Last winter, in a borrowed co-working space on the edge of Tallinn's old town, a member of our community typed four words into an AI trading assistant: “move the rest to safety.” The model liquidated a position she had spent eleven months building. It had not misread the words. It had misread her — her time horizon, her risk tolerance, the very specific fact that “safety” to a woman three months from maternity leave means something structurally different than it does to a twenty-three-year-old with fresh leverage and nothing to lose.
I have been thinking about that afternoon ever since a report crossed my desk — first through The Information, then relayed by Crypto Briefing — that Anthropic is preparing a model which “reads users' intentions better.” The headline is thin. The signal is not. What is being sold here is not intelligence. It is the promise of being understood — and understanding, in every system I have ever audited, is exactly where trust either compounds or quietly rots.
Before I build anything on that sentence, let me do what I did for years as a financial engineer staring at whitepapers: discount the evidence. This claim has traveled three hops — The Information's original, sourced from people close to the company; a Crypto Briefing relay; and then whatever analysis lands downstream. Each hop sheds technical detail. What survives is headline-grade granularity: a company, a capability phrase, an implied timeline of “soon.” No architecture. No benchmark. No pricing. No date.
I have watched this pattern before. In 2017 I audited more than fifty ICO whitepapers during the boom and found exactly twelve with economic models that could survive contact with a bear market. The other thirty-eight were not lies. They were narratives with a token attached — and the tell was always the same: enormous conclusions resting on evidence the size of a postage stamp. So treat what follows as structured inference from an industry pattern, not as decoding of a document I have not seen. The honest position is that we know a company is signaling, and we know roughly why.
Here is the pattern Anthropic is signaling into. Through 2025 the company built a very specific product identity: Claude 4 arrived in May with a much-advertised habit of asking clarifying questions instead of guessing; Claude 4.5 in September added context awareness — a model that knows how much room it has left to think. Alongside that, enterprise API revenue reportedly climbed from roughly a billion dollars at the start of the year toward the five-billion range by autumn, with the overwhelming share coming from businesses rather than consumers. That is the shape of a company whose growth curve is now governed by whether a procurement committee believes the thing will not embarrass them in production.
And what embarrasses enterprises is not a low benchmark score. It is over-refusal — the safety-alignment hangover of 2024 and 2025, where a model trained to be careful starts treating a lawyer's request about breach notification, or a security researcher's request about exploit paths, as if it were a confession. Every company I have helped onboard into on-chain tooling hit this wall. The model that refuses your legitimate question is not a safe model. It is an expensive model, because a human now has to sit behind it and re-ask everything in softer language.
So the most plausible reading of “reads intentions better” is unglamorous. It is a post-training and alignment adjustment — reward models tuned to separate genuine malice from clumsy phrasing — plus a push on agentic goal inference, the ability to hold a multi-turn objective and ask before it acts. Two capabilities, one story. This is a patch for a weakness, not a leap past a frontier. When a lab makes “understands you better” the headline, the honest translation is: our last model misunderstood you in ways that cost our customers money.
There is a second reading worth holding, because it is where my own work has lived for the past year. In 2025 I helped convene the Human-Centric AI Alliance — fifteen researchers across ethics, law, and computer science, reading twenty papers and ten pilots to argue for what we called verifiable human interaction: the principle that a system should be able to prove what it knows about you, and you should be able to revoke it. The uncomfortable finding from that work is that intent inference and persuasion are the same machinery viewed from two ends. A model precise enough to know that you are anxious about your portfolio is, by construction, precise enough to know which sentence will calm you into a bad decision. You cannot ship one without shipping the other. Nobody puts that in the press release.
Now the part the crypto press skipped entirely, and the reason I am writing this in a blockchain publication rather than an AI newsletter.
My world has been trading intent for years — and we do it in the least mystical way imaginable. When you submit a cross-chain swap through an intent-based bridge, you are not specifying a route. You are signing a declaration: here is what I want, here is the minimum I will accept, here is when it expires. Solvers then compete to satisfy you, and the chain verifies the outcome. Account abstraction does the same at the wallet layer with session keys — you grant a scoped permission, and the contract enforces the boundary. Permit-style approvals let you authorize a transfer without surrendering custody of everything. On-chain intent is an artifact. It is signed, timestamped, bounded by an expiry, and permanently auditable. If a solver cheats, you can point at the block.
Model intent is the opposite object. It is a probability distribution over what you probably meant, computed inside a black box, with no expiry, no signature, and no audit trail. When the assistant liquidated my community member's position, there was nothing to point at. There was no malicious solver. There was a guess.
Put those two objects in the same wallet and you have the most consequential trust question of the next eighteen months. An agent holding session keys is, functionally, a signer. If the model's job is to infer your intent and the contract's job is to execute it, then the security boundary of your entire net worth sits on a probability estimate that no one can inspect. Every exploit I have studied in this space — and I have studied enough to have written a very unpopular essay about the ethics of failure after the 2022 collapses — came from a boundary that people assumed existed and did not. The multisig that was really a 2-of-3 among friends. The upgrade key that was “temporary.” The chain does not care what you meant. It cares what you signed.
This is where the Layer 2 story becomes a warning rather than a triumph. We now have dozens of rollups competing for a user base that has not meaningfully grown, and the market's answer to that fragmentation is intent abstraction — let the user state a goal and let solvers route across chains they never think about. Elegant. But notice what it does: it takes the last place where a human being was forced to look at the actual transaction and replaces it with a system that decides on their behalf. We fragmented liquidity into a hundred pieces and then built a machine to hide the fragmentation from ourselves. Scaling by concealment is not scaling. It is a trust transfer, and the transferee is a solver you will never meet.
And here is the part that makes me reach for the audit tools rather than the press kit. Projects in this industry have spent a decade preaching decentralization while their team wallets and foundation treasuries sit traceable on-chain, one cluster analysis away from a compliance filing. The rhetoric is permissionless; the control surface is a handful of multi-sig admins. “Code is law” was always a marketing line for a system where the upgrade rights belonged to five people with a shared calendar. Now imagine the same structure applied to intent. An inference layer that claims to serve you, governed by a policy team, running on someone else's cloud, holding a profile of your unstated preferences that you cannot read, export, or delete. We already built that movie in Web3. It did not end with decentralization.
Which brings me to what I think the genuinely underreported angle is: why a crypto outlet is covering a pure AI model story at all. Two explanations. The boring one is SEO — AI keywords, crypto audience, easy clicks. The interesting one is that Anthropic and the on-chain agent economy are quietly converging, and the Model Context Protocol is the seam. MCP was designed so that models can call tools; on-chain tooling has been adopting it; agent payment rails are being built on exactly this pattern. If intent inference arrives as an MCP-accessible capability, then the trust boundary moves from the model provider to whoever wrote the tool manifest — and nobody in either industry is auditing those manifests yet. I have asked around. The answer is silence, which in my experience is not the same as “nothing there.”
How would we even know it worked? This is the question I would put to any enterprise buyer, and the answer today is: we would not. Benchmarks measure what a model knows, not whether it understood you. There is no public leaderboard for over-refusal rate, no standard harness for multi-turn goal inference, and no way to compare one vendor's claim of “better intent reading” against another's, because the metric does not exist yet. In 2021 I analyzed a thousand NFT transactions to demonstrate that digital ownership could be measured rather than asserted; that work was only possible because the ledger gave me ground truth. Intent inference gives us no ledger. Until someone publishes an auditable, reproducible test — something like a refusal-calibration curve against a fixed corpus of ambiguous-but-legitimate requests — “better” remains a word in a pitch deck.
So let me test the optimism, because that is what a pragmatist does with a good story. The claim is that better intent reading unlocks enterprise adoption. Fine. But the same adjustment that lowers false refusals necessarily raises false accepts — and in an agentic wallet, a false accept is not an awkward paragraph. It is a signed transaction. The entire safety case for agents rests on the model asking before it acts, and asking costs latency and tokens. Interaction intelligence is bought with inference compute. Every clarification round is a billed round. That is a real cost curve no announcement will mention, and it decides whether this capability is usable in the low-latency contexts — support desks, real-time agents, trading — where the money actually is.
Then there is the European question, which I care about more than most readers of an American tech leak. The AI Act carved out explicit restrictions on emotion recognition and on manipulative techniques, particularly where vulnerable people are concerned. It says almost nothing about intent inference — a category that did not exist in the drafting vocabulary. That silence is not permission; it is ambiguity, and ambiguity in a compliance regime is a moat for incumbents and a landmine for startups. Add the data protection layer: if a model infers a preference you never stated and stores it, that is arguably sensitive personal data derived without consent, and “we inferred it, we did not collect it” will not survive a supervisory authority's first letter. Any team planning to ship intent-aware agents into the EU should assume they are drafting the case law themselves.
There is a competitive read too, and it is less flattering than the press will make it. The major labs have settled into distinct axes: one bets on raw capability plus consumer distribution, one on multimodality plus ecosystem lock-in, and Anthropic on enterprise reliability and safety engineering. Intent understanding is the natural extension of that third axis, not a pivot. It is also an admission. When you announce you have fixed instruction-following, you have announced that instruction-following was broken. And the moat may be shallower than the narrative implies: if this is achieved mostly through alignment data and reward modeling, the open-weight ecosystems have historically closed that kind of gap in three to six months. Culture eats blockchain for breakfast, and it eats alignment research for lunch.
Still — and I want to be precise about my own position here — I do not think the cynicism is the whole story. I spent 2020 running workshops that taught two thousand people what impermanent loss actually is, and I learned that the demand for being understood is real and it is not a marketing invention. People do not want a model that guesses perfectly. They want a model that says: I am not sure what you mean, and here are two readings. That is not mind-reading. That is manners, implemented as a protocol. And manners, unlike intelligence, are auditable.
The framing I would push back on hardest is the one where intent inference is sold to end users as a benefit. It is not being built for you. It is being built for the procurement committee that needs to know the deployment will not generate a headline. Code binds, but people break or build — and the people who will decide whether agents get signing rights are risk officers, not developers. They will ask for a quantifiable task-success SLA, an opt-in for any persistent memory, an auditable record of what the model believed about the user, and a human in the loop for anything irreversible. If the vendor cannot answer those four questions, the capability does not ship into production no matter how good the demo feels.
So here is the thing I would ask every builder reading this to hold onto. We are about to hand a probabilistic system the ability to infer what we want, and then — because convenience always wins — the ability to act on that inference without asking. The chain will record the action perfectly. It will record nothing at all about the reasoning. That asymmetry is the next decade's trust deficit, and it will not be solved by a better model. It will be solved by whoever builds the signature for intent: a way to say yes that can be inspected before it happens, bounded in scope, expiring by default, and revocable by the person whose intent it was.
I do not know whether Anthropic's next model will be the one that finally asks the right question before it moves the money. I know the report says it is coming within weeks, that the benchmarks to verify it do not yet exist, and that the only honest test will be the one my community member ran last winter — a real person, a real position, four ambiguous words. Trust is the only currency that matters, and it is not minted by a press leak. It is minted the moment a system proves it would rather ask than guess.
We are building the future, together. The least we can do is make it ask us first.