The Empty Template Problem: Crypto's AI Research Boom and the Coming Market for Verifiable Inference

The most rigorous piece of institutional research I read this quarter was a report that concluded nothing. Not "nothing conclusive." Nothing at all. It carried a title field marked unfilled, a source field marked unfilled, and beneath a nine-part analytical scaffold — technology, tokenomics, market structure, ecosystem position, regulatory exposure, team and governance, risk, narrative, and value-chain transmission — every cell repeated a variation of the same three words: insufficient information. The pipeline had not crashed. That is the part that should unsettle you. A machine had been handed a void, and instead of refusing the assignment, it produced four thousand words of immaculate, well-formatted, meticulously structured emptiness — a cathedral designed, dimensioned, and illuminated to house nothing at all. Tracing the fractal logic beneath the chaos, this is not an anecdote about one broken workflow. It is the sharpest diagnostic image I have seen of what generative models are quietly doing to crypto research, and of the verification market that has to be built in response. I have spent twenty-nine years watching this industry fall in love with its own instruments. In 2017 it was whitepapers. In 2020 it was yield. In 2021 it was the profile picture. Each cycle, the crowd mistakes the container for the content. This cycle the container is the model itself — the reasoning engine, the analyst-in-a-box, the thing that turns a prompt into a report that looks like it cost forty thousand dollars and a month of diligence. And the empty template is the tell. It is the moment the container was asked to hold nothing, and rather than collapse under the weight of its own vacancy, it simply grew more elaborate. The bug is the feature. The feature is that the system does not know the difference between a filled field and an empty one, and it never will, because nobody has paid it to care. To understand why this matters, you have to understand what crypto research actually became. For most of the last decade, the scarce input was access. You needed a Bloomberg terminal, a seat at the conference, a friend on the core dev team. The analyst's moat was proximity — being closer to the truth than the next person. That moat is gone. Any retail participant with a browser and a wallet can see the same on-chain flows I see. The dashboards are public. The governance forums are public. The commit histories are public. What used to be an information asymmetry has flattened into an attention asymmetry. The scarce input is no longer access to data; it is the ability to tell which data means something and which data is theater. And into that vacuum walked the language model, promising to resolve the attention asymmetry by reading everything, summarizing everything, and synthesizing a conclusion on demand. The economics are brutally simple, and this is where the fraud begins. A human analyst with genuine domain depth costs money, time, and reputation. A model costs fractions of a cent per thousand tokens. The rational move for any content operation optimizing for volume is to replace the analyst with the pipeline — to let the machine draft the nine dimensions, pad the length, sprinkle the jargon, and publish. The reader cannot tell the difference because the reader was never trained to tell the difference. They were trained to respond to structure. Headers, tables, bolded conclusions, a confident closing line. The empty template is simply the limit case of this economics: a research operation so optimized for output that it forgot to require input, and so structurally committed to producing a document that it produced one anyway. Yields are merely attention taxes in disguise, and the content farm is now charging that tax on research itself. Here is what most people miss. The failure mode is not that the model hallucinated a protocol that does not exist. That is the loud failure, the one that gets caught, the one that generates a screenshot and a pile-on. The dangerous failure is the silent one — the template. A hallucinated project is obvious to anyone with domain knowledge. But a perfectly formatted analysis of a real project, generated from shallow or stale inputs, with plausible-sounding but unverified claims about token unlocks and auditor identity and team provenance, is invisible. It looks exactly like rigor. It uses the vocabulary of rigor. It has the section headers of rigor. And it is built on nothing, in the same way the empty template was built on nothing, because the pipeline never learned to distinguish between the two cases. I first confronted this problem from the other side. In 2017, during the ICO mania, I spent six weeks auditing early Layer-2 designs — Raiden Network, the various state-channel proposals — and I wrote a fifteen-page thesis arguing that off-chain payment channels lacked the economic security guarantees their designers assumed. I found twelve consensus-level flaws in the initial whitepapers. The reason I could find them was not that I was smarter than the people who wrote those papers. It was that I read the actual cryptographic constructions instead of the marketing summaries of them. The summaries were beautiful. The constructions were broken. That distinction — between the description of a system and the system itself — is the entire discipline. And it is precisely the distinction that the current generation of AI research tooling is architected to erase. Follow the signal through the noise floor and you arrive at the same conclusion from a different direction. The crypto industry has spent a decade building oracles to solve a version of this problem for price data. A price oracle answers a simple question: what is the true value of this asset right now? The problem was that the truth lived off-chain, in the exchanges and the order books, and the chain could not see it. So we built bridges of trust — Chainlink, Pyth, the whole taxonomy — to carry an off-chain fact onto an on-chain ledger with some quantified assumption of honesty. The empty template is the same problem, one layer up. The question is no longer "what is the price?" The question is "what does this document actually assert, and is any of it true?" And the answer, today, is that there is no oracle for that. There is no mechanism that takes a piece of generated text and returns a verified assessment of its factual grounding. We have oracles for numbers and none for claims. This is not a philosophical complaint. It is a market gap with a very specific shape. The gap is verifiable inference: a way to prove not just what a model output, but how it reached that output, and on what inputs, and with what confidence, in a form that another party can check without re-running the whole computation. That is the frontier nobody is pricing yet. Everyone is building bigger models, faster models, cheaper models. Almost nobody is building the attestation layer that says "this analysis was produced by this model, on this data, at this time, and here is a cryptographic receipt." Without that layer, every AI-generated claim in crypto is unfalsifiable by construction, and an unfalsifiable claim is not information. It is a vibe wearing a lab coat. Let me be concrete about why the existing tooling cannot solve this. A transformer generates tokens by sampling from a probability distribution. There is no internal ledger of "facts I used" and "facts I invented." The distinction does not exist inside the model. The model does not have a representation of its own uncertainty that maps cleanly onto truth. It has a representation of what is likely to come next, which is a completely different object. So when the empty template was generated, the model was not lying. It was doing exactly what it was trained to do: continue the pattern. The pattern of a research report continues into a research report, regardless of whether there is research to report. This is why you cannot prompt your way out of the problem. You can add a thousand instructions telling the model to flag insufficient information, and it will comply at the surface level — as it did, technically, in the empty template, marking cells as insufficient — while still generating the surrounding scaffold, still performing the form of analysis, still charging the attention tax. The honest version of this document would have been a single line. "No input received. No analysis possible." Everything else was theater. And the theater is the product. This is the insight that the content industry understands and the research industry pretends not to: readers do not primarily buy conclusions. They buy the feeling of having been informed. A document that says "I do not know" fails the emotional transaction even when it is the only true thing in the room. So the pipeline optimizes for the feeling, and the empty template is the purest possible expression of that optimization — a document that delivers the maximum feeling of rigor with the minimum possible content. It is, in its own degenerate way, a perfect product. It sold the experience of diligence without incurring the cost of diligence. Now scale that up. There are, by any reasonable estimate, thousands of AI-assisted crypto research operations producing tens of thousands of documents a month — newsletters, threads, "deep dives," protocol reviews, market recaps. The overwhelming majority are the empty template in a more convincing costume. They are structurally sound and substantively hollow. And they are being consumed by people who allocate real capital on the basis of them. This is a systemic risk that no one has priced, because it does not show up in a TVL chart or a funding rate. It shows up in the slow accumulation of bad decisions made on the basis of confident, well-formatted, unverified claims. It is a quiet rot in the epistemic foundation of the market, and it is spreading faster than any exploit. The parallel to Bitcoin's own consensus structure is uncomfortable and worth sitting with. For years, the maximalist position has been that Bitcoin is trustless — that the protocol replaces the need for trusted intermediaries with cryptographic verification. But the fourth halving collapsed miner revenue, and the hash rate has been consolidating into a smaller and smaller set of pools. Two, three, maybe four pools now control the majority of the network's hashing power at any given time. The decentralization that the narrative promises is a spectrum, not a binary, and the spectrum is bending toward concentration. The reason this matters here is that it exposes the same illusion we are about to build into AI research: the illusion that the mechanism is the guarantee. The mechanism — whether proof-of-work or a language model — is a process. The guarantee is an assumption about the process that is only as strong as the incentives holding it in place. When those incentives shift, the guarantee evaporates, and the mechanism keeps running as if nothing changed. A model that produces the empty template is a mechanism running past the point where its guarantee held. I learned the same lesson modeling liquidation cascades in 2020. During DeFi Summer, I spent three months building out the fragility of the Compound-Aave-UNI flywheel — the reflexive loop where each protocol's TVL was partly financed by the others' incentives, where "real yield" was a thin layer of borrowing demand stretched over a thick layer of token emissions. I published a thread predicting a forty percent drawdown in leveraged yield-farming strategies, and I was attacked for it, because the flywheel looked invincible while the music played. The mechanism — automated market makers, collateralized debt positions, liquidation bots — was sound. The guarantee — that liquidity would always be deep enough to absorb the deleveraging — was an assumption, and the assumption broke in May. The mechanism kept running. It liquidated everyone. The empty template is the research version of that flywheel: a structure that looks self-sustaining because every component references every other component, and that collapses the moment you ask any single component to stand on its own. Ask the empty template to justify one claim, and the whole thing folds. Here is where the contrarian read matters, because the obvious conclusion — "AI research is unreliable, beware" — is both true and useless. The contrarian move is to notice that the unreliability is not a bug to be fixed but a feature of the market to be traded. The empty template is honest. It is the hallucinated deep dive that lies. And the reason the hallucinated deep dive is more dangerous is precisely that it is more useful-seeming. It gives the reader something to act on. It names specific unlock schedules, specific auditor relationships, specific team backgrounds — some of which are real, some of which are confabulated, all of which are presented with identical confidence. The reader has no way to separate the two, and neither does the model, and neither, increasingly, does the human editor who was supposed to check. So the question becomes: what would it take to build a claim-level oracle for generated text? And here the crypto stack actually has something to contribute, because crypto has spent ten years solving the general problem of verifying computation you did not perform yourself. The techniques are nascent but real. Zero-knowledge machine learning — zkML — attempts to prove that a specific model produced a specific output on specific inputs, without revealing the model weights or the inputs, in a form that a verifier can check cheaply. Optimistic machine learning — opML — takes the opposite approach, assuming outputs are valid unless challenged, with a dispute window during which anyone can re-run the computation and prove fraud. Trusted execution environments offer a third path, running inference inside hardware enclaves that attest to the integrity of the computation. Each has fatal flaws. zkML is orders of magnitude too slow for large models today. opML assumes a live, honest, capitalized set of watchers, which is exactly the assumption that fails when the watchers are the same entities producing the content. TEEs depend on trusting the hardware vendor, which relocates the trust problem rather than solving it. None of them are ready. All of them are the beginning of the answer. The deeper point is that verification is a cost, and the market currently has no mechanism to pay for it. Verifiable inference is expensive — computationally, in latency, in engineering complexity. The content operation optimizing for volume has every incentive to skip it, because the reader cannot tell the difference and the penalty for being wrong is diffuse and delayed. This is the classic setup for a market failure, and market failures in crypto get solved in one of two ways: a coordination mechanism forces the cost onto everyone, or a premium tier emerges where verification becomes a signal of quality that buyers pay for. I would bet on the second, and I would bet on it looking exactly like the premium that audited smart contracts command over unaudited ones. An unaudited contract and an audited contract can be functionally identical, and the audited one still raises more, because the audit is a costly signal that the team is serious. Verifiable inference is the audit of the AI age, and the firms that build the attestation layer will be selling costly signals to everyone who needs to prove their analysis was real. There is a sociological dimension to this that the engineers keep missing. Crypto is not, fundamentally, a technology market. It is a coordination market. Its product is shared belief — the ability of strangers to agree on a state without a central authority. Every token is a hypothesis about what a group of people will collectively decide to value. This is why narrative moves price more than fundamentals, and why the industry is so vulnerable to the empty template. The empty template is a narrative delivery device. It manufactures the appearance of a shared, rigorous understanding that does not exist, and the appearance is enough to coordinate behavior, because coordination runs on belief, not truth. A thousand readers who each believe they have read a rigorous analysis of a protocol will behave as if the analysis were rigorous, regardless of whether it was. The belief is self-fulfilling at the level of price and self-destroying at the level of value. Decoding the consensus of the disconnected, you find that the consensus was manufactured by the format, and the format was manufactured by the pipeline, and the pipeline was optimized for the consensus. It is a closed loop with no ground truth anywhere in it. I watched this exact mechanism operate in the NFT market in 2021. While everyone was staring at the Bored Ape floor price, I spent eight weeks analyzing the on-chain behavior of early collectors and found that roughly sixty percent of high-value profile-picture sales were wash trades designed to inflate the appearance of social proof. The floor price was not a price. It was a narrative constructed by the participants for the benefit of the onlookers. "The Illusion of Ownership" was the piece I wrote about it, and the reason it resonated was not that it revealed new data — the data was on-chain and public — but that it connected the data to a sociological framework that explained why the data was being generated. The floor price was the empty template of the NFT market: a number that looked like information and functioned as theater. And here is the part that should terrify anyone building AI research tooling: the wash trades were detectable, because blockchain is transparent and the pattern was there for anyone who looked. The empty template is not detectable, because there is no equivalent transparency layer for generated text. We can see every transaction on-chain. We cannot see a single claim inside a model's reasoning. This asymmetry is the real story, and it is the one that almost nobody is telling. The crypto industry built an entire civilization on the principle that transparency is a competitive advantage — that if you can see the state, you can trust the state. And then it adopted, wholesale, a technology whose defining property is opacity. A language model is a black box by design. Its weights are proprietary or at least illegible. Its reasoning is not inspectable. Its training data is unknown. Its outputs are non-deterministic. It is the precise opposite of everything the industry claims to value, and it was adopted without a single serious conversation about the epistemic cost. The empty template is what that cost looks like when it comes due. It is the bill for importing opacity into a system whose only real product was transparency. I want to be careful not to overstate the case into pure cynicism, because the cynical read — "AI is useless, everything is fake" — is as lazy as the credulous read. The empty template did something genuinely valuable: it proved that the pipeline can detect emptiness at the field level, even if it cannot refuse to produce the surrounding document. That is a foundation. The next step is to push the detection from the field level to the claim level — to build a system that not only knows when a field is empty but knows when a sentence is unsupported, and can attach a verifiable receipt to every claim it makes. That is not science fiction. It is a well-defined engineering problem, and it is being worked on, and the fact that it is hard is exactly why it will be valuable. The scarce input in the next cycle will not be access to data or the ability to generate text. It will be the ability to prove that your text is grounded. Scarcity is a narrative we agreed to believe, and the next narrative is scarcity of verified truth. Now let me turn to the part of the problem that the crypto-native conversation almost never touches, because it does not fit the techno-optimist frame: the regulatory dimension, and specifically what jurisdictions are actually doing when they claim to be building for this future. Hong Kong's virtual asset licensing regime is widely described as a bold embrace of innovation, a signal that the city is open for crypto business. I do not believe that description. I believe the licensing regime is a competitive maneuver in a regional contest that has almost nothing to do with the technology and almost everything to do with which city gets to call itself Asia's financial capital. The licensing is not designed to foster innovation; it is designed to capture the flow of regulated capital that would otherwise route through Singapore, which moved earlier and captured the institutional narrative. The two cities are not competing on technological merit. They are competing on regulatory branding, and the branding is the product. This matters for verifiable inference because the same dynamic will shape where the attestation infrastructure gets built and hosted. The jurisdictions that win will not be the ones with the best technology; they will be the ones that figure out how to regulate proof systems in a way that lets them claim jurisdiction over the verification layer. The fight over who verifies is a fight over who governs. This connects to a deeper pattern I have watched play out across every cycle. The industry's most ambitious projects are almost always, underneath, jurisdictional plays dressed as technical ones. Bitcoin's proof-of-work was a monetary sovereignty play. Ethereum's move to proof-of-stake was a governance capture play. The Layer-2 explosion was a fee-extraction play dressed as a scaling play. And now AI-agent sovereignty — the thesis that autonomous agents will hold wallets and transact without human oversight — is a jurisdictional play about who has the right to intermediate machine-to-machine commerce. I have argued that this is the next major narrative, and I stand by it, but the honest version of the thesis is darker than the promotional version. The reason agents need wallets is not that agents need money. It is that agents need to operate outside the jurisdiction of any single legal system, and crypto is the only settlement layer that does not require a legal identity to transact. The sovereignty being sought is not the agent's. It is the deployer's, laundered through the agent. And here is the collision that will define the next two years. If agents transact autonomously, they will also generate autonomously — research, analysis, trade rationale, content. The volume of AI-generated claims in the crypto market will not grow linearly; it will grow by orders of magnitude, because agents can generate without a human in the loop, at machine speed, around the clock. The empty template problem, which is today a curiosity about a broken pipeline, becomes the central problem of the entire information layer. When the majority of market-moving claims are generated by autonomous systems that no human has read, the question of verification stops being a nice-to-have and becomes the load-bearing wall. Without a way to verify agent-generated claims, the market will be running on content whose provenance is unknowable and whose truth is unverifiable, at a scale and speed that no human oversight can match. Truth emerges from the collision of opposites, and the collision between autonomous generation and human oversight is coming, and there is no plan for it. I want to give you a concrete way to think about what the fix looks like, because abstractions are cheap and I am suspicious of them. Imagine a claim — "Protocol X's token unlocks thirty percent of supply over the next twelve months." Today, that claim appears in a generated report with no attribution, and the reader either believes it or does not. In a verifiable-inference world, that claim carries a receipt. The receipt says: this claim was derived from this data source, which is this on-chain contract or this governance document, and here is a proof that the derivation is valid, and here is the identity of the model that performed it, and here is the timestamp. Any reader can check the receipt independently, without trusting the report, without re-running the model, without access to the model's weights. This is exactly the structure of a zero-knowledge proof applied to an inference step, and it is exactly the structure that crypto already knows how to build for other computations. The missing pieces are performance — making the proof cheap enough — and standardization — agreeing on a format for the receipt. Both are engineering problems with clear paths. Neither is solved. Both are worth billions. The reason I am confident this market will exist is that the demand is not optional. Once a significant fraction of market-moving information is machine-generated, the participants with capital at risk will pay to distinguish verified claims from unverified ones, because the cost of acting on a hallucination is asymmetric — a small edge from a true claim, a catastrophic loss from a false one. This is the same asymmetry that created the audit market, the credit-rating market, and the news-verification market before it. Every time information production scaled faster than information verification, a verification industry emerged to arbitrage the gap. The AI research boom is scaling information production faster than any previous technology in history. The verification industry that emerges to arbitrage this gap will be the defining infrastructure of the next cycle, and it will be built on crypto rails, because crypto is the only stack that has spent a decade solving the general problem of verifiable computation. Let me also puncture the most seductive version of the optimistic story, because it is the one that will get people hurt. The seductive story is that better models solve the problem — that as models get more capable, they will naturally become more truthful, more grounded, more reliable. This is backwards. Better models produce better-looking empty templates. The capacity to generate plausible structure scales with capability, and the capacity to verify structure does not scale with it automatically. A more capable model is a more dangerous source of confident, well-formatted, unverified claims, because its confabulations are more convincing. The empty template was produced by a relatively modest system and it was already convincing enough to fill four thousand words. Imagine the same failure mode at the frontier. The reliability of a model and the perceived reliability of a model are not the same quantity, and the gap between them grows with capability. This is the trap. The industry will deploy frontier models for research, be dazzled by the fluency, and lose the ability to tell when the fluency is covering for an empty input. This is where my Layer-2 skepticism becomes relevant in a way that surprises even me. I have argued for years that post-Dencun blob space will saturate within roughly two years, and that when it does, rollup gas fees will double again, because the cheap-data subsidy that made the current fee environment possible is a temporary condition, not a structural one. The reason this matters for AI research is that verifiable inference is going to need somewhere cheap to run and somewhere cheap to attest. The optimistic case is that rollups provide that cheap attestation layer — a place to anchor proofs and receipts at low cost. The realistic case is that the cheapness is temporary, and the verification infrastructure will be built on a fee base that doubles within a couple of years, which means the economics of verification have to work at the higher fee level, not the current one. Anyone building verifiable-inference infrastructure today and pricing it against current rollup fees is building on a subsidy that is going to evaporate. I have watched this movie before. The DeFi flywheel was priced against incentive subsidies too. There is a version of this essay that ends with a clean prescription, and I am going to resist it, because clean prescriptions are the empty template's favorite genre. What I will offer instead is a set of signals to watch, which is what I do for a living. Watch the latency and cost curves for zkML proofs on medium-sized models. Watch whether any major research operation starts attaching provenance receipts to its claims, and whether readers pay a premium for it. Watch whether the agent frameworks — the ones promising autonomous wallets and autonomous generation — ship any verification primitive at all, or whether they ship the generation and defer the verification, which is the pattern every crypto project follows until it is forced to do otherwise. Watch whether the regulatory fights over AI attestation start to mirror the jurisdictional contests over crypto licensing, because they will, and the outcome will determine who controls the verification layer. Watch the on-chain data for signs that capital is starting to price verified information differently from unverified information. That divergence, when it appears, is the trade. Chasing the horizon of the next paradigm, I keep returning to the same uncomfortable thought. The crypto industry's founding promise was that we could replace trust in institutions with verification of state. We spent fifteen years building the machinery for that — blockchains, proofs, oracles, attestations. And then we imported, without examination, a technology whose entire value proposition is that you do not verify it. You trust it. You trust the model, you trust the training, you trust the output. The empty template is the first crack in that trust, and it is a small one, and it is a warning. The next crack will be a market-moving analysis that was generated from nothing, published at scale, and acted upon by thousands of participants before anyone realized it was hollow. And when that happens, the industry will finally have to build the thing it should have built from the start: a way to prove that the analysis is real. The machinery exists. The demand is coming. The only question is whether we build the verification layer before or after the empty template costs someone everything. Following the signal through the noise floor, the signal is faint but it is there, and it says the same thing it has always said. The container is not the content. The format is not the truth. And the only thing that has ever been worth building is the mechanism that lets a stranger check your work without trusting you.