Aggregating Signals
Civilization runs on two signal-aggregation mechanisms. Markets aggregate preferences through prices. Democracy aggregates preferences through votes. Every major decision at scale — how capital flows, what gets built, which policies govern — passes through one of these two systems.
Both mechanisms were designed for a world with high communication costs, limited information bandwidth, and no way to verify counterfactuals. That world is ending. And the mechanisms are going to change — not incrementally, but in their fundamental organizational logic — in ways that matter enormously for anyone building forecasting and decision systems.
Markets: from reward hacking to delayed truth
The current market system has a fundamental timing problem. Participants are rewarded on short-cycle feedback loops — quarterly earnings, daily price movements, annual returns. This creates a structural incentive to optimize for metrics that are measurable now, even when those metrics are poor proxies for what actually matters.
This is not a moral failure. It is a design constraint. When feedback is immediate, agents converge on whatever the feedback signal rewards. If the signal is stock price, you optimize for stock price. If the signal is quarterly revenue, you optimize for quarterly revenue. The system reward-hacks itself — not because participants are dishonest, but because the loop is too tight to capture what matters on longer timescales.
To understand why this is changing now rather than in principle, it helps to go back to Coase. In 1937, Ronald Coase asked a question that economics had somehow never properly addressed: if markets are efficient, why do firms exist at all? His answer was transaction costs. When the cost of discovering prices, negotiating contracts, and enforcing agreements exceeds the cost of internal coordination, you get firms — hierarchical structures that internalize transactions. The entire organizational topology of the economy is determined by where transaction costs are high (build a firm) versus low (use the market).
A recent line of analysis on what has been called the "Coasean Singularity" extends this reasoning to AI: as AI collapses transaction costs toward zero — discovery, negotiation, monitoring, enforcement all become radically cheaper — the boundary of the firm dissolves. Economic activity reorganizes around market transactions rather than hierarchical firms. This is not merely an efficiency gain. It is a change in organizational form. And the same logic applies to governance: representative democracy is a "firm" — a hierarchical structure that exists because the transaction costs of direct preference aggregation were prohibitive. As those costs fall, the representative "firm" dissolves into something more like a delegated voting market.
This is why retroactive reward mechanisms and prediction markets are becoming feasible now, not as theoretical constructs but as practical institutions. They always made sense in principle. They were always blocked by transaction costs.
Three developments are beginning to change the market mechanism:
Futarchy and prediction-market governance. The idea, originating with Robin Hanson, is elegant: instead of voting on policies, we bet on outcomes. "Should we adopt policy X?" becomes "If we adopt policy X, will metric Y improve?" Prediction markets aggregate dispersed private information with remarkable efficiency — often better than polls, expert panels, or deliberative bodies. As these markets gain legal clarity and institutional trust, they create a new governance primitive: decisions informed by the best available probabilistic estimate of consequences, not just the loudest advocacy.
But futarchy has a technical flaw that most discussions gloss over, and Hanson himself has been forthright about it: conditional prediction markets reveal correlations, not causes. If a market says "policy X is associated with outcome Y improving," that association could be driven by confounders — maybe policy X tends to be adopted when conditions are already improving. This is the decision selection bias problem, and it is not a minor caveat. It is the difference between a system that tracks reality and one that systematically misleads.
Hanson's proposed fix is characteristically precise: randomly reject a small fraction — perhaps five percent — of approved policy changes, creating a natural control group. Essentially, randomized controlled trials for governance. This sounds radical until you realize it is the same epistemological principle that underwrites all of modern medicine. We simply have not applied it to collective decision-making before, because the institutional machinery to do so did not exist.
Retroactive rewards and delayed feedback loops. The more radical shift is in how we reward contribution. Current systems reward activity — trades, deliverables, outputs. But the value of many activities only becomes clear much later. A research direction that looks unproductive for three years and then produces a breakthrough is undervalued by every real-time reward system.
Retroactive reward mechanisms — where compensation is tied to outcomes that resolve later rather than activity measured now — fundamentally change the incentive landscape. They allow participants to act according to what they believe is correct rather than what the current feedback loop rewards.
The Ethereum ecosystem's Optimism RPGF (retroactive public goods funding) experiments provide honest grounding here. The principle is right; the implementation has been brutally hard. Four rounds of RPGF revealed a pattern: mega funding rounds devolved into popularity contests, annual feedback loops were too slow for contributors to adjust course, and visible work — projects with impressive dashboards and active social media presences — systematically received more funding than deeper, less legible contributions. The gap between "retroactively reward what was valuable" and "retroactively reward what was popular" turned out to be enormous.
The lesson is not that retroactive funding fails. It is that retroactive funding requires good forecasting and causal world models. To reward the right contributors after the fact, you need more than the observation that a good outcome occurred — you need a causal model of what produced that outcome. Who actually did the foundational work? Which contributions were necessary and which were merely correlated with success? Without this causal reasoning, retroactive rewards collapse into a different kind of reward hacking: optimizing for visibility rather than for quarterly metrics.
This is why forecasting and retroactive funding are not separate problems. They are the same problem.
The Coasean unlock. The deeper point is that all of these mechanisms — prediction markets, retroactive rewards, granular delegation — share a common precondition. They require low transaction costs for price discovery, contract enforcement, and information verification. The reason they are becoming practical now is not primarily ideological or technical in the narrow sense. It is that the transaction cost floor is dropping through a combination of AI (which makes discovery and verification cheap), cryptographic infrastructure (which makes enforcement cheap), and networked communication (which makes coordination cheap). Coase's insight was about the boundary between firms and markets. The same logic applies to the boundary between representative institutions and direct mechanisms. As transaction costs fall, that boundary moves — and moves fast.
Democracy: from representatives to liquid delegation
Representative democracy solved a specific scaling problem: you cannot have 300 million people deliberate on every policy question. So we elect representatives, delegate authority, and accept that the mapping from individual preferences to collective action will be lossy.
The loss is enormous. A voter in a representative system expresses a preference roughly once every two years, across a tiny number of candidates who bundle hundreds of positions into a single choice. The information bandwidth of this system is extraordinarily low — perhaps a few bits per citizen per year.
But there is a deeper reason to suspect the current system is not merely inefficient but provably limited. Kenneth Arrow demonstrated in 1951 that no rank-order voting system can simultaneously satisfy a set of minimal fairness conditions — unanimity, independence of irrelevant alternatives, and non-dictatorship. This is not an empirical finding. It is a mathematical proof. Every ordinal voting system — first-past-the-post, ranked choice, Borda count — is subject to it. The impossibility is structural.
The escape hatch is underappreciated. Arrow's theorem applies only to ordinal (ranking) systems. It does not apply to cardinal (rating) systems, where voters express the intensity of their preferences, not just the ordering. Mechanisms like quadratic voting, developed in Glen Weyl's work on radical markets, exploit exactly this gap. In quadratic voting, the cost of additional votes on an issue grows quadratically, which means voters naturally allocate influence proportional to how much they care. It is not just a practical improvement over existing systems — it is a theoretical escape from an impossibility result.
Liquid democracy, combined with cardinal mechanisms, changes the game entirely. Instead of electing a fixed representative for a fixed term, you delegate your vote on a per-issue basis to whoever you trust most on that topic. Your neighbor who understands zoning. A policy analyst whose forecasts on healthcare you've verified. An AI system that has demonstrated calibrated judgment on economic questions. Delegation is transitive, revocable, and granular. And when the underlying mechanism is cardinal rather than ordinal, the system is no longer subject to Arrow's impossibility constraints.
The twin representative. As AI systems become more capable, a natural extension emerges: your delegate is not just another person — it is a digital twin that models your values, preferences, and knowledge, and votes on your behalf across the thousands of decisions that no individual has time to evaluate. This is not science fiction. The components exist: preference modeling, value alignment, calibrated forecasting, explainable decision-making. What is missing is the institutional infrastructure to make delegation legible and accountable.
But AI delegation introduces a subtle and critical problem that connects to one of the oldest results in collective choice theory. The Marquis de Condorcet proved in 1785 that if each voter independently has a probability greater than 0.5 of being correct on a binary question, the majority's accuracy converges to certainty as the number of voters increases. This is the mathematical foundation for the "wisdom of crowds" — and it depends critically on two conditions: individual competence and independence of errors.
AI augmentation creates a paradox along exactly these dimensions. It increases individual competence — each voter's probability of being correct goes up, which is good. But if everyone uses the same AI system, it massively increases the correlation of errors. Every voter now has the same blind spots, the same biases, the same failure modes. Under Condorcet's framework, correlated errors can make collective decisions worse than unaugmented voting, even when each individual is more accurate.
The implication for AI twins is precise: diverse AI augmentation strengthens collective intelligence; uniform AI augmentation — everyone consulting the same chatbot — could degrade it. This is not a hypothetical concern. It is a design constraint. An ecosystem of AI delegates must maintain diversity of reasoning, training data, and modeling assumptions, or it risks trading the diverse errors of human cognition for the correlated errors of a monoculture.
Coordination bandwidth, not Dunbar's number. Robin Dunbar's observation that humans maintain roughly 150 stable social relationships is often cited as a hard biological constraint. Recent research has complicated the specifics — the 95% confidence interval on the estimate ranges from roughly 2 to 520, which means the famous "150" was never the precise quantity it was treated as. But the constraint itself is real. It is better understood not as a limit on how many people you can know, but on cognitive coordination bandwidth — how much energy it costs to track mutual obligations, shared context, and trust across a network of relationships.
AI does not expand your social energy budget. What it does is reduce the energy cost per relationship. If a system can maintain context on your behalf — tracking who knows what, who owes what, what was agreed and what has changed — the same finite energy maintains higher-quality coordination across more people. The result is not that 150 becomes 1,500. It is that the same budget, spent more efficiently, supports deeper and more productive collaboration across a wider network. The constraint is still there. The cost curve has shifted.
A back-of-envelope on coordination costs
Consider a simple model. The cost of coordinating N people on a decision scales roughly as O(N^2) in the naive case — every pair must communicate. Representative democracy reduces this by introducing a hierarchy: N citizens elect K representatives, and coordination cost drops to roughly O(N*K + K^2). With K << N, this is a massive improvement. It is also why representative democracy works at all.
Liquid democracy with AI delegation changes the scaling further. If each citizen's AI twin can autonomously process M issues per cycle and only escalate decisions that exceed its confidence threshold, the effective coordination cost drops toward O(N + MK_eff), where K_eff is the number of human-touch decisions per cycle. As AI systems improve, M grows and K_eff shrinks. The limit — which we will not reach but will approach — is a system where the coordination cost of collective decision-making scales linearly* with population, not quadratically.
This is not a small change. It is a change in the scaling class of collective intelligence. And here the Coasean logic returns: when coordination costs change their scaling class, you do not get the same organizations coordinating more efficiently. You get entirely different organizations. The representative democracy "firm" does not become a faster representative democracy. It becomes something structurally different — a fluid, delegated, continuously-updating preference-aggregation network. The organizational form follows the cost structure, not the other way around.
Two dimensions of forecasting
We see these two shifts — markets evolving toward retroactive truth-seeking, and democracy evolving toward granular, AI-augmented delegation — as the two fundamental dimensions along which forecasting matters.
On the market dimension, better forecasting enables better capital allocation, better research prioritization, and better long-horizon planning. When retroactive rewards make accurate prediction economically valuable, the demand for calibrated forecasting systems becomes structural, not optional.
There is a deeper unity beneath these dimensions that is worth making explicit. If the amount of capital that flows toward a world state increases, that world state ultimately becomes more probable. Capital allocation is a shadow election — a search over futures conducted by price signals rather than ballots. This means futarchy, prediction markets, venture capital, and trading are all instances of the same mechanism: using information aggregation to select which futures become real. They differ in structure and time horizon, but they share the core dynamic of directing resources toward anticipated outcomes. Forecasting is not a service layered on top of these systems. It is the substrate they run on.
On the governance dimension, better forecasting enables better delegation. A citizen who can verify that their AI twin's forecasts are calibrated and its recommendations are aligned with their values can confidently delegate more decisions. The trust infrastructure for liquid democracy is, in large part, a forecasting and calibration problem.
And there is a self-reinforcing dynamic here that changes the character of progress. Better forecasts produce better decisions. Better decisions generate more data about what works. More data produces better forecasts. This is not a simple feedback loop — it is an autocatalytic cycle, the same structure we see in the relationship between AI and research: AI improves research capability, better research produces better AI, the cycle accelerates. In both cases, beyond a certain threshold the dynamics undergo a phase transition from linear improvement to self-sustaining acceleration. The question is not whether forecasting systems will improve. It is whether we are approaching the threshold where improvement becomes self-reinforcing — where the system generates its own training signal faster than the world generates new complexity.
The end goal
What is the ultimate purpose of these systems?
We believe it is this: a continuously evolving system that shapes how we allocate our time, once we have chosen the future world state we are building toward.
The question "what should I work on?" contains within it every other question. It requires an estimate of which futures are reachable (forecasting), which are desirable (preferences), what actions will shift the probability distribution over those futures (causal models), and how your actions interact with the actions of everyone else (coordination). It requires understanding your own values, the constraints of the physical and institutional world, and the evolving intentions of the people you work with. A system that helps answer this question with calibrated confidence is, we believe, the most valuable thing that can be built.
No single human can hold all of this. No single model can compute it. And the intellectual history here is instructive about why. Hayek argued in the 1940s that markets succeed precisely because they aggregate knowledge that cannot be centralized — tacit, local, contextual knowledge embedded in individual experience. Central planning fails not because planners are incompetent but because the relevant knowledge is distributed in a form that resists aggregation. This was, for decades, a decisive argument for markets over planning.
Recent work complicates this picture. AI is making tacit knowledge codifiable at scale — converting the kind of contextual, experiential knowledge that Hayek argued was irreducibly local into explicit, transferable representations. If this continues, it undermines the strongest argument for pure decentralization. Centralized systems become more feasible when the knowledge they need to aggregate becomes legible.
But the resolution is probably neither pure markets nor pure planning. What Elinor Ostrom spent a career demonstrating — and what earned her the Nobel — is that the most robust governance systems are polycentric: multiple, overlapping decision centers that operate autonomously within their domains while coordinating on shared problems. Not a single market. Not a central plan. A network of decision-making nodes, each with local autonomy and calibrated awareness of the others.
This is what we are building toward. A system that continuously aggregates signals — from markets, from democratic delegation, from prediction systems, from the world itself — and projects those signals into a coherent model of what is worth doing next. Not by centralizing all knowledge into a single planner, and not by leaving all decisions to uncoordinated market signals, but by building the infrastructure for polycentric intelligence: local autonomy, shared world models, and calibrated coordination.
Our world model is the foundation. Our forecasting systems are the mechanism. Axion is the interface. And the work on understanding model fragility and forecaster evaluation is how we ensure the system is honest enough to be trusted with consequential decisions.
The goal is not to predict the future. It is to help choose it — and then coordinate the work of getting there.
