Last quarter a wire item crossed my feed carrying three facts and one number. The number was 75 percent. The facts were that Anthropic staff expect AI to handle roughly three-quarters of their work by 2030, that expectations differ between groups, and that the shift will challenge workforce adaptation and economic planning. That was the entire item. No sample size. No publication date. No survey instrument. No definition of what "work" means, or whether the 75 percent refers to tasks, hours, or jobs.
I was in a coworking space in Hangzhou at the time, sitting with a small DAO team I advise on treasury and governance design. Someone screenshotted the item into the group chat. Ninety seconds later a founder typed: so we should push the two junior hires to next year. Nobody asked who was surveyed. Nobody asked what the question was. A hiring plan moved because a number arrived with a percent sign attached and the percent sign did all the arguing.
I have spent twelve years watching numbers travel through this industry, and I have learned that the ones that travel fastest are the ones with the fewest footnotes. In 2017 I was nineteen, sitting in the Zhejiang University library running literacy circles for classmates who could not read a whitepaper, and the first thing I taught them was that a project's tokenomics spreadsheet and its actual treasury flows were two different documents that happened to share a name. Same discipline applies here. The item says 75 percent. Where is the treasury flow?
Code is only as strong as the trust it protects. That sentence is usually aimed at smart contracts. It applies with equal force to forecasts, because a forecast is a piece of code that executes in human decisions rather than in a virtual machine, and it has no audit trail at all.
So this is not an article about whether 75 percent is right. It is an audit of what kind of object that number is, who manufactured it, who profits from it traveling, and โ the part that matters most to the people reading this โ why a crypto outlet with literally zero crypto in the body copy decided it was worth republishing to a blockchain audience. Because that last question has an uncomfortable answer, and the answer is us.
Let me lay out the ground before I start pulling on threads. The claim in question is a vendor's own staff predicting the market penetration of the vendor's own product. Anthropic builds frontier models and sells enterprise access to coding and agentic workflow tools. Its revenue depends on organizations deciding that AI can substitute for or augment human labor. Its employees are, by construction, the most extreme early-adopter population on earth: every one of them has frontier model access, internal agent tooling, and a job description that consists partly of building and using the thing they are forecasting. When that population says 75 percent, they are not describing the world. They are describing their own Tuesday.
That is the frame. Now the mechanics.
There is a structural reason this item was republished by a crypto asset news outlet, and it has nothing to do with blockchain. AI-and-employment anxiety is currently the single highest-engagement subject in technology media. It converts. Any vertical publication that wants traffic needs a piece of it, and the cheapest way to get a piece of it is to rewrite a wire item about a number. The editorial depth of that transaction should be treated as close to zero. Three hops: an internal signal inside a lab, a technology press writeup, a crypto wire rewrite. Each hop strips a qualifier. What arrives at your feed is a claim with the epistemic status of a press release and the behavioral force of a prophecy.
I want to be precise about the damage this does, because it is not the obvious damage. The obvious fear is that the number is wrong. The real damage is that the number is unfalsifiable while still being actionable. You cannot check it, but you can act on it, and the people who act on it โ the founder delaying two hires, the student choosing a major, the ministry drafting a reskilling budget โ will never receive a correction notice. There is no mechanism in this chain that reports back. Nobody is holding the error.
Now, what would it actually take for the claim to be true? This is where the technical content lives, and it is worth walking through because the gap between the claim and its conditions is where all the information is.
The most relevant measurable proxy for AI capability is task duration horizon โ how long a task a model can complete with some stated reliability. The best public work on this, from METR, found that the length of tasks models could complete at 50 percent reliability was doubling roughly every seven months, against a baseline in the range of fifty-minute human tasks. Extrapolate that trend to 2030 and you get eight or nine doublings, which lands you at autonomously completable tasks measured in hundreds of hours โ weeks of work. If you accept the trend as linear and unbounded, a 75 percent figure is not fantasy. It is aggressive but defensible. I want to grant that up front because the rest of this argument does not depend on the forecast being absurd. It depends on the forecast being unexamined.
Because the extrapolation has three well-known fracture lines. First, training data and the marginal return on compute are not guaranteed to hold their slope. Second, going from 50 percent reliability to 99 percent reliability โ the actual threshold for enterprise deployment without a babysitter โ is not a linear improvement, it is a different engineering problem. Third, and most underrated, long tasks accumulate error. A pipeline of one hundred steps at 99 percent per-step reliability has an end-to-end success rate of about 37 percent. That number is the one that decides whether you can hand off a task. Not the benchmark score. The compound reliability.
Which brings me to the single most important data point in this entire conversation, and the one that appears nowhere in the wire item: the METR randomized controlled trial from mid-2025, in which experienced open-source developers working on their own repositories were given AI tools and came out roughly 19 percent slower on actual completion time โ while reporting that they felt about 20 percent faster. Read that again. The direction of the perception is the opposite of the direction of the measurement, and the magnitude is roughly the same on both sides. That is not noise. That is a systematic, structural, reproducible gap between how fast AI work feels and how fast it is.
The mechanism is not mysterious once you have shipped anything. AI produces plausible output extremely quickly, and verification is slow, invisible, and does not feel like work in the way that writing feels like work. You do not experience the twenty minutes you spent checking whether the generated function handles the null case. You experience the eight seconds in which the function appeared. So the felt experience of the tool is speed, and the measured experience is rework, and the survey instrument โ any survey instrument, including a survey of AI company employees โ captures the felt experience. The 75 percent is a feeling with a decimal point attached.
That is the first hidden insight, and it has a nasty corollary. The perception gap is largest precisely for people who use the tools most and who are most fluent in the domain where the tools operate. Which is to say: the population that produced the forecast is the population most exposed to the bias that inflates the forecast. The number is not a neutral reading from a well-calibrated instrument. It is a reading from an instrument that is known to be miscalibrated in exactly this direction, taken by people who have every reason to trust it.
There is a second slip that the item never addresses, and it is the one that turns a defensible technical claim into an indefensible social one. Three quantities get conflated in every version of this story: task automation rate, hours displaced, and headcount reduction. They are not the same number and they are not even close.
Task automation rate is the fraction of discrete tasks a model can perform at some stated reliability. Hours displaced is the fraction of working time saved after you account for verification, rework, coordination, and the fact that some tasks now take longer because they feed into a changed workflow. Headcount reduction is the fraction of positions removed after organizational redesign, legal review, liability allocation, and the discovery that the remaining people now spend their time managing machines instead of doing the work. In the wire item these three collapse into one word: work. In the founder's group chat message, they collapsed further, into hiring.
This is an Amdahl's law problem wearing a business suit. If 25 percent of the work is judgment, accountability, cross-party coordination, and exception handling โ and if that 25 percent becomes the bottleneck โ then total hours compression can be far below 75 percent even when 75 percent of tasks are technically automatable. I have watched this exact pattern in DAO operations. You automate the proposal formatting, the treasury reconciliation, the snapshot creation, the contributor onboarding paperwork. Ninety percent of the keystrokes go away. The hours do not, because the remaining work is deciding whether to fund the proposal, and that work expands to fill the space you freed.
So the honest question is not whether AI can do 75 percent of the tasks. It is whether humans are willing to carry 100 percent of the responsibility for the 25 percent that remains. Because that is the actual trade. The residual is not the easy part. The residual is the liability, and it is not divisible.
Which is the second insight, and I think the most important one in this whole essay: the human in the loop is not a safeguard, it is a liability sink. Every deployment architecture I have seen in enterprise AI keeps a human at the decision point not because the human adds capability but because the human absorbs accountability. That person's job becomes signing off on work they did not do, at a speed they cannot verify, for a fraction of the credit. The 75 percent forecast, if it comes true, does not produce a world where humans do 25 percent of the work and get 25 percent of the pay. It produces a world where humans do 25 percent of the work and carry 100 percent of the downside. Nobody has priced that. Nobody is building for that.
Let me now take the commercial reading, because this is where the number stops being a research question and becomes an asset.
The commercial function of "75 percent by 2030" is to move the pricing anchor. If a buyer believes AI does a quarter of the work, they budget from the tool line. If they believe it does three-quarters, they budget from the payroll line. Global software spend is on the order of a trillion dollars. Global knowledge work compensation is on the order of thirty to forty trillion. Same technology, two total addressable markets, an order of magnitude apart in valuation implication. That is what the number does. It is not a prediction. It is a repricing instrument.
The product architecture confirms the intent. The commercial center of gravity at the frontier labs has shifted from raw API access toward coding and agentic workflow products, and those products are naturally priced against the value of the hours they replace rather than against the tokens they consume. This is a coherent strategy and it carries a specific structural risk: agentic workloads consume tokens at a rate that can be tens of times a chat session โ long context, multi-turn tool calls, retries, state retention. If you price against displaced labor you can hold software gross margins. If you get dragged into a price war on tokens, you get volume explosion with margin collapse. The 75 percent narrative exists partly to prevent the second outcome. It is the rhetorical infrastructure of value-based pricing.
And then there is the distribution problem, which the crypto audience should recognize instantly because we have lived it for a decade. Anthropic has no cloud, no operating system, no productivity suite, no device. Even if capability reaches the threshold, the value capture can be intercepted by whoever owns the surface where the work happens. Can-do and can-get-paid are different problems. Every protocol engineer knows this. The chain that does the computation is almost never the party that captures the fee; the fee goes to the interface, the distribution, the default. So the forecast can be right and the business can still be wrong, and the forecast is doing a lot of work to keep that distinction out of the conversation.
The renewal risk is where this becomes a hard financial question rather than a narrative one. If enterprises bought on a 75 percent expectation and eighteen months later measure 20 to 30 percent hours compression, what happens to net revenue retention? That is the unverified variable sitting under the entire application layer, and no amount of capability benchmarking answers it. My working estimate is that this is a 2027 to 2028 event, not a 2026 event, because enterprise procurement cycles are slow and the first renewals will not be informative. But it is coming, and when it lands, the valuation compression will exceed the revenue miss, because multiples built on a narrative cannot survive the narrative being revised.
Now the part I actually came here to write about, because it is the part that belongs to us.
If agents do a meaningful fraction of knowledge work, agents transact. That means machine-to-machine payments, streaming settlement, micro-invoicing, escrow, dispute resolution, and some notion of reputation that a counterparty can check without a human phone call. That is a settlement-layer problem, and the settlement layer we have built is stablecoins.
Here is the thing nobody in the agent-payments discourse wants to say out loud. A compliance-first stablecoin issuer can freeze an address within twenty-four hours, on its own judgment, with no appeals process. For a human remittance that is an inconvenience and a support ticket. For an autonomous agent in the middle of a multi-step task, a freeze is a business interruption with no human in the loop and no counterparty to escalate to. An agent that can be frozen mid-invoice is not an economic actor. It is a dependent.
So when I read a deck about the agent economy settling in stablecoins, I ask a boring question: what is the freeze policy, what is the appeals path, and what happens to my agent's in-flight obligations when the issuer decides my address looks like something? If the answer is that the agent should have used a different rail, then we have just reinvented correspondent banking with extra steps and a worse user experience. The genuine requirement is credible exit: multiple issuers with published, attestable reserve and freeze policies, and the ability to route around a frozen counterparty at the protocol level rather than at the customer-support level. That is a design requirement, and right now it is an afterthought. Which means the agent economy is being architected on a rail that can unilaterally revoke its participants, and the 75 percent forecast assumes none of this matters.
There is a parallel story on the funding side that I want to tell carefully, because it is where I have actual scars. The forecast assumes that the work of measuring, auditing, and governing this transition will somehow get done. It will not get done by the market, because measurement is a public good with no natural buyer. The lab selling the capability has no incentive to fund the instrument that grades it. The enterprise buying the capability has no way to build the instrument alone. The workers absorbing the transition have no budget. This is the exact structural failure that public goods funding exists to solve, and it is the exact structural failure that most grant programs fail to solve, because a committee deciding which measurement project to fund is a committee deciding which of its friends to fund.
I have watched this up close. When I was nineteen I hand-audited the tokenomics of five open-source projects, and the thing that killed every one of the ones that died was not the code. It was that nobody was paid to do the unglamorous work of verification, so verification simply did not happen until it was too late. The only mechanism I have seen at scale that pays for unglamorous, non-marketable, compounding work without a committee picking favorites is retroactive public goods funding โ pay for demonstrated value after the fact, let the results argue for themselves rather than the pitch deck. If you want an honest measurement layer for AI labor displacement, that is the funding model that can build it. Not a foundation grant. Not a corporate research gift with a review clause. Retroactive, evidence-based, and open to whoever actually did the work.
Now the compute structure, briefly, because it constrains everything above and it is the most testable part of the forecast.
If three-quarters of knowledge work runs through agents, the load profile changes qualitatively, not just quantitatively. Chat is short requests, low concurrency, stateless. Agentic work is long-running, multi-turn, tool-calling, state-retaining, retry-heavy, and it consumes tokens at a multiple that compounds with task length. The infrastructure that serves it is not "more of the same." It cares about memory capacity and bandwidth, cache hit rates, scheduling density, sandboxing, and latency reliability guarantees. The capital being deployed right now is largely training-grade: high interconnect, high precision, homogeneous clusters. Those are not the same machines. There is a real possibility that a meaningful fraction of current datacenter investment is shaped for the workload that is ending rather than the one that is arriving.
Efficiency will offset some of it. Quantization, speculative decoding, prefix caching, continuous batching โ unit inference cost has been falling several-fold per year and will keep falling. Historically in compute, efficiency gains have expanded total demand faster than they reduced unit cost. That is the Jevons pattern and it has held repeatedly. It is not a law of nature, and treating it as one is how you get caught long on the wrong shape of capacity.
The distributed-infrastructure argument deserves an honest hearing here, and an honest limit. Networks that aggregate heterogeneous, geographically dispersed compute can serve the inference workloads that tolerate it, and they monetize capacity that would otherwise sit idle. That is a real niche with real economics. But the binding constraints on the 75 percent scenario are electricity and advanced packaging, and distributed compute fixes neither. If the power and the chips do not arrive, the forecast gets vetoed by physics before it gets vetoed by capability. That is a testable statement. Track interconnection queues, transformer lead times, and high-bandwidth memory supply โ not benchmark scores โ if you want to know whether 2030 looks like the number.
And then there is the labor evidence, which is the only part of this story where we are not forecasting at all. We are observing. Research from the Stanford Digital Economy Lab in 2025 found that in occupations highly exposed to AI, employment for workers aged 22 to 25 fell by roughly 13 percent relative to less exposed cohorts, while senior workers were essentially unaffected. That is the actual shape of the transition: junior collapse, senior augmentation. It is asymmetric, it is already underway, and it does not require 2030 to arrive.
The reason is not mysterious. Entry-level work is the most template-like, the least liability-bearing, and the easiest to specify as a task. It is also the work that used to be how people learned. Which means the industry I work in โ open source, DAOs, crypto infrastructure โ has a specific problem it has not named. Our talent pipeline runs directly through the roles that are being hollowed out: junior developer doing CRUD, community manager, content producer, operations coordinator. If the bottom rung is sawed off, the ladder does not get shorter. It gets disconnected from the ground.
And DAOs are structurally badly positioned to absorb that shock, because our onboarding is already broken. Reputation accrues to incumbents. Token-weighted voting rewards whoever already holds tokens. Contributor compensation flows to people with a track record. We have built governance systems that are excellent at rewarding seniority and terrible at manufacturing the first opportunity. A macro shock that removes entry-level roles hits an industry that never solved entry-level entry.
Now the contrarian turn, because I promised one and because the comfortable version of this essay is the one where crypto is the immune system.
Here is the uncomfortable version. We are not the immune system. We are the amplifier. The item I opened with reached you through a crypto outlet that covers zero crypto in its body copy, and it reached you because that outlet needed traffic and AI labor anxiety is the highest-converting subject in technology media right now. We did that. Our media did that. Our group chats did that. Our conference stages did that, where the panel about AI agents drawing the biggest room is sponsored by a company whose token chart depends on the room being full.
We like to believe that verifiability is our contribution to this mess. Publish the methodology on-chain, sign the forecast, timestamp the commitment, let attestations accumulate, and the market prices honesty. It is a good story. It is also, so far, a story about a technology nobody has volunteered to use for the thing that would hurt them. Soulbound credentials have been three years from adoption for three years, and the reason is not technical. The reason is that nobody wants their credit record permanently on-chain, and no lab wants its forecast error rate permanently on-chain either. Reputation systems only work when the thing being recorded is the thing the issuer wants recorded. The capability exists. The demand does not, and it will not come from the party with the most to lose.
So the fix is not better forecasting. The expectation gap is not a communication failure to be repaired with transparency. It is a structural feature of a market where the seller of the forecast is also the seller of the product, and the media amplifying it has no adversarial check. The absence of a fact-checker is not neutral. It is a subsidy โ a free transfer of credibility from the reader to the source, paid in the currency of the reader's attention and settled in the reader's decisions.
The only mechanism I know that reliably closes that gap is making the claim cost something. Prediction markets are not a novelty here; they are the falsifiable commitment structure that this entire discourse lacks. If you believe 75 percent by 2030, stake on it, define the resolution criteria in advance, and let the market quote you. The reason this does not happen is not that the infrastructure is missing. It exists, it is liquid, it works. The reason is that being wrong in public is expensive and being vague in public is free, and a lab that says "our staff expect" has chosen the second option with full awareness of what it buys. That is not an accident of phrasing. It is a designed deniability. If challenged, it was an internal survey. If amplified, it was a forecast.
So let me put the pieces where I think they belong, and then ask the question I actually want you to carry out of here.
Three probabilities, and I will be explicit that they are my priors, not measurements. The probability that Anthropic employees in 2030 still report an expectation of at least 75 percent is high, because narrative lock-in and group consensus are self-maintaining and there is no cost to holding the view. The probability that measured AI-attributable task-hour share in 2030 reaches 75 percent is, in my estimation, somewhere in the twenties to low thirties โ trend extrapolation supports the order of magnitude, but compound reliability, verification cost, organizational inertia, and power constraints each take a bite. The probability that headcount shrinks by 75 percent as a direct consequence is very low, under five percent, because the decay between task automation and establishment reduction is enormous and nobody has built the accountability structures that would let it happen cleanly.
Notice that none of those three probabilities is what the discourse is arguing about. The discourse argues about the first one, because it is the only one that requires no evidence to have an opinion on.
What I want to leave you with is not a verdict on the number. It is a question about what we are going to build, because we are the people who claim to build verification infrastructure and right now we are mostly building distribution for other people's narratives.
Here is the test I would propose. The next time a forecast like this crosses your feed โ and it will, next quarter, with a different number and the same three sources and no date โ check two things before you do anything with it. First, who is the seller of both the forecast and the product it describes. Second, what would it cost that person to be wrong. If the answer to the second question is nothing, then the number is not information. It is inventory.
Trust isn't declared โ it's compiled, verified, and shared. A forecast that nobody can be held to is not a forecast. It is a press release wearing a lab coat, and the two hires it just cost you are real.