What should the new financial infrastructure look like in the AI era?

Original author: Matt Liston

Original translation by: AididiaoJP, Foresight News

In November 2024, prediction markets predicted the election results before anyone else. When polls showed a close race and experts were hesitant, the market gave Trump a 60% chance of winning. When the results were announced, the prediction markets outperformed the entire prediction establishment—polls, models, expert opinions, everything.

This proves that markets can aggregate fragmented information into accurate beliefs, and risk-sharing mechanisms are at work. Since the 1940s, economists have dreamed that speculative markets could outperform expert predictions, and now that dream has been validated on the grandest stage.

But let's examine the underlying economic principles.

The bettors at Polymarket and Kalshi provided billions of dollars in liquidity. What did they get in return? They generated a signal that the whole world could see instantly and for free. Hedge funds watched it, campaigns absorbed it, and journalists built data dashboards around it. Nobody had to pay for that intelligence; the bettors effectively subsidized a global public good.

This is the dilemma that prediction markets are mired in: the information they generate, which is also their most valuable part, is leaked the moment it's created. Savvy buyers won't pay for publicly available information. Private data providers can charge hedge funds exorbitant fees precisely because their data is inaccessible to competitors. Conversely, publicly available prediction market prices, no matter how accurate, are worthless to these buyers.

Therefore, prediction markets can only exist in areas where enough people want to "gamble": elections, sports, and internet memes. The result is a form of entertainment disguised as information infrastructure. The truly critical questions for policymakers—geopolitical risks, supply chain disruptions, regulatory outcomes, and technological timelines—remain unanswered because no one will bet on them for entertainment.

The economic logic of market prediction is upside down. Correcting this is part of a much larger transformation. Information itself is the product; betting is merely a mechanism for producing information, and a limited one at that. We need a different paradigm. Here is a preliminary outline of "cognitive finance": an infrastructure redesigned around information itself, starting from first principles.

Collective intelligence

Financial markets are a form of collective intelligence. They aggregate dispersed knowledge, beliefs, and intentions into prices, thereby coordinating the behavior of millions of participants who never communicate directly. This is remarkable, but also extremely inefficient.

Traditional markets operate slowly due to limitations in trading hours, settlement cycles, and institutional friction. They can only express beliefs broadly through the crude tool of price. The scope they can represent is also extremely limited; the space for tradable claims is negligible compared to the space of issues that truly concern humanity. Furthermore, participants are severely restricted: regulatory barriers, capital requirements, and geographical constraints exclude the vast majority of people and all machines.

The emergence of the crypto world has begun to change this, with features including uninterrupted markets, permissionless participation, and programmable assets. Modular protocols that can be assembled without central coordination. DeFi (Decentralized Finance) has proven that financial infrastructure can be rebuilt as open, interoperable foundational components born from the interaction of autonomous modules, rather than gatekeeper decrees.

However, DeFi is largely just a copy of traditional finance with better "pipelines". Its collective intelligence is still based on prices, focused on assets, and slow to absorb new information.

Cognitive finance is the next step: rebuilding intelligent systems from first principles for the era of artificial intelligence and encryption. We need markets that can "think" to maintain probabilistic models of the world, absorb information with arbitrary granularity, be queried and updated by AI systems, and allow humans to contribute knowledge without understanding the underlying structure.

The components that enable it are not mysterious: using private markets to refine economic models, using combinatorial structures to capture correlations, using an intelligent agent ecosystem to process information at scale, and using human-computer interfaces to extract signals from the human brain. Each component can already be built today, and when they are combined, they will create something new that is qualitatively transformative.

Private Market

If prices are not made public, economic constraints will be easily resolved.

A private prediction market only allows the entity subsidizing liquidity to see prices. This entity thus gains exclusive access to the signal, proprietary intelligence, rather than a public good. Suddenly, the market becomes viable for any question where "someone needs an answer," regardless of whether anyone is willing to bet for entertainment.

I've discussed this concept with @_Dave_White_.

Imagine a macro hedge fund seeking continuous probability estimates of Federal Reserve decisions, inflation outcomes, and employment data—as signals for decision-making, not as betting opportunities. They'd pay for this as long as the intelligence is exclusive. A defense contractor wants probability distributions of geopolitical scenarios, a pharmaceutical company wants forecasts of regulatory approval timelines. However, these buyers don't exist today because information, once generated, is immediately leaked to competitors.

Privacy is the foundation upon which economic models are built. Once prices are made public, information buyers lose their advantage, competitors begin to free-ride, and the entire system regresses to a point where it can only rely on entertainment needs.

Trusted Execution Environments (TEEs) make all this possible; they are secure computing enclaves where operations are invisible to the outside world (even system operators). Market state exists entirely within the TEE. Information buyers receive signals through verified channels. Multiple non-competing entities can subscribe to overlapping markets; tiered access windows strike a balance between information exclusivity and wider distribution.

TEE is not without its flaws; it requires trust in the hardware manufacturer. However, it offers sufficient privacy for commercial applications, and the related engineering technology is now quite mature.

Combination Market

Current prediction markets treat events as isolated phenomena. "Will the Fed cut rates in March?" is in one market. "Will inflation exceed 3% in the second quarter?" is in another. A trader who understands the intrinsic connections between these events—for example, knowing that high inflation may increase the probability of a rate cut, or strong employment may decrease it—must manually arbitrage between these disconnected pools of funds, attempting to rebuild the correlations that have been destroyed by the market structure itself.

It's like building a brain where each neuron can only fire in isolation.

Unlike other markets, combined prediction markets maintain a "joint probability distribution" of multiple outcomes. A trade expressing "interest rates remain high and inflation exceeds 3%" will create ripples in all relevant markets within the system, synchronously updating the entire probability structure.

This is similar to how neural networks learn: during training, each gradient update simultaneously adjusts hundreds of millions of parameters, and the entire network responds holistically to every piece of data. Similarly, when a portfolio predicts every transaction in the market, it updates its entire probability distribution; information propagates through the correlation structure, rather than simply updating isolated prices.

What ultimately emerges is a "model," a probability distribution that is continuously updated across the world's event state space. Each transaction optimizes this model's understanding of the correlations between events. The market is learning how the real world is interconnected.

Smart ecosystem

Automated trading systems have already taken over Polymarket. They monitor prices, detect mispricing, execute arbitrage, and aggregate external information far faster than any human.

Current prediction markets are designed for human bettors using web interfaces. Intelligent agents participate "barely" under this design. An AI-native prediction market would completely reverse this logic: intelligent agents would become the main participants, while humans would be accessed as information sources.

Here is a crucial architectural decision: complete isolation must be achieved. An agent that can see prices must never also be an information source; and an agent responsible for acquiring information must never have access to prices.

Without this "wall," the system would self-destruct. An agent capable of both acquiring information and observing prices could deduce valuable information from price fluctuations and then seek it out itself. In this scenario, market signals would become "treasure maps" guiding others. Information acquisition would degenerate into a complex form of "forward-looking trading." The isolation mechanism ensures that information-acquiring agents can only profit by providing truly novel and unique signals.

On one side of the "wall" are trading agents that compete in complex combinatorial structures to identify mispricing; and evaluation agents that assess incoming information through adversarial mechanisms to distinguish between signals, noise, and manipulation.

On the other side of the "wall" are information-acquiring agents, operating entirely outside the core system. They monitor data flows, scan documents, contact individuals with unique knowledge—and unilaterally transmit information to the market. When their information proves valuable, they receive compensation.

Compensation flows backward along the chain. A profitable transaction rewards the agent that executed the transaction, the agent that evaluated the information, and the agent that initially provided the information. This ecosystem thus becomes a platform: on the one hand, it allows highly specialized AI agents to monetize their capabilities; on the other hand, it serves as a foundational layer for other AI systems to gather intelligence to guide their actions. The agent is the market itself.

Human intelligence

Much of the world's most valuable information exists only in the human mind. For example, engineers who know their products are behind schedule; analysts who perceive subtle shifts in consumer behavior; and observers who notice details invisible even to satellites.

An AI-native system must be able to capture these signals from the human brain without being overwhelmed by massive amounts of noise. Two mechanisms make this possible:

Intelligent agent mediation: This allows humans to "trade" without seeing prices. A person simply expresses their belief in natural language, such as "I believe the product launch will be postponed." A dedicated "belief-translating agent" analyzes this prediction, assesses its confidence level, and ultimately translates it into a market position. This agent coordinates with the system that has access to prices to construct and execute the order. Human participants only receive rough feedback: "Position established" or "Unfavorable position." Compensation is settled after the event based on the accuracy of the prediction, and price information remains confidential throughout the process.

Information Markets: These markets allow intelligent agents to directly pay for human signals. For example, an agent wanting to understand a technology company's profitability can identify an engineer with relevant insider knowledge, purchase an assessment report, and then verify and pay based on the information's subsequent market value. Humans are paid for their knowledge without needing to understand complex market structures.

Take analyst Alice as an example: Based on her professional judgment, she believes a certain merger will not receive regulatory approval. She inputs this view through a natural language interface. Her "belief translation agent" analyzes the prediction, assesses her confidence from the details of the language, checks her historical records, and constructs an appropriate position, without ever accessing the price. The "coordination agent" located at the TEE boundary then determines whether her view has an informational advantage based on the implied probability of the current market and executes the trade accordingly. Alice only receives notifications that "position established" or "insufficient advantage." The price remains confidential at all times.

This architecture views human attention as a scarce resource requiring careful allocation and equitable compensation, rather than a public resource that can be freely exploited. As such interfaces mature, human knowledge will become "flowing": the information you possess will converge into a global model of reality and be rewarded when proven correct. Information trapped in the mind will no longer be trapped.

Future Vision

If we broaden our perspective enough, we can see where all of this will lead us.

The future will be an ocean of fluid, modular, and interoperable relationships. These relationships will spontaneously form and dissolve between human and non-human participants, without any central gatekeeper. This is a kind of "fractal autonomous trust."

Intelligent agents negotiate with each other, and humans contribute knowledge through natural interfaces. Information flows continuously into a constantly updated reality model that anyone can query, but no one can control.

Today's prediction markets are merely a rough sketch of this picture. They validate the core concept (that risk-sharing generates accurate beliefs), but are trapped in flawed economic models and structural assumptions. Sports betting and election predictions are to cognitive finance what ARPANET (the early internet) is to today's global internet: a "proof of concept" mistakenly considered the ultimate form.

The true "market" is actually every decision made under uncertainty—virtually all decisions. Supply chain management, clinical trials, infrastructure planning, geopolitical strategies, resource allocation, personnel appointments… the value of reducing uncertainty in these areas far outweighs the entertainment value of betting on sporting events. We simply haven't yet built the infrastructure to capture this value.

What's coming is the "OpenAI moment" in the field of cognition: an infrastructure project on a civilizational scale, but its goal is not individual reasoning, but collective belief. Large language model companies are building systems that "reason" from past training data; cognitive finance aims to build systems that "believe"—systems that maintain a calibrated probability distribution about the state of the world, continuously updated through economic incentives (rather than gradient descent), and integrating human knowledge with arbitrarily high granularity. LLM encodes the past; prediction markets aggregate beliefs about the future. Only by combining the two can a more complete cognitive system be formed.

When fully expanded, this will evolve into an infrastructure: AI systems can query it to understand the uncertainty of the world; humans can contribute knowledge to it without understanding its internal mechanisms; it can absorb partial knowledge from sensors, domain experts, and cutting-edge research, and synthesize it into a unified model—a self-optimizing, predictive model of the world; a foundation where uncertainty itself can be traded and combined. The resulting intelligence will transcend the sum of its parts.

The computer of civilization—this is precisely the direction that cognitive finance strives to build.

Stakes

All the pieces of the puzzle are in place: intelligent agents have crossed the threshold of predictability; confidential computing has moved from the lab to production environments; and prediction markets have demonstrated large-scale product-market fit in the entertainment industry. These threads converge on a specific historic opportunity: building the cognitive infrastructure needed for the age of artificial intelligence.

Another possibility is that prediction markets will forever remain at the entertainment level, performing accurately during elections but ignored at other times, never touching upon truly important issues. In that case, the infrastructure upon which AI systems rely to understand uncertainty will cease to exist, and the valuable signals imprisoned in the human mind will be forever silenced.