Fundamental question

Given all this, what happens next?

This series has spent ten investigations assembling a model — of perception, consciousness, self, suffering, behaviour, identity, systems, events and cosmos. The payoff question is whether a better model of reality lets you see forward. The honest answer is: partly, for a measurable distance, and not in the way most people assume. The most interesting finding is that the limit is not a limit of intelligence.

Overall confidence in the core claim moderate (3/5) — the forecasting-accuracy and neuroscience findings are robust and replicated; the predictability-limit mathematics is strong but its interpretation is actively disputed, and the AI-parity results are two months old and rest on an extrapolated human baseline

Output 1 · Research synthesis

What the evidence currently says

"What happens next?" is four questions wearing one coat. How far ahead can anyone see? — an empirical question with an answer. How far ahead can anything be seen, even in principle? — a mathematical question with a partial answer. What is the machinery doing when it looks forward? — a neuroscientific question with a surprising answer. And who gets to decide which future counts? — a political question that turns out to move outcomes more than accuracy does.

superforecastingpredictability limits episodic future thinkingcomputational irreducibility ForecastBenchFuture Design

Strong / convergent

Forecasting skill is real, and it has a horizon. Philip Tetlock's Good Judgment Project — the largest forecasting tournament ever run — established that a small group of identifiable "superforecasters" beat professional intelligence analysts with access to classified material by roughly 30%. The same body of work established the ceiling: accuracy is strong over the first year, then declines, and by a horizon of three to five years it is difficult to distinguish elite forecasting from noise. Skill is genuine but bounded in time, and the bound arrives sooner than institutional planning cycles assume.
The horizon is a property of the system, not only of the forecaster. A May 2025 PNAS paper by Dong, Faranda, Gualandi, Lucarini and Mengaldo introduced Time-Lagged Recurrence (TLR), a method that estimates, from observational data alone and without knowing the governing equations, how far into the future a given system state can be predicted. Its central result is that predictability is state-dependent: the same system has windows of high forecastability and windows where the horizon collapses. Validated on idealised systems and on large-scale atmospheric fields. This reframes the question from "are we good enough?" to "is the system, right now, in a predictable state?"
Prediction runs on the memory hardware. Episodic future thinking — imagining specific events one may personally experience — recruits the same core brain network as episodic recollection, and typically engages those regions more strongly than remembering does. The constructive episodic simulation hypothesis explains why: imagining a future event requires recombining stored details into a novel scenario, which is computationally more expensive than retrieving one. Two interacting subsystems carry it: a medial-temporal/hippocampal scene-construction system, and a dorsomedial-prefrontal/temporoparietal self-referential and social system, with ventromedial prefrontal cortex integrating valuation. You do not have a prediction organ. You have a memory organ running in reverse.
Machines have arrived at the boundary of elite human forecasting — this year. ForecastBench, run by the Forecasting Research Institute, is the only major benchmark that pits AI forecasters directly against human superforecasters. As of 16 July 2026, several systems are statistically indistinguishable from superforecaster-level accuracy on the tournament leaderboard, and one system (Cassi AI) ranked above the superforecaster median on market questions — the harder category, drawn from Manifold, Metaculus, Polymarket and the RAND Forecasting Initiative, requiring judgement about novel one-off events rather than base-rate lookup. AI matched superforecasters on the easier time-series "dataset questions" back in May 2026. The Institute's own caveats matter and are given below.

A fifth strand, weaker but persistent. Imagining a future in more concrete, personally continuous terms appears to change present behaviour. A 2025 systematic review (Grekin, Thomas, Souweidane and Stidham) screened 1,256 articles and found 14 papers containing 23 studies of future-self-continuity interventions, reporting effects on behavioural outcomes ranging from small to large. That range is the finding: the mechanism is real and malleable, the effect size is not yet stable, and the samples are overwhelmingly undergraduates and online panels.

Underseen · Non-English research

Japan stopped trying to predict the future and started installing it

The most consequential work on this question in the last decade is largely absent from Anglophone forecasting literature, because it is not forecasting at all.

1 · Future Design (未来デザイン) and "imaginary future generations"

Since 2012, a group led by Tatsuyoshi Saijo (Kochi University of Technology) and Keishiro Hara (Osaka University) has developed Future Design, built on a deliberately simple observation: future generations bear the consequences of present decisions and have no seat at the table. The intervention is not better prediction. It is to designate living people to occupy that seat — the imaginary future generation (仮想将来世代) — and have them deliberate as residents of 2060. The capacity this is meant to unlock is named futurability (将来可能性): the disposition to feel satisfaction from forgoing present benefit for the benefit of a generation one will never meet.

2 · The Yahaba result

In 2015, in the town of Yahaba, Iwate Prefecture, randomly selected residents were split into present-generation groups and imaginary-future-generation groups and asked to deliberate on town administration to 2060 — including the municipal water utility, then running a surplus. The present-generation groups argued to return the surplus by cutting water charges. The imaginary-future-generation groups argued to raise them, to accumulate funds for infrastructure renewal they themselves would not benefit from. Yahaba subsequently raised the rates. Laboratory versions report that more than twice as many groups made investments favouring future generations when an imaginary future person was present. Similar deliberations have run in Suita, Matsumoto and Kyoto; Kyoto's applied the method to decarbonisation planning for 2050.

Researchers associated with the programme are candid about its status: the experimental protocols have not been standardised across sites, and the quantitative evidence that role assignment changes individual cognition is not yet sufficient. It is a promising mechanism, not a settled one.

3 · Why the German tradition names it differently

German Zukunftsforschung — institutionalised at Fraunhofer ISI and IAO, the IZT, and FU Berlin — has for decades organised itself around Szenariotechnik (scenario technique), the Delphi method and Backcasting: planning backwards from a specified future state to the present. The grammar is instructive. Anglophone practice says forecast — singular, extrapolated, and something the world does to you. The German tradition says Zukünfte — futures, plural, consistent images to be constructed and chosen between. The 2025 direction of travel there is toward AI-assisted trend analysis married to immersive scenario work: making a possible future something a decision-maker stands inside rather than reads.

This is precisely the kind of finding an English-only review misses. Both traditions have independently converged on the same move — from predicting the future to occupying it — and neither converged on it by improving accuracy.

Output 2 · Insight generation

What follows if this is true

The forecasting question is badly posed, and its own evidence says so. Event-level accuracy decays toward noise within a few years, but the constraints that generate events decay far more slowly. The productive question is therefore not "what will happen?" but "what is this system becoming structurally unable not to produce?" — a question whose answer is stable over exactly the horizons where event forecasting fails.
Your forecast horizon is bounded by your memory repertoire. If prospection is recombination of stored episodes, then a person or institution can only imagine futures assembled from parts it has already encountered. This makes model-building a forecasting activity in the literal sense: acquiring more shapes — more structural patterns, from more domains — extends the space of futures that can be simulated at all. It also predicts the characteristic failure: novel events are not mispredicted, they are unimaginable, because no stored parts assemble into them.
If machines reach parity on the number, human advantage moves upstream. ForecastBench measures accuracy on questions someone else wrote. Nothing in the benchmark measures deciding which question is worth asking, or noticing that the resolution criterion smuggles in a frame. Comparative advantage moves from producing probabilities to specifying what should be forecast — and from event-level output to structural framing.
The strongest lever on the future found in this research is not accuracy. It is standing. Yahaba changed the water rate not by predicting 2060 more precisely but by putting 2060 in the room. If that generalises, the binding constraint on long-horizon decisions is representational, not epistemic — a claim with uncomfortable implications for anyone selling forecasts, this project included.

These are interpretive implications drawn from the evidence above, not established findings — flagged as such.

Show the work · Contradictions & competing theories

Where it's contested

Do we systematically underestimate how much we will change? The "end of history illusion" (Quoidbach, Gilbert & Wilson, Science, 2013; 19,000+ participants) reported that people at every age recall having changed a great deal and predict changing very little. It has been widely repeated — and challenged since publication. Christian Jarrett's critique for the BPS notes that the design contains no data on actual change: it compares remembered change against predicted change, so memory distortion alone could produce it. Worse, if a person expects change but cannot predict its direction — more extraverted, or less? — then reporting "no change" is the mathematically correct answer, not a bias. The authors' own longitudinal check used a different personality instrument and permitted no direct comparison. Unresolved. Treat the effect as suggestive, not established.

Does computational irreducibility actually entail unpredictability? Wolfram's principle — that for a sufficiently rich computational process there is no shortcut, and the only way to know the outcome is to run it — is the strongest available argument that some futures are closed to prediction in principle. But a Philosophy of Science paper, "Setting the Demons Loose," argues directly that computational irreducibility does not guarantee either unpredictability or emergence: irreducibility concerns the cost of exact simulation, while useful prediction is often statistical or coarse-grained and may remain available. This is a live and load-bearing dispute — it decides whether the forecasting horizon is a hard wall or an economic one.

Has AI really reached superforecaster parity? The Forecasting Research Institute publishes its own caveats, and they are substantial: the superforecaster baseline was last elicited in 2024 and is now a statistical extrapolation that degrades with time; the 95% confidence intervals overlap heavily, so the data are "more consistent with superforecaster parity than with outperformance"; and results are stochastic in the resolved question set. A fresh superforecaster round is scheduled for autumn 2026. Separately, work in Philosophical Transactions of the Royal Society B (2026) on crowd versus LLM forecasting identifies an accuracy–correlation effect — LLM forecasts are more correlated with one another than human crowds are, which means aggregating many models buys less error cancellation than aggregating many people. Parity on a leaderboard is not equivalence in an ensemble.

Does imagining the future reliably change what people do? The future-self- continuity literature reports effects from small to large across 23 studies — a spread wide enough that the honest summary is "the mechanism exists, the magnitude is unknown." Most samples are undergraduate. Future Design's own researchers say the same of the collective version: real effects, non-standardised protocols, insufficient data to quantify. Anyone quoting a specific number here is over-reading.

Through the RFT lens

The same shape, in the project's own terms

In Recursive Field Theory (RFT) terms, the distinction the forecasting literature keeps rediscovering is the distinction between an event and a field. Events are outputs; fields are what generate them. A forecast at event level decays fast because events are the noisiest layer of the system; a forecast at field level — which constraints are activating, coupling, closing — holds for longer, because constraint changes more slowly than incident. The TLR result that predictability is state-dependent reads, in this frame, as the claim that a system's forecast horizon is itself a variable: it expands when fields are stable and collapses when several are reorganising at once. And the correct output of a long-horizon forecast is not a point but a distribution over attractors — which structural order the system is becoming unable not to produce — with event-level probabilities stated only conditional on that trajectory, and always with less confidence.

The neuroscience lands on the same shape from the other side. If prospection is recombination of stored episodes, then memory is not a record of the past; it is the generator of available futures. RFT's distinction between history — everything that happened — and memory — only what still exerts force — is the same distinction the hippocampal literature draws when it says the brain does not retrieve the past, it rebuilds from active fragments. Which suggests the sharpest version of this week's question is not "what happens next?" but "which parts of the past are still exerting enough force to shape what can happen next?"

This is David's own interpretive framework, offered as a way of seeing — not established cognitive science, not a validated forecasting method, and not a claim that the analogy has been tested. It describes a structural correspondence, nothing stronger.

Output 3 · Recursive investigation

What to investigate next

Can a TLR-style predictability index — "how forecastable is this system right now" — be computed for social and psychological systems, not just physical ones? If so, honest forecasting would report its own horizon alongside every estimate.

Does the "prediction is recombined memory" mechanism imply that deliberately acquiring structurally different models measurably widens the range of futures a person can imagine — and can that be tested rather than asserted?

If AI systems reach and pass superforecaster accuracy on event questions, what exactly is left that humans do better — and is question-generation genuinely a separate skill, or does it fall next?

Does Future Design's effect survive contact with high-conflict, low-trust settings, or does the imaginary-future-generation role only work where a shared civic frame already exists?

Is the end-of-history illusion real? A prospective design — predict, then measure actual change with the same instrument a decade later — would settle a thirteen-year-old dispute. Does one exist?

Sources

Dong, C., Faranda, D., Gualandi, A., Lucarini, V. & Mengaldo, G. — "Time-lagged recurrence: A data-driven method to estimate the predictability of dynamical systems." PNAS, 16 May 2025. pnas.org · PubMed 40377987 · code: github.com/MathEXLab/TLR
"Data-driven method reveals how (un)predictable complex systems can be." Phys.org, June 2025. phys.org
"Predictability of Complex Systems." Physics Reports / arXiv:2510.16312. arxiv.org
Tetlock, P. & Gardner, D. — Superforecasting: The Art and Science of Prediction (2015); Good Judgment Project results. Overview: "The Limits of Applied Superforecasting," Commoncog
Forecasting Research Institute — "AI models have likely reached parity with superforecasters on ForecastBench," 16 July 2026. forecastingresearch.substack.com
Forecasting Research Institute — "LLMs Are Closing the Gap on Human Superforecasters" (January 2026). forecastingresearch.substack.com
Karger, E. et al. — "ForecastBench: A Dynamic Benchmark of AI Forecasting Capabilities." arXiv:2409.19839. arxiv.org · leaderboards: forecastbench.org
"Crowdsourced versus large language models forecasting: evidence for the accuracy–correlation effect." Philosophical Transactions of the Royal Society B, 381(1948), 2026. royalsocietypublishing.org
"Episodic Future Thinking." Open Encyclopedia of Cognitive Science, MIT. oecs.mit.edu
Schacter, D. & Addis, D. — constructive episodic simulation hypothesis; Buckner, R. & Carroll, D. — "Self-projection and the brain," Trends in Cognitive Sciences (2007). sciencedirect.com
Schacter, D. et al. — "Episodic future thinking: mechanisms and functions." Current Opinion in Behavioral Sciences. sciencedirect.com
Grekin, E. R., Thomas, H. A., Souweidane, M. A. & Stidham, J. L. — "The Effects of Future Self-Continuity Interventions on Behavioral Outcomes in Adults: A Systematic Review of the Literature" (2025). journals.sagepub.com
Hershfield, H. — future self-continuity research index. halhershfield.com
Quoidbach, J., Gilbert, D. T. & Wilson, T. D. — "The End of History Illusion." Science, 339(6115), 96–98 (2013). science.org · PDF: harvard.edu
Jarrett, C. — "'The end of history illusion' illusion." BPS Research Digest. bps.org.uk
Kobayashi, K. — "Future Design: A new policymaking system for future generations." RIETI (Japanese original and English translation). rieti.go.jp
Saijo, T. (ed.) — Future Design: Incorporating Preferences of Future Generations for Sustainability. Springer (2020). link.springer.com
Hara, K. et al. — "Policy design by 'imaginary future generations' with systems thinking: a practice by Kyoto city towards decarbonization in 2050." Futures (2023). sciencedirect.com
United Nations University — "'Future Design': an innovative approach to decision-making." unu.edu
IZT — Methoden der Zukunfts- und Szenarioanalyse: Überblick, Bewertung und Auswahlkriterien (German). izt.de
iit Berlin — 7 Foresight-Methoden zur erfolgreichen Strategieentwicklung (German, 2024). iit-berlin.de
Fraunhofer IAO — "AI und XR im Foresight" (German, 2025). iao.fraunhofer.de
Wolfram, S. — computational irreducibility. A New Kind of Science, p.739. wolframscience.com
"Setting the Demons Loose: Computational Irreducibility Does Not Guarantee Unpredictability or Emergence." Philosophy of Science, Cambridge University Press. cambridge.org
Szpunar, P. M. & Szpunar, K. K. — "Collective future thought: Concept, function, and implications for collective memory studies." Memory Studies (2016). journals.sagepub.com
"Collective future thinking in cultural dynamics." European Review of Social Psychology (2025). tandfonline.com

Primary sources linked where available; books and encyclopedia entries cited by title. Always consult the originals — this synthesis describes emphasis and findings, not verbatim claims.

CPD · Structured learning

Make this count as CPD (≈30–45 min)

Learning outcomes. After completing this unit you will be able to: (1) state the empirical forecasting horizon established by tournament research and explain why accuracy decays with time; (2) distinguish intrinsic predictability — a property of the system's current state — from realised predictability, and explain why a forecast without a stated horizon is incomplete; (3) describe the shared neural basis of episodic memory and episodic future thinking, and explain what the constructive episodic simulation hypothesis predicts about the futures a person can and cannot imagine; (4) evaluate at least one contested claim in this piece — the end-of-history illusion or AI forecasting parity — against its published critique.

Format & time. Reflective structured learning · ~30–45 minutes.

To complete the unit:

Read the associated references. At minimum, the MIT OECS entry on episodic future thinking and the Forecasting Research Institute post including its caveats section — both linked in Sources above.
Reflect (write 3–5 lines each): When I form an expectation about a client, a colleague or my own practice twelve months out, which stored past cases am I recombining — and what future is therefore structurally unavailable to me? · Where in my work do I state a prediction without stating its horizon or the conditions under which it stops holding? · If I put the "imaginary future generation" of my own practice or organisation in the room — the people affected in twenty years — which current decision would they most object to?
Log it. Record time spent and these reflections against your professional body's CPD requirements.

CPD-eligible structured learning; not statutory-regulator endorsement — practitioners self-assess relevance and log accordingly.

How this was made. One fundamental question per week, taken in order from a published queue. Every claim is traced to a primary or near-primary source, checked to exist before it is cited, and separated into strong/convergent evidence and contested evidence. Non-English literature is searched deliberately, because an English-only review reliably misses things — this week, most of the interesting material. Competing theories and direct critiques are shown rather than smoothed over, confidence is stated as a number with a reason, interpretation is labelled as interpretation, and the RFT lens is flagged throughout as David's own framework rather than established science. A human reviews every word before publication. Reality is the arbiter.

Note on the series. This closes the founding arc of eleven questions. From here the engine returns to the oldest investigations and deepens them: what has changed, what confidence has moved, what broke.