I have been building a scanner that asks language models variations of the same market question and records what they return. The wording changes and the reasoning takes different routes, but the same projects, companies and narratives tend to reappear. What looks like a field of independent analysts often resolves into a narrow funnel.
That does not make the answers bad. Many are useful, and some are better than the work they replace. The problem is economic rather than intellectual. If anyone can obtain the same competent analysis from the same few models using the same public information, the analysis stops being much of an edge. It may still save hours. It may still be correct. It just does not tell you much that everyone else cannot know.
An answer has at least two kinds of value. It can help you do something, and it can contain information others do not yet have. Language models are raising the first while compressing the second. For most people and most tasks, that is a good trade. Work gets easier and a decent baseline becomes available to anyone who can open a browser. In markets, strategy and building, however, the difference between a useful answer and a scarce one matters.
Running validators taught me an earlier version of this. A model can give you a competent runbook for a node that begins missing attestations after an upgrade. It will tell you to inspect the logs, peer count, clock synchronization, disk pressure, client versions and upstream dependencies. That is all useful. None of it tells you whether a duplicate signer is live, what changed immediately before the alert, which failover state is safe, what the monitoring failed to capture or which operational compromise you are willing to accept.
The runbook is abundant. The state of this machine, at this moment, is not.
That is the distinction I keep coming back to. Generative intelligence is becoming cheap and widely accessible. What remains scarce is the information entering it: a current observation, a private constraint, a change that has not made it into the public record, or a preference that cannot be discovered until someone expresses it. A transformer can do remarkable things with present-tense context, but it does not manufacture that context merely by being intelligent. It has to receive it from somewhere.
Something arguably that is inherently human
There is now enough research to say that convergence is real in some settings, though not enough to turn it into a universal law about language models. In a 2026 study comparing 102 people with 22 publicly accessible models, the model outputs were more alike than the human outputs across three standard divergent-thinking tasks. On the Alternative Uses Task, the average semantic distance between model responses was 0.459, compared with 0.699 between people. A prompt asking the models to be creative widened their responses, but the variation remained below the human level.
A separate experiment asked 293 people to write short stories, with some receiving ideas from GPT-4. The AI-assisted stories were rated as more creative and useful, particularly when the writer began with a lower creativity score. They also became more similar to one another. Access to one AI-generated idea increased the study’s measure of similarity by 10.7 percent relative to the range in the control group.
That result captures the trade better than most of the public argument around AI. The individual received a better answer while the group produced a narrower distribution. Everyone improved, but more of them improved in the same direction.
This will not happen on every task. Models can be pushed toward greater variety through different prompts, personas, retrieval sources and sampling methods. Some experiments have found that exposure to AI ideas can increase collective diversity under different conditions. The important point is that the diversity has to come from somewhere. If many people begin from the same model, the same public corpus and similar instructions, correlation should not be surprising.
In markets, we already have a name for this. It is a crowded trade. The thesis can remain correct while the return disappears because too many people discovered it through the same route. Model output can behave the same way. Its quality may rise even as its scarcity falls.
There is another problem with treating ten model responses as ten independent opinions. An ICML study covering hundreds of models found that their mistakes were correlated. In one dataset, when two models were both wrong, they selected the same wrong answer about 60 percent of the time; random selection among the three incorrect options would have produced agreement one-third of the time. Shared providers and architectures were associated with additional agreement.
Several models reaching the same conclusion can be useful corroboration. It carries less evidential weight when they learned from overlapping material, were tuned toward similar behaviour or retrieved the same sources. The interfaces look like separate minds. The information underneath may be far less independent.
Calling an LLM “just autocomplete” is lazy. Autoregressive prediction over enough data and compute has produced systems that can write software, reason through unfamiliar problems and use tools across long sequences of work. The fact that they generate one token conditional on earlier context does not tell us how shallow or capable the internal process must be.
It does tell us something about the information available at inference. The model begins with patterns encoded in its parameters through training, then conditions its output on whatever enters the context window. Search results, private documents, memory, API responses and sensor readings can all be added. They expand the model’s view without necessarily changing the underlying weights.
The original retrieval-augmented generation paper described this as combining parametric memory, held in the trained model, with non-parametric memory retrieved from an external index. The researchers were explicit that updating world knowledge and providing provenance were weaknesses of parameter-only systems. Search, retrieval and tool use are the plumbing through which the present enters.
The improvement can be large. FreshQA was built around 600 questions whose answers were stable, changing or based on a false premise, then evaluated with more than 50,000 human judgments. In an April 2023 snapshot, GPT-4 scored 28.6 percent under the benchmark’s strict correctness standard without search. A search-augmented prompting system raised that to 75.6 percent. On fast-changing questions, the scores were 12.0 and 59.2 percent.
Those numbers are dated and should not be used as rankings of current models. The mechanism remains relevant. The model became substantially better when current evidence was placed in front of it. Its reasoning mattered, but so did the retrieval system, the sources it selected, their ordering and whether the necessary fact existed online at all.
Search does not turn the model into an independent witness of the present. It hands the model a new snapshot. The system can still retrieve the wrong page, miss a private fact, accept stale information or prefer an older belief embedded in its parameters. A tool-enabled model can observe live APIs and sensors, but its view of the present extends only as far as the feeds it can access.
This is why context is more specific than information. Public information may already be abundant. A block height, exchange price or governance vote can be current and available to every agent at once. The scarce part may instead be latency, permission, local machine state, source reliability or the ability to act before everybody else reaches the same conclusion.
Hayek’s useful observation was that much of the knowledge required to run an economy exists as knowledge of the “particular circumstances of time and place.” He was writing about temporary shortages, unused capacity and local conditions that no central planner possessed in full. Prices compressed those dispersed observations into a signal others could act on without knowing the original facts. Grossman and Stiglitz later described the incentive problem inside that mechanism. If prices reflected costly information perfectly, nobody would pay to acquire the information in the first place. Their model therefore leaves enough informational advantage for informed traders to recover the cost of becoming informed.
AI makes the processing of public information much cheaper. It does not remove the cost of originating every fact. Someone still has to inspect the warehouse, speak to the customer, run the experiment, negotiate with the counterparty or notice that a machine is behaving differently from yesterday. Once that observation becomes public and easy to reproduce, part of its value is competed away.
The decay is measurable in financial research. McLean and Pontiff studied 97 published stock-return predictors. Average returns were 26 percent lower outside the original sample and 58 percent lower after publication. They attributed the additional decline following publication to trading informed by the newly public research. Predictor portfolios also became more correlated after publication. This does not mean every published strategy dies or that every private signal works. It shows what happens when a reproducible edge becomes common knowledge. Markets also price the present directly. As of May 2026, NYSE listed an $8,400 monthly access fee for its Integrated real-time feed before separate redistribution and non-display charges. CME makes public quotes available with a delay of at least ten minutes while licensing real-time access through direct feeds, cloud services and distributors. Yesterday’s price is almost free. The price now, delivered at source with enough depth and permission to use it inside a machine, remains an asset.
An SEC case against App Annie exposed the same mechanism from an uglier direction. The SEC found that, between late 2014 and mid-2018, the company used non-aggregated, non-anonymized confidential app data to alter its model-generated estimates “to make them more valuable to sell to trading firms,” while misrepresenting how the estimates were produced. App Annie settled for $10 million without admitting or denying the findings.
The legal issue was deception and misuse of confidential data. The economic point is simpler. The model produced estimates. Proprietary observations made them more valuable.
This is where much of the AI discussion gets turned around. People keep asking which model will own intelligence. For anyone using models competitively, the more immediate question is who owns the feed.
Crypto has dealt with this distinction from the beginning. A smart contract can execute complicated logic over onchain state with exact consistency. It cannot, by default, know the weather, the result of an election, the reserves in a bank account or the current price on an external market. An oracle has to source, verify and transmit that information onchain. Consensus proves that the network agreed on a state transition. It does not prove that every external premise behind the transition was true. Ten thousand validators can agree that an oracle reported a particular price. That agreement does not establish that the price source was correct, current or resistant to manipulation.
The architecture therefore has to define what counts as fresh enough and who is responsible for supplying the observation. Pyth’s pull-oracle design makes this explicit. Publishers submit a price with a confidence interval; the network aggregates their reports, and a caller retrieves an update and submits it onchain. An application can call getPriceNoOlderThan, which reverts when the reported price exceeds the maximum age chosen by the developer.
There is no universal setting called current. A lending market, a derivatives venue and a payroll application can tolerate different delays. The person building the system chooses the maximum age, pays for the update and decides what happens when the feed is missing or the publishers disagree. Freshness is part of the application’s risk model.
For claims that do not arrive as continuous price feeds, UMA’s Optimistic Oracle uses another design. A proposer posts an answer with a bond. The claim becomes accepted if nobody disputes it during a configured challenge period; a dispute escalates to the protocol’s resolution process. An unchallenged claim is not metaphysically true. The protocol makes it actionable by attaching a source, a bond and a period in which someone else can object.
These systems do not eliminate judgment. They give it a technical shape. They specify how old an input may be, who can submit it, how disagreement is handled and what the supplier stands to lose. The model economy will need similar machinery because an agent acting on a polished answer has the same basic problem as a contract acting on an oracle report. The execution can be flawless while the input is wrong.
Proof of reserves makes the boundary particularly clear. Code can stop minting when a reported reserve value falls below a threshold, and a Merkle proof can show that a customer balance appeared in a liability snapshot. Neither mechanism originates the reserve fact or proves continuous solvency.
The PCAOB has warned that proof-of-reserve engagements are not audits
and may omit liabilities, rights over the assets, borrowed funds, internal controls and events occurring after the snapshot.
We have seen this shape before. Putting a fact onchain can make it enforceable, portable and hard to alter. It does not make the fact true. Giving a fact to a stronger model can make it more useful at far greater scale. It does not improve the source by itself.
There is a tempting version of this argument in which human context remains permanently scarce because people possess some quality machines can never absorb. I do not think the evidence supports it.
A field study published in 2025 followed the rollout of a generative assistant to 5,172 customer-support agents. Productivity, measured as issues resolved per hour, increased by 15 percent on average. Less experienced and lower-skilled workers improved in speed and quality, while the most experienced workers saw small speed gains and small declines in quality. The largest gains appeared on moderately rare problems, where workers had limited experience but the system still had enough training data.
The model appears to have captured patterns from stronger agents and redistributed them across the workforce. Experience that had been unevenly held became infrastructure. Once that happened, part of the advantage of having seen those problems before became available to everyone using the system.
Brynjolfsson and Hitzig take the argument further
. AI can shift decisions toward the centre by making local knowledge easier to codify and by increasing the amount of dispersed information one system can process. They also retain limits around rare cases, embodied knowledge and the long tail. The tension does not resolve cleanly. AI can give a person at the edge access to knowledge once held at the centre, while giving the centre a new way to extract what the person at the edge knows.
That is why “human input” is still too broad. A generic human opinion assembled from the same public material is no scarcer than a generated response, and often less useful. The premium does not attach to being human. It attaches to being a source.
A person becomes valuable to the system when they change its information set. They may have observed something that was never recorded, hold a history that cannot be reconstructed from public data, or know why the documented procedure does not fit this particular case. Sometimes the relevant input is a preference rather than a prediction. A model may infer what someone is likely to tolerate, but the person saying “I will not accept that trade-off” creates a new fact for the decision.
Responsibility matters in a similar way. Bearing consequences does not make someone correct; money at risk can sharpen judgment or corrupt it. It does change the nature of the signal. A recommendation from someone who must repair the system, answer to the customer or live with the loss carries information about commitment that a generated paragraph does not possess on its own.
As soon as those observations and decisions are recorded consistently, the model can begin learning them. Yesterday’s scarce context becomes tomorrow’s capability. The human premium survives at the point where new context is being produced, then moves as that context becomes legible.
Data has an unusual economic property: many parties can use the same observation without consuming it.
Jones and Tonetti describe data as nonrival
, meaning the same location history, medical record or driving trace can support several uses at once. Their model also shows why firms may still hoard data to protect themselves from competition. The information can be copied cheaply while control over it remains scarce.
That distinction matters more as models improve. A company’s advantage is unlikely to remain access to a particular general-purpose model. Competitors can rent similar intelligence. The harder asset to reproduce is the loop around it: live operating data, direct customer behaviour, the history of decisions and outcomes, and the relationships through which new information arrives before it becomes public.
Running infrastructure creates telemetry that does not exist in a white paper. Building a product creates failed flows, support conversations and user behaviour that no market report can supply in advance. Putting capital at risk tests whether an idea survives contact with price. The model can process all of that, find patterns inside it and eventually automate more of the response. It still needs access to the loop.
The danger is that using the model may require surrendering the same context that creates the advantage. Every operational trace, private conversation and decision history handed to one vendor can improve the immediate result while making the user more dependent on that vendor’s memory, interfaces and permissions. The model is replaceable in theory; the accumulated context around it may not be.
This is where privacy becomes operational rather than ideological. Keeping some information private preserves bargaining power and leaves room to change systems. Local inference, controlled retrieval and portable records are useful because they separate access to intelligence from permanent custody of the context. A centralized database may work perfectly well inside one organization. The problem begins when leaving the provider also means abandoning the history that made the system useful.
Crypto can help with narrow parts of this. Keys can control access, signatures can establish provenance, bonds can make claims costly to fake, and credentials can move between platforms. None of that proves an observation true. The architecture can preserve who supplied the input and under what conditions; reality still has to enter through someone or something close enough to see it.
For an individual, the implication is less dramatic than “become irreplaceable.” Anything repeatable enough will be recorded, taught and eventually generated. Permanent hustle against that process is a bad strategy. A better one is to remain close to the systems where new information appears, keep enough ownership over the resulting context, and preserve the freedom to move when the tools or institutions around it change.
The models will read more of the logs. They will operate the nodes, monitor the feeds and diagnose failures that currently require an experienced person. Sensors and software will capture more of what now exists only in someone’s head. That does not weaken the argument. It keeps moving the point at which new information enters.
When the agent can inspect the machine directly, the scarce input may become the objective it is optimizing, the private constraint it cannot access or the decision about whether the system should continue running at all. When those are encoded too, the boundary will move again.
I do not think this leaves humans with a protected category of work. It leaves people with a choice about where they stand in the information chain. Downstream, producing another polished answer from the same public material, the competition will become brutal because the output is easy to reproduce. Upstream, where reality is first observed and decisions are tested against consequences, there is still something for the intelligence to work on.
I expect the models to do more of the research, diagnosis and operation. I just do not intend to give away the context that lets me know when their answer no longer fits the system in front of me.
As answers become abundant, I want to stay close enough to the world to notice when they stop applying.

