertainties which hold outsized influence on the topic at large. The authors of the paper selected long-term AI risk as an example issue for the study, but it’s easy to see how the approach could generalise. It involves first setting out an overall hypothesis, and then asking experts to identify what events or trends would significantly change the outcome. This is used to calculate a ‘value of information’ score, thus revealing the most informative questions - as well as how they connect to each other.
Doing this manually is pretty resource-intensive. The authors point to LLMs as one potential solution - and for sure, there will be a role for agents here. But in the light of relevance realisation - needing skin in the game, building the model yourself, and so on - I think some level of human effort here is the price of impact. The forecasts produced by this approach are not significantly different in format than those produced by prediction markets. But questioning, mapping, and constructing a model, by contrast, is relevance realisation made manifest. The process of building the model is more valuable than either the model or the forecast themselves.
A new goal for the community should be finding ways to scale the model-building process to make it intuitive and incisive for people who make decisions. The current approach (consulting) is too bottlenecked and expensive; but retaining some level of effort and social involvement will be critical for this work to be meaningful. This is an interesting design challenge: we’re shifting upstream to include the generative, hypothesis-creation space as well as the evaluative work of forecasting.
This will not be either/or, but both/and. Remember, cognition works by cycling between the generative work of deciding what’s relevant, and the evaluative work of making predictions in order to act. Each moves the other, and each is necessary for the other. This is a good opportunity for the two schools of foresight and forecasting to integrate their relative strengths. The world of foresight has typically been good at generative work, widening the hypothesis space; while the forecasting community, as we have seen, is measurably good at evaluative work. Bringing the two together - into a tool, platform, or process - might genuinely (and literally) close the loop on this weird problem of irrelevance. There are signs of life here. Early experiments point to the emergence of parts of such a system. This is encouraging: the answer might well be right in front of us.
In practice: the living thesis
My modest proposal for what this may look like is a ‘living thesis,’ a way to continuously create relevance, find leverage, and test the viability of potential actions. As we’ve seen, trying to replace parts of this process from the outside simply doesn’t touch the sides. The generative-evaluative cycle is already running in the minds of the people you want to influence. This default state will always win out because it forms a fully functioning, complete loop. No one would claim this is optimal, but it is locally optimal: a complete, bad process will always beat an incomplete good one. The natural experiment of trying to make forecasting impactful shows that the mind just routes around better information if it’s not part of this already-running loop. You can either upgrade the entire cycle, or none of it.
Any tool must externalise the Whole Thing, in one place. to work across the cycle in its entirety. It should help you:
Determine what’s relevant for you for the issue at hand. It should be easy for a non-specialist to build their own model of the system, i.e. (a) what are the material factors at play and (b) how do they influence each other? This is not a million miles off a successful method in foresight practice called ‘futures landscapes,’ which involves carrying out a horizon scan and then mapping trends into a system. Over many years, this was found to be more accessible than traditional scenario methods, because people conceptualise a system as a kind of ‘emergent present’ rather than distinct, incompatible futures. The approach set out in conditional trees is a good complement: the specific mode of decomposition, finding high value-of-information questions, will help to identify crux issues in the system. Where the FRI’s work here was manual and resource-intensive, I expect we will soon be at a point where LLMs can carry a significant amount of the load, under direction. The function of this step is widening the hypothesis space and connecting the pieces together. But what really matters is that this ongoing process gets significant involvement from the people you want to influence. This is the generative work of relevance realisation made explicit.
Surface decisions and actions you might want to take. Analysing the dynamics of the system will do two things. First, it will give you a coherent and credible view of ‘how things will play out.’ This is a bit like the model underlying AI 2027. But second, it will also surface points of leverage which matter to you. This is a bit like Michael Story’s pandemic modelling in Downing Street, or Boyd’s OODA loop creating a “menu of responses.” The question becomes ‘what happens if we push on this bit?’ or ‘how do we make [objective] more likely?’ In other words, you can start to see which decisions and actions could meaningfully push the system one way or another. This step is where forecasting activities start to transition into better decision-making, i.e. the work of real strategy.
Test the viability of those actions. Finally, you’ll want to make a robust assessment of how your decision or action will perform under different conditions. This is a great task for judgemental forecasting operating within a larger model: ‘what happens if we do X? How would it affect the system at large?’ Here, you treat your decision or action as a new variable to be inserted into the model, and re-run the forecast(s) to see how everything else changes. On the flip side, an advantage of having a model of the system is you can also play with the uncertainties outside of your control. If an important factor swings one way or another, it changes the overall scenario. In this way, you can stress-test how well your plans perform; how brittle, resilient, or exposed you are to the system at large.
These are, I think, the bare bones of what’s structurally necessary for the practice of forecasting to be useful and relevant. It’s a better, more illuminating way to construct a thesis about the context you’re in and the actions you may wish to take.
What makes this a living thesis is that the context is always changing, the ‘frame’ of relevance realisation is always degrading, and so, the cycle must always keep running. Step three is not the end of a linear process, it is just an element in a continuous exercise of monitoring the situation, placing irons in the fire, and gathering enough confidence to commit. Taking action changes the dynamics of the system from within; and so, the frame degrades, the model updates, by which point new actions appear possible, and on, and on, and on.
In the Before Times, maintaining and updating a live world model was a highly resource-intensive activity. In practice, you’d pay a team of consultants to do it every few months (if you even did it at all), which is both expensive and ineffective. Massively reducing this cost - and making model updates daily - is non-negotiable for this approach to scale. I think this involves hitting the ‘fix everything easily’ button and using agents to do some, but not all, of the work. Specifically, agents should scan for evidence that the model is becoming unfit for purpose: perhaps an overnight decision by a regulator is perturbing the dynamic of the system, or an uncertain issue has swung one way rather than the other. But the decision on whether to incorporate that evidence and update the model must rest with the human. This is a form of meaningful friction that ensures relevance is maintained. We should design for skin in the game.
If it sounds never-ending, that’s because it is. The edges of the system are always exposed to new things. You can’t pause the world while you decide what to do, and yet the tooling we have today assumes that this can ever be the case. Instead, our tools should accept both the ever-changingness of the world and the constant cycling of the cognitive process we’ve evolved to avoid getting stuck in place.
Parts of this are already emerging
Doing any amount of this work, at scale, is going to require highly-reliable predictions and reasoning on tap. Probably the best-known (and perhaps best-funded?) AI forecaster is Mantic, a London-based startup which has achieved impressive results in various tournaments on Metaculus. As far as I can tell, they’re in the midst of trying to find commercial applications beyond winning forecasting competitions - although that seems to be going okay for the Polymarket bots which make up 5% of wallets but over 75% of trading volume. I’m reminded here of Scott Alexander’s quip, that perhaps “prediction markets’ role in God’s plan was only to provide the foundation for AI superforecasters - the training data, the benchmarking arena, and the pot of money that rewards innovation.” Mantic and others are solving the supply-side problem of turning really good forecasts into a commodity - it’s just a weird quirk that the demand-side doesn’t really need more of the same. Instead, I think AI forecasters may just become APIs powering living theses, or something similar.
In February this year, Metaculus announced Radiant, “a networked forecasting tool for strategic decision-makers.” The tool is like a whiteboard where you can link forecasts together to visualise how their relationships affect a domain: in essence, building a living map of a system. This is supported by AI-assisted question generation[1], which makes it easier to formulate questions and identify connections you might otherwise have missed. Radiant, though still in prototype form, gets closer to what I think is necessary for forecasting to touch relevance realisation. It’s also good to see Metaculus move a level of abstraction upwards beyond the traditional format of a prediction market. What’s missing is still skin in the game: you need to be able to form your own plan and throw it into the model to see how it performs under uncertainty. Otherwise, Radiant will remain a systems-mapping tool, rather than becoming a strategy and leverage tool.
Perhaps the closest thing to what I’ve articulated is Deep Future, another early prototype which bills itself as “Deep Research for scenario planning.” Uncannily, its creator Gordon Brander seems to have landed on a similar diagnosis of what needs to change to make forecasting useful: “AI agents map out your strategic landscape, identify the forces driving change, explore multiple trajectories, and discover leverage points you can use to shape the space of possibility.” So, this looks like foresight-based scenario analysis on steroids, which is cool if you like that sort of thing. I’m particularly interested in the final step, in which “AI synthesizes insights from scenario testing to recommend adaptive strategies with built-in flexibility, creating decision frameworks that perform regardless of which future materializes.” Now, this all sounds great, but we… just don’t know yet, because Deep Future is currently waitlisted. It would be good to try it out and see if I could have saved myself (and you) several thousand words.
What may be more difficult than building the right tools, or interfaces, or even just backend integrations into Claude, is letting go of Tetlock-style forecasting and prediction platforms as ends in themselves. Perhaps you have come to believe that forecasting is way overrated and we should stop funding it, or maybe you take the firm view that forecasting is not overrated and it’s probably funded appropriately. Both positions assume that forecasting as currently practiced is in its terminal form. I think this would be a mistake. It will be hard to admit, but forecasting needs to become a boring utility like water or intelligence; something which powers stuff other people actually want to do. Yes, perhaps the role of prediction markets in God’s plan was to provide the training data for AI forecasters, but so too the role of AI forecasters will be to enable useful, beautiful, living theses: Bloomberg terminals for the rest of us.
Epilogue
While writing this piece I often wondered: why don’t Anthropic or OpenAI just eat this? It doesn’t seem beyond the capability of the engineers and designers at either organisation to build a system which helps you make radically better decisions under uncertainty. Surely, at least for enterprise customers, such a capability - if well-executed - would be enormously useful? Perhaps they’re both so focused on leapfrogging each other’s coding products that ‘IDEs for strategy’ are just further down the priority list. But I think that sharp, clear-eyed systems mapping and decision-making will only become more important as AI capabilities improve; not least because the speed and scale at which things may fuck up will increase by ~orders of magnitude over the next two years.
There is another alternative, though. Perhaps the final form for all technology is simply a chat window. I’m sceptical of this view - but what if it may never be necessary to build specialised interfaces for decision-making: living theses, scenario tools, IDEs, whatever - all by the wayside. What if, instead, forecasting finally finds its impact by diffusing into billions of conversations with generalised AI, invisibly, silently improving their accuracy without ever stating a probability? Putting fluoride in the water supply worked pretty well. You may never ask for a forecast again - and prediction markets may well devolve into casinos - but the essence of forecasting will still be there, buried in the system prompt for Opus 5 or GPT 6.
Earlier, I suggested that “if you could bring someone’s orientation more in line with reality, then everything downstream might fall more neatly into place.” Maybe that’s possible, and all it takes is a few edits to a markdown file, hidden deep underground, somewhere beneath the Bay Area. In the future, AI aligns you.
It at least looks like ‘AI for question decomposition’ is solvable. It will need to be, to lower the cost and experience barriers enough for non-experts to start directing their own forecasting and scenarios. Future Tokenizer, a prototype from Jordan Rubin, is a very interesting stab at a solution. He built the tool as an “IDE for human-AI collaborative thinking” which gets around the weirdly limiting problem of most AI interactions occurring in linear conversations. You start by articulating a question or a goal, and Future Tokenizer helps you to map out load-bearing assumptions, blind spots, and hidden cruxes, in (you guessed it) a system of nodes and branches. You can then take things further: red-teaming your map for alternative hypotheses before ultimately consolidating down into a robust plan of action. I think this level of decomposition and testing is necessary for any sort of scenario or systems tool to be useful; otherwise you risk hypothesis lock-in - forecasting well only on the issues you assume are important.
LLMs are beginning to outperform human forecasters. If the trajectory holds, AI will rival superforecasters on all but the most complex of questions by some point next year. Even if all progress ceases tomorrow, we would still have in our hands today a set of tools and processes which produce accurate, accessible, abundant information about the near future.
All else being equal, it won’t matter much. Increasing the scale, speed, and accuracy of forecasting will not make forecasts useful to people who make decisions. I want to offer my reflections on why this is the case as someone paid to think about the future by people who make decisions. In doing so, I will propose some routes forward which might make forecasting, especially AI forecasting, more useful.
tl;dr
No one wants my golden goose
We have been able to produce well-calibrated forecasts for decades. People can learn how to make good predictions. We also know that with time, you can improve: judgement is a teachable skill. Surely every government, military, and the entire S&P 500 would be jumping at this by now? Not only has this not happened, but some institutions - Downing Street, for example - have gone as far as creating and then disbanding their forecasting operations.
This is a weird problem, and no one really knows what to do with it. Explaining why forecasting has not been adopted en masse is something of a cottage industry. The main ideas are not hard to find:
Beyond gambling, judgemental forecasting remains a niche activity disconnected from most decision-making. The community’s response has been quite rationalist-shaped: double down, make it more complex, try harder. But more of the same won’t bring about product-market fit. A major category error is causing us to solve the wrong problem.
First, all decisions are predictions
One of the obvious places where forecasting should be well-established by now is corporate strategy. In theory, companies create strategies to stay competitive and respond to the external environment. A good strategy requires (1) a useful model of what’s happening, (2) a judgement of what needs to change, and (3) a coherent plan to make it change. By default, lots of organisations are pretty bad at this: much of what passes for ‘strategy’ is, instead, mere assumption and wishful thinking.
What’s striking is that even bad strategy rests on judgements about the future. Whenever you make a decision about what you ‘should’ do, you unconsciously form a view of future events and the effects your actions will have. Those judgements may be partial, poorly-calibrated, or just flat-out wrong, but they’re still there. You may only see the decision, but look between the lines, and you’ll find an unarticulated view of the future which coheres with the actions you’re planning to take. In other words, all decisions encode predictions.
Once you see this, and especially with ‘should’ decisions - those which express some change you want to make, or something you need to do - it’s hard to unsee! This matters, because it means forecasting has been competing with an unconscious, deeply-ingrained process all along.
Judgemental forecasting has failed to take hold, but prediction itself is already ubiquitous. It happens automatically and below the level of awareness. Trying to make decisions - especially normative decisions - without some element of prediction is like trying to outrun your own legs. So why hasn’t forecasting been the obvious upgrade we’d all hoped?
Betting against BlackRock
I think judgemental forecasting is kind of like CBT: surface unconscious beliefs, make predictions explicit, and challenge them in the light of day. Incrementally, and by making this process conscious, you improve; albeit on the narrow dimension of calibrating the forecasts you’ve already decided to track. The end result - the forecast itself - is what matters, but it barely touches the process that determines what questions to ask in the first place. Forecasting hands you a very precise laser, but where you point it is up to you.
Figuring out what’s relevant is a harder problem to solve. It can’t be done computationally: to decide what’s irrelevant, you must first take the whole search space into consideration and assess everything, making the exercise pointless. Neither people, nor machines, can work this way without grinding to an indefinite halt. This is the ‘frame problem’ in cognition: how do you turn the ill-defined ‘large world’ into a tractable model of a ‘small world?’ Which parts of the territory should go on the map? What’s worth your attention?
In Naturalizing Relevance Realization, Johannes Jaeger, Anna Riedl, and their coauthors describe how the mind resolves this tension. Relevance emerges from an ongoing negotiation between opposing cognitive priorities: accuracy vs. flexibility, exploration vs. exploitation, generalisation vs. specialisation. Relevance is essentially your frame; the boundary you draw between what matters and what can be ignored for now.
According to this explanation, there is no final, objectively-correct frame through which to view the world. The mind sees only temporarily-useful frames, and they’re in a constant state of emergence and degradation. You continuously renegotiate from wherever you happen to be, in a never-ending cycle. As your context or your goals change, the frames you were using become less fit for purpose. Realising you need a better frame and updating is a moment of insight which allows for more effective action or a better grip on the situation.
For relevance, there is no ‘view from nowhere.’ Relevance is an inherently embodied process; it is created, not discovered. It’s specific to your context, your goals, and your position in the system. It requires skin in the game. Most forecasting, however, takes place downstream of this process, after the frame has been set. This means forecasting, as practiced, is structurally detached from relevance.
This all sounds a bit theoretical, so let’s look at an example. In A stock price is not a prediction, Matt Levine contrasts the purpose of stock markets with prediction markets, taking the bidding war for Warner as an example. While participants in prediction markets were able to bet on the likely winner of the acquisition, doing so was, in Levine’s view, adjacent to the needs of real economic actors. To a Warner shareholder like BlackRock, the question of who wins such a war is basically besides the point. They’re interested in the consequences of the deal; how it affects the share price, restructures the sector, changes their exposure, and so on. They’re not interested in general questions for general observers. BlackRock’s context and goals are specific; their frame is specific, and so, what’s relevant is specific to them alone.
Chief Investment Strategist Wei Li made this point directly: “point forecasting for strategic horizons is basically pointless, which is why we’re rolling out scenarios to navigate the longer term.” That doesn’t mean forecasts aren’t interesting or accurate. But BlackRock - like any actor with skin in the game - cannot take a view from nowhere, and especially not on events in general. Creating scenarios is a more involved process; they’re built from somewhere; you have to decide which dimensions matter most to you and to your decisions. As Levine says, “people with real exposure to economic activity are not making clean bets on isolated events, disaggregated from their consequences. The consequences matter.”
If you can't change how people think, change what they think with
When constructing relevance, people implicitly bring their experiences, expertise, and worldview to the task of identifying what matters. We could imagine this like weights and biases in an LLM: deep parts of the machinery which determine overall behaviour. Or, they might be like emotional schemas: the unconscious rules which shape your interactions with the world. Either way, relevance doesn’t come from thin air - it’s formed on top of a substrate: your fundamental orientation to the world.
This matters for forecasting, because intervening after relevance has already been locked in is just too little, too late. A higher-leverage, or even just plainly necessary, place to play is deeper in the system, at the level where relevance is created. If you could bring someone’s orientation more in line with reality, then everything downstream might fall more neatly into place.
Squint a little, and the steps we’ve discussed map remarkably well to Col. John Boyd’s OODA loop, another cognitive framework for decision-making:
Observe ↔ Orient ↔ Decide ↔ Act
Boyd was a US Air Force pilot who developed OODA to understand how decision-making works in dogfights. OODA describes a “nonlinear process with constant feedback and feed-forward channels,” a continuous negotiation between an actor (in this case, a pilot), their environment, and the actions they might take. For Boyd, orientation was the most important step: he even called it his ‘Big O.’ He saw that a pilot’s orientation was the product of their culture, beliefs, past experiences, personality, as well as their analysis of new information. As well as providing a complex set of filters producing potential decisions, orientation also fed backwards into observation - how the pilot screened unfolding events to form a working model of ‘what’s going on.’
Boyd's OODA loop
Tetlock-style forecasting today aims to intervene at the decision step. The assumption is that better forecasts naturally trickle through to better decisions. The expectation, I believe, is that better information causes people to update their worldview accordingly, which then leads to more effective decisions and actions. If only we were so rational!
Remember that there’s a two-way flow between orientation and observation, and it’s entirely possible that better observations don’t automatically follow through to better orientation. This is where I think forecasting has got stuck. Instead, forecasting is firmly evaluative work: it depends on the question and resolution criteria already being set, often by someone else. This means, by default, it’s downstream of deciding what matters.
Now, this can still be useful, but only if a) you’ve asked for a forecast on a specific question, b) you’re sure you’re asking all the right questions, and c) you plan to take action on the basis of a forecast. If these conditions aren’t met, it’s much harder to retroactively make the forecast useful. This isn’t what we’d want, but it’s the actually-existing state of play. This is a structural problem. Forecasting as currently practiced is not designed to intervene at the level of relevance or orientation, and so, it fails to influence decisions.
The category error
This takes us to the heart of the issue. Relevance realisation is generative work; it’s the business of determining where to pay attention, or, what should be inside your hypothesis space. It produces the frame. It is intimately connected to your goals, dependent on your worldview, and utterly specific to you. Moreover, it’s non-optional. Relevance realisation happens one way or another. If you’re not directly intervening at this stage then it will still happen automatically.
Judgemental forecasting, by contrast, is evaluative work. It’s remarkably good at testing claims and answering questions. But most often those questions are disconnected from any coherent goal, and they don’t come from you: they float unmoored on prediction markets or in forecasting competitions. Forecasting happens outside of relevance.
The category error, then, is in trying to make an evaluative tool do generative work. Forecasting lacks relevance because it is not involved in the process that creates relevance in the first place.
Well-calibrated narratives
A few weeks after launching AI 2027, Scott Alexander wrote that “almost a million people visited our webpage; 166,000 watched our Dwarkesh interview. People even made memes about us.” By the end of the year, according to Oliver Habryka, “5+ million read [it], 20+ million watched Youtube videos about it, and ~80 million people read about it in articles.”
As ‘infrastructure for communicating forecasts’ goes, AI 2027 is kind of the inverse of a prediction market. This makes for a useful comparison. Where prediction markets offer broad aggregation of (basically) any topic, AI 2027 is a deep investigation of precisely one topic. More importantly, the forecast itself is habitable: through narrative, scenarios, and sheer drama, you can climb inside and live through it, year by year. It’s easier to integrate, and generous, too: beyond headline forecasts, the project offers a rich, explicit model of how the world works today.
As Zvi wrote, “if you want to know what a recursive self-improvement or AI R&D acceleration scenario looks like in a way that helps you picture one, and that lets you dive into details and considerations, this is the best resource available yet and it isn’t close.”
I think habitability is hugely important for making forecasting more useful. Living through the drama of the forecast means you can easily internalise new, critical hinge points: competition with China, regulatory signals, the sights and sounds of scenario pathways. You might have decomposed the project into a standard set of forecasting questions for prediction markets, and got similar results, but absent any narrative meaning they too would have little broad appeal. I’d argue the headline forecast - superhuman AI in 2027 - only holds weight because the rationale is both rigorous and habitable to readers. The same prediction stated solely as a probability would fall flat. People wouldn’t know what to do with it; nor would it mean that much.
Instead, reading AI 2027 is closer to a generative act. It introduces people to new information they can actually parse; and, at a deeper level, offers a more powerful frame through which to understand and engage with the problem.
This does not mean abandoning rigorous forecasting. Quite the opposite: the research is immensely detailed. The team (including Daniel Kokotajlo and Eli Lifland) could hardly be described as amateurs! Narrative does not replace reasoning; it sits atop it. It does more than make the forecast accessible, it compresses its many variables and trade-offs into a format human beings can internalise. How you create the information is one thing; how you make it potent is very much another.
All (world) models are wrong, some (world) models are useful
This is a big step in the right direction. The habitable, narrative format is great for sharing and internalising a forecast, but it still has its limits. It’s someone else’s story, set within someone else’s frame. You become aware of new issues you might not have considered, but the decision about what to include in the first place is out of reach. You, reader, are still downstream.
(I'm not criticising AI 2027 here, quite the opposite: I think the team have landed on well-calibrated narratives as an important, and enduring foresight artefact for a specific purpose; where communication is more important than active decision-making. I’m excited to see what they might look like for the other issues of our time.)
That said, the process of building the underlying model must have been incredibly illuminating for the AI 2027 team. It often is, because you’re grappling with the core process of relevance realisation: deciding what’s worth your attention. This is an active process, and good models are useful tools. Here’s Michael Story reflecting on his time in Downing Street:
A good qualitative model enables people “to think more clearly;” which means they internalise, and develop, a better frame for decision-making. Isn’t that exactly what we’re after? As Kay & King set out in Radical Uncertainty, decision-makers rely on ‘abductive explanations,’ not forecasts. These are handy narratives which explain ‘what’s really going on here,’ stories which point to the material factors swinging an issue one way or another. Good forecasting, incidentally, does the same; it involves building a model of how things work today - which often turns out to be way more useful than a fixed view of ‘where things are going.’ That’s because a conditional model which highlights points of leverage and uncertainty lends itself to action and decision. And pound for pound, discovering a new, important crux is probably more useful than forecasting on the issues you already know about.
We don’t have to start from scratch here. Figuring out how to make this process participatory is something the foresight discipline (forecasting’s older cousin) has emphasised for decades - with some success.
It might sound like ‘forecasting’ and ‘foresight’ are just different words for the same thing. The two actually represent quite distinct approaches to thinking about the future, which are “often seen to be at philosophical loggerheads.” Judgemental forecasting is rigorously quantitative, as we’ve seen, while foresight is much more qualitative and speculative. The foresight discipline has been established for longer and has had - it’s fair to say - more success in getting close to decision-making, especially in government and industry. Views differ as to why. Uncharitably, it might be that the qualitative haziness (or, uncertainty) of traditional foresight questions, which often look to five- or ten-year time horizons, serves as useful cover to people in power - who can just do whatever they were going to do anyway. Charitably, it might be that it gives non-specialists a practical, challenging, horizon-widening way to work through fundamental uncertainty. You can pick your poison. That the two disciplines aim to solve the same problem, and yet regard each other with suspicion, continues to be amusing to me. Call it academic politics, or the narcissism of small differences. But as we shall see, and even in the opinion of one Philip E. Tetlock, we really need both to figure out how to get along.
What the foresight discipline figured out is that if you want to influence people, they must build the model. In a lot of foresight work, the process is the product. The influence comes from a subtle transfer of ownership: as the decision-maker gets more involved in constructing, testing, and playing with a model, they increasingly feel like it’s theirs. They develop their own theory of change and causality; at which point, any decisions which follow, follow naturally.
Remember, of course, that this will happen with or without you. The problem has never been that people in power have no model of the world or a theory of how their actions will affect things. The problem is that they do, it’s just unconscious, untestable, and often wrong.
Handing someone a single point forecast merely explains what happens if they do nothing. But conditional models are a good way to capture uncertainty and to identify points of leverage for people to use. The downside, of course, is that the traditional way of doing this work is pretty costly. It still operates on a consultancy basis, which is hard to scale. Is there a better alternative?
They must build the model
A couple of years ago, the Forecasting Research Institute published a paper on Conditional Trees. The paper describes a method for “generating informative questions about complex topics,” by constructing simple Bayesian networks of highly-informative forecasts. It’s a way to build more illuminating models which reveal crux issues: the uncertainties which hold outsized influence on the topic at large. The authors of the paper selected long-term AI risk as an example issue for the study, but it’s easy to see how the approach could generalise. It involves first setting out an overall hypothesis, and then asking experts to identify what events or trends would significantly change the outcome. This is used to calculate a ‘value of information’ score, thus revealing the most informative questions - as well as how they connect to each other.
Doing this manually is pretty resource-intensive. The authors point to LLMs as one potential solution - and for sure, there will be a role for agents here. But in the light of relevance realisation - needing skin in the game, building the model yourself, and so on - I think some level of human effort here is the price of impact. The forecasts produced by this approach are not significantly different in format than those produced by prediction markets. But questioning, mapping, and constructing a model, by contrast, is relevance realisation made manifest. The process of building the model is more valuable than either the model or the forecast themselves.
A new goal for the community should be finding ways to scale the model-building process to make it intuitive and incisive for people who make decisions. The current approach (consulting) is too bottlenecked and expensive; but retaining some level of effort and social involvement will be critical for this work to be meaningful. This is an interesting design challenge: we’re shifting upstream to include the generative, hypothesis-creation space as well as the evaluative work of forecasting.
This will not be either/or, but both/and. Remember, cognition works by cycling between the generative work of deciding what’s relevant, and the evaluative work of making predictions in order to act. Each moves the other, and each is necessary for the other. This is a good opportunity for the two schools of foresight and forecasting to integrate their relative strengths. The world of foresight has typically been good at generative work, widening the hypothesis space; while the forecasting community, as we have seen, is measurably good at evaluative work. Bringing the two together - into a tool, platform, or process - might genuinely (and literally) close the loop on this weird problem of irrelevance. There are signs of life here. Early experiments point to the emergence of parts of such a system. This is encouraging: the answer might well be right in front of us.
In practice: the living thesis
My modest proposal for what this may look like is a ‘living thesis,’ a way to continuously create relevance, find leverage, and test the viability of potential actions. As we’ve seen, trying to replace parts of this process from the outside simply doesn’t touch the sides. The generative-evaluative cycle is already running in the minds of the people you want to influence. This default state will always win out because it forms a fully functioning, complete loop. No one would claim this is optimal, but it is locally optimal: a complete, bad process will always beat an incomplete good one. The natural experiment of trying to make forecasting impactful shows that the mind just routes around better information if it’s not part of this already-running loop. You can either upgrade the entire cycle, or none of it.
Any tool must externalise the Whole Thing, in one place. to work across the cycle in its entirety. It should help you:
These are, I think, the bare bones of what’s structurally necessary for the practice of forecasting to be useful and relevant. It’s a better, more illuminating way to construct a thesis about the context you’re in and the actions you may wish to take.
What makes this a living thesis is that the context is always changing, the ‘frame’ of relevance realisation is always degrading, and so, the cycle must always keep running. Step three is not the end of a linear process, it is just an element in a continuous exercise of monitoring the situation, placing irons in the fire, and gathering enough confidence to commit. Taking action changes the dynamics of the system from within; and so, the frame degrades, the model updates, by which point new actions appear possible, and on, and on, and on.
In the Before Times, maintaining and updating a live world model was a highly resource-intensive activity. In practice, you’d pay a team of consultants to do it every few months (if you even did it at all), which is both expensive and ineffective. Massively reducing this cost - and making model updates daily - is non-negotiable for this approach to scale. I think this involves hitting the ‘fix everything easily’ button and using agents to do some, but not all, of the work. Specifically, agents should scan for evidence that the model is becoming unfit for purpose: perhaps an overnight decision by a regulator is perturbing the dynamic of the system, or an uncertain issue has swung one way rather than the other. But the decision on whether to incorporate that evidence and update the model must rest with the human. This is a form of meaningful friction that ensures relevance is maintained. We should design for skin in the game.
If it sounds never-ending, that’s because it is. The edges of the system are always exposed to new things. You can’t pause the world while you decide what to do, and yet the tooling we have today assumes that this can ever be the case. Instead, our tools should accept both the ever-changingness of the world and the constant cycling of the cognitive process we’ve evolved to avoid getting stuck in place.
Parts of this are already emerging
Doing any amount of this work, at scale, is going to require highly-reliable predictions and reasoning on tap. Probably the best-known (and perhaps best-funded?) AI forecaster is Mantic, a London-based startup which has achieved impressive results in various tournaments on Metaculus. As far as I can tell, they’re in the midst of trying to find commercial applications beyond winning forecasting competitions - although that seems to be going okay for the Polymarket bots which make up 5% of wallets but over 75% of trading volume. I’m reminded here of Scott Alexander’s quip, that perhaps “prediction markets’ role in God’s plan was only to provide the foundation for AI superforecasters - the training data, the benchmarking arena, and the pot of money that rewards innovation.” Mantic and others are solving the supply-side problem of turning really good forecasts into a commodity - it’s just a weird quirk that the demand-side doesn’t really need more of the same. Instead, I think AI forecasters may just become APIs powering living theses, or something similar.
In February this year, Metaculus announced Radiant, “a networked forecasting tool for strategic decision-makers.” The tool is like a whiteboard where you can link forecasts together to visualise how their relationships affect a domain: in essence, building a living map of a system. This is supported by AI-assisted question generation[1], which makes it easier to formulate questions and identify connections you might otherwise have missed. Radiant, though still in prototype form, gets closer to what I think is necessary for forecasting to touch relevance realisation. It’s also good to see Metaculus move a level of abstraction upwards beyond the traditional format of a prediction market. What’s missing is still skin in the game: you need to be able to form your own plan and throw it into the model to see how it performs under uncertainty. Otherwise, Radiant will remain a systems-mapping tool, rather than becoming a strategy and leverage tool.
Perhaps the closest thing to what I’ve articulated is Deep Future, another early prototype which bills itself as “Deep Research for scenario planning.” Uncannily, its creator Gordon Brander seems to have landed on a similar diagnosis of what needs to change to make forecasting useful: “AI agents map out your strategic landscape, identify the forces driving change, explore multiple trajectories, and discover leverage points you can use to shape the space of possibility.” So, this looks like foresight-based scenario analysis on steroids, which is cool if you like that sort of thing. I’m particularly interested in the final step, in which “AI synthesizes insights from scenario testing to recommend adaptive strategies with built-in flexibility, creating decision frameworks that perform regardless of which future materializes.” Now, this all sounds great, but we… just don’t know yet, because Deep Future is currently waitlisted. It would be good to try it out and see if I could have saved myself (and you) several thousand words.
What may be more difficult than building the right tools, or interfaces, or even just backend integrations into Claude, is letting go of Tetlock-style forecasting and prediction platforms as ends in themselves. Perhaps you have come to believe that forecasting is way overrated and we should stop funding it, or maybe you take the firm view that forecasting is not overrated and it’s probably funded appropriately. Both positions assume that forecasting as currently practiced is in its terminal form. I think this would be a mistake. It will be hard to admit, but forecasting needs to become a boring utility like water or intelligence; something which powers stuff other people actually want to do. Yes, perhaps the role of prediction markets in God’s plan was to provide the training data for AI forecasters, but so too the role of AI forecasters will be to enable useful, beautiful, living theses: Bloomberg terminals for the rest of us.
Epilogue
While writing this piece I often wondered: why don’t Anthropic or OpenAI just eat this? It doesn’t seem beyond the capability of the engineers and designers at either organisation to build a system which helps you make radically better decisions under uncertainty. Surely, at least for enterprise customers, such a capability - if well-executed - would be enormously useful? Perhaps they’re both so focused on leapfrogging each other’s coding products that ‘IDEs for strategy’ are just further down the priority list. But I think that sharp, clear-eyed systems mapping and decision-making will only become more important as AI capabilities improve; not least because the speed and scale at which things may fuck up will increase by ~orders of magnitude over the next two years.
There is another alternative, though. Perhaps the final form for all technology is simply a chat window. I’m sceptical of this view - but what if it may never be necessary to build specialised interfaces for decision-making: living theses, scenario tools, IDEs, whatever - all by the wayside. What if, instead, forecasting finally finds its impact by diffusing into billions of conversations with generalised AI, invisibly, silently improving their accuracy without ever stating a probability? Putting fluoride in the water supply worked pretty well. You may never ask for a forecast again - and prediction markets may well devolve into casinos - but the essence of forecasting will still be there, buried in the system prompt for Opus 5 or GPT 6.
Earlier, I suggested that “if you could bring someone’s orientation more in line with reality, then everything downstream might fall more neatly into place.” Maybe that’s possible, and all it takes is a few edits to a markdown file, hidden deep underground, somewhere beneath the Bay Area. In the future, AI aligns you.
It at least looks like ‘AI for question decomposition’ is solvable. It will need to be, to lower the cost and experience barriers enough for non-experts to start directing their own forecasting and scenarios. Future Tokenizer, a prototype from Jordan Rubin, is a very interesting stab at a solution. He built the tool as an “IDE for human-AI collaborative thinking” which gets around the weirdly limiting problem of most AI interactions occurring in linear conversations. You start by articulating a question or a goal, and Future Tokenizer helps you to map out load-bearing assumptions, blind spots, and hidden cruxes, in (you guessed it) a system of nodes and branches. You can then take things further: red-teaming your map for alternative hypotheses before ultimately consolidating down into a robust plan of action. I think this level of decomposition and testing is necessary for any sort of scenario or systems tool to be useful; otherwise you risk hypothesis lock-in - forecasting well only on the issues you assume are important.