The Gist: Replacing people is the least imaginative thing you can do with the most powerful technology of our lifetime. The AI industry is betting the farm on it anyway: point a swarm of non-deterministic agents at the average worker, call speed success, and hope nobody checks the bill. Somebody should. The biggest labs have only just scraped into adjusted profit, more reasoning can make a model less accurate, and most enterprises still cannot find the return. The replacement thesis has crypto’s fingerprints all over it: concentrated gains, collapsing trust, doubtful durability. Every revolution worth the name, from the railroad to the internet, paid off by making people more capable, not by making them unnecessary. Coding agents already turn great engineers into small armies. Nobody has done that for the analyst, the planner, or the scientist. The right context layer can. Next week at Oracle AI World, we stop arguing and start showing.
1. Replacement is the least imaginative use of AI, and the numbers don’t back it. The biggest labs have only just reached adjusted profit, tokens per task keep climbing, more reasoning can make models less accurate, and 95% of organizations report zero return.
2. Every revolution that mattered made people more capable, not unnecessary. The railroad, the dynamo and the internet paid off when work was redesigned around the new tool, not when the worker was swapped out. Brute force finds answers. It doesn’t create insight.
3/ The prize is enablement. Coding agents already turn great engineers into small armies because code comes with built-in context. Give information workers the same, through a governed context layer with real instruments and receipts, and everyone in the enterprise gets that boost.
In April 2014, Mark Zuckerberg walked onstage at Facebook’s F8 developer conference and retired the most famous motto in Silicon Valley. “Move fast and break things” became “move fast with stable infra.”
His reason was refreshingly honest: breaking things, it turned out, “wasn’t helping us move faster.” Fixing them ate more time than the speed had bought. The company that made velocity a religion figured out, a decade in, that speed is a tactic. It is not a strategy. Twelve years later, the AI industry has quietly brought the old motto back, with seemingly one word changed:
Move fast and replace people.
Nobody puts that version of the motto on a keynote slide. The industry is remarkably shy about saying it out loud. But read the success stories closely and it is the undercurrent of nearly every one. Klarna launched its AI assistant in 2024 by announcing it was doing the equivalent work of 700 full-time workers, worth a projected $40 million in profit. In September 2025, Marc Benioff explained that Salesforce had cut about 4,000 support roles because it “needs less heads.” The headline says productivity. The subtext says headcount.
A brief aside, because the public has noticed. Data centers have become a bipartisan punching bag: Bernie Sanders and Alexandria Ocasio-Cortez want a federal moratorium to halt a race “to eliminate hundreds of millions of jobs,” Texas Democrat James Talarico says “we can’t afford to waste water” on them, and Florida Republican Byron Donalds doesn’t want one “outside of my front door.”
Local opposition blocked or delayed about $130 billion of projects in the first quarter of 2026 alone, and when an industry’s own pitch is replacement, it should not be shocked that voters hear “they’re taking our water and our jobs.” (I have argued that this is a zero-sum argument about a non-zero-sum problem. The industry is not helping.)
I have spent the better part of a year watching the smartest industry on earth pursue one idea with the focus of a man looking for his car keys. Take a thing a human does, at roughly average skill. Replace it with a swarm of non-deterministic agents. Then measure success by how much got done and how fast. Rarely by what it cost. Almost never by whether anybody learned anything new.
Next week I’ll be at Oracle AI World in Las Vegas. We have things to show you, and I’m not going to front-run them here. But I want to set the table, because what we’re building only makes sense if you believe something that is currently out of fashion:
The point of AI is not to do what we already do, cheaper.
It is to let us do what we could not do at all.
So let me make the case. I’m going to be fair to the other side, I’m going to show my work, and then I’m going to tell you what I think we ought to be building instead.
More, faster: the oldest promise in the book
Let me start by conceding the accelerationists’ best point, because it is a good one.
Every great industrial revolution has let humanity do more in less time. In 1830, getting from New York to Chicago took about three weeks; by 1857 the railroad had cut it to about two days. The railroads moved so fast they broke time itself. With an estimated 80 local time standards in use by the early 1880s, railroad clocks across the country were reset to a new standard at noon on November 18, 1883. Congress did not get around to making it law until 1918. When the personal computer and the internet finally showed up in the numbers, U.S. productivity growth, which had run about 1.5 percent a year since the early 1970s, more than doubled between 1995 and 2003. Speed compounds. A faster economy is a richer economy, and I am not here to argue otherwise.
Richer for whom, though? Two weeks before the midterms, that is the question. Workers’ share of national income has fallen to 52.8 percent, the lowest since the government began counting in 1947, even as corporate profit margins hit a record.
An NBC News poll found 63 percent of Americans expect AI to reduce overall employment, and neither party has made the sale: on AI policy, 51 percent of registered voters say they agree with neither. The faster economy has been very good to capital. Whether it is reaching the middle class is genuinely contested, and an industry whose loudest pitch is needing fewer people should not be surprised when voters are skeptical.
But notice what the speed was for. In 1840, about 63 percent of American workers labored in agriculture. In 2022, direct on-farm employment was about 1.2 percent. The railroad and the engine did not just move wheat faster. They moved the grandchildren of the people who grew it into classrooms, offices, and laboratories. The PC and the internet then took that intellectual work and wired it to the entire planet.
Each revolution pushed us up, from muscle toward mind.
And here is the part the replacement crowd keeps skipping: the gains never came from one-for-one substitution. In 1987 Robert Solow complained that “you can see the computer age everywhere but in the productivity statistics.” Paul David answered him with history. Electric motors were less than 5 percent of factory mechanical drive in 1899, and only slightly more than half by the early 1920s, four decades after the first central power station opened. Factories that simply swapped a dynamo in for the steam engine got very little for it. The payoff came when engineers stopped replacing the old power source and redesigned the factory around the new one: single-story, linear layouts, a motor at every machine. Then the productivity statistics moved.
Substitution is the least interesting thing you can do with a general-purpose technology. It was true of the dynamo. It is true of the model.
Brute force is not a breakthrough
Now let me give the other side its due, because it has earned it.
AI is doing real science. Half of the 2024 Nobel Prize in Chemistry went to Demis Hassabis and John Jumper for AlphaFold, which has been used by more than two million people in 190 countries. A Google DeepMind system earned an officially graded gold-medal score at the 2025 International Mathematical Olympiad. And in May, an OpenAI model disproved Erdős’s 1946 conjecture on the planar unit distance problem, a result Tim Gowers called “a milestone in AI mathematics.” I am not going to pretend that is nothing. It is remarkable.
But look at how most of the progress is being made. DeepMind’s AlphaEvolve “uses an evolutionary framework to improve upon the most promising ideas”: generate, test, keep the winners, repeat. Its AlphaProof Nexus agent resolved 9 of 353 open Erdős problems “at the per-problem cost of a few hundred dollars.” When another team swept 700 Erdős problems marked open, it concluded that for the ones it cracked, the open status was “through obscurity rather than difficulty.”
Terence Tao, who knows a thing or two about this, put it plainly. The tools are “becoming capable enough to pick off the lowest hanging fruit,” and there is “an empirical tradeoff between the level of AI involvement in the solution, and the difficulty or novelty of that solution.”
On Epoch’s FrontierMath: Erdős benchmark, 68 significant open problems with $300 of compute apiece, the best model at launch solved two. And after its third round last week, the First Proof team raised a worry that deserves more airtime than it is getting: that “access to computational resources may become a barrier for basic research in mathematics.” It added, almost in passing: “Exposition remains generally weak.”
Exposition. That is mathematician for explaining what it means.
Asking a thousand agents to grind out a thousand attempts at a proof is impressive. It is also, mostly, a very expensive search. It finds answers inside a space we already know how to describe. It does not, by itself, make the space any bigger.
Which is why I think the next frontier is not the search. It is the search space itself: how a problem is described, what the model is told the world contains, which questions it is even able to ask. Improve the space, and every search inside it gets cheaper, more accurate, and more likely to surprise you.
And making the space bigger is where the real returns have always been. On June 10, 1854, Bernhard Riemann gave a lecture on the hypotheses that underlie geometry.
It was pure mathematics, and it waited the better part of sixty years for its moment. In 1912, wrestling with gravity, Albert Einstein wrote that he had “gained enormous respect for mathematics, whose more subtle parts I considered until now, in my ignorance, as pure luxury.” Riemann’s luxury became the language of general relativity.
In 1961, Murray Gell-Mann used the representation theory of a unitary group to organize the particle zoo into the Eightfold Way, and predicted a particle nobody had seen. Brookhaven found the omega-minus in 1964. Eugene Wigner called the fit between mathematics and physics “a wonderful gift which we neither understand nor deserve.”
None of those were faster answers to existing questions. They were new languages that made new questions askable.
And we need new questions. Nicholas Bloom, Charles Jones, John Van Reenen, and Michael Webb found that U.S. research productivity falls by half every 13 years, and that sustaining Moore’s Law now takes more than 18 times as many researchers as it did in the early 1970s.
Ideas are getting harder to find.
The binding constraint on progress is not throughput. It is insight, and you do not cure an insight shortage with more brute force.
Which brings me to my own job. My challenge at Oracle is not to produce the right answer at the right time, faster than an analyst could. Any decent agent can be made to do that eventually, at some price.
My challenge is to help an organization reach an insight it could not reach before: to make capability, not just efficiency, a bedrock property of the platform.
The arithmetic nobody wants to do
The economics of latency are settled. Every morning I marshal the output of dozens of engineers from my keyboard, and a sprint becomes an afternoon. The economics of cheaper are not, and this industry has been strangely incurious about them.
The sellers are subsidizing you. OpenAI reportedly posted a $12.3 billion operating loss on $6.7 billion of revenue in the second quarter, $1.84 out for every dollar in, and forecasts $278 billion of negative free cash flow through 2030. Anthropic lost more than $8 billion operating on $4.6 billion of 2025 revenue before telling investors it turned an adjusted profit this year. One lab’s adjusted quarters are a hopeful data point, not a law of economics. As the space race for model intelligence continues, labs will continue to show a loss and unsustainable economics.
The meter runs faster than the price falls. Per-token prices drop roughly tenfold a year, but reasoning models now consume more than half of all tokens and the average prompt has quadrupled. GitHub called its old pricing “no longer sustainable” for agents. Anthropic capped its $200 plan after one user burned through tens of thousands of dollars. An Nvidia vice president says his team’s compute “is far beyond the costs of the employees.”
A price set below cost is not a market price. It is a promotional period.
More horsepower, less accuracy. Researchers have built tasks where longer reasoning makes models worse, because they “get occupied by the distractors.” OpenAI’s o3 hallucinated on 33 percent of PersonQA questions, double o1’s rate. A 2026 study found that past a moderate budget, “marginal returns diminish substantially.” That is a V12 that will not drive in a straight line, and every extra token lands on somebody’s bill. And the problem exists on today’s models.
The buyers cannot find the return. MIT’s NANDA team found 95 percent of organizations getting zero return on $30 to $40 billion of generative AI spending. Gartner expects over 40 percent of agentic projects to be canceled by the end of 2027. On the Remote Labor Index, the best agent automated 2.5 percent of real freelance projects. Even measuring is slippery: METR found experienced developers 19 percent slower with AI while believing they were faster, and its 2026 rerun produced “only very weak evidence” because so many developers refused to work without it. We have built something people will not give up and still cannot price.
Substitution keeps backfiring. Klarna’s CEO admitted cost had been “a too predominant evaluation factor” and quality suffered. The Commonwealth Bank of Australia reversed 45 job cuts it had handed to a voice bot. Forrester expects half of AI-attributed layoffs to be quietly reversed. Daron Acemoglu puts AI’s ten-year productivity boost at no more than 0.66 percent, and he and Pascual Restrepo warn of “so-so technologies” that displace workers without making anything much more productive. Their textbook example: automated customer service.
So let me cross-examine the proposition, the way I would want someone to cross-examine mine.
Can you run an entire business on AI? Yes.
Can it scale revenue to remarkable heights? Absolutely.
Can it be a profitability machine once the tokens stop being subsidized? I don’t know, and neither does anyone selling it to you.
Can it be a counterparty that human beings trust with their health, their money, and their families? Not today.
Do its customers produce durable output, the kind of value that cascades through the rest of the economy?
That is the question I most want answered, and the one I hear asked least.
We have seen this movie
There is a crossover I cannot stop noticing, between this moment in AI and the last decade of crypto.
Crypto made a segment of the population spectacularly rich. It made a lot of other people poorer, which is what tends to happen in a game far closer to zero-sum than its promoters admit. Researchers at the Bank for International Settlements estimated that 73 to 81 percent of retail investors had likely lost money on their initial bitcoin investment. The model behind that estimate is crude. The direction is not.
Did people come to trust it? In 2024, 63 percent of Americans told Pew they had little or no confidence that the ways to invest in, trade, or use crypto were reliable and safe. The Federal Reserve found that in 2025 just 2 percent of U.S. adults used crypto to buy something, pay someone, or send money. I still pay in dollars, yen, and euros. So do you.
Did it produce durable new output? A year after its record high, bitcoin is down about a third, having been down by more than half in June. Seventeen years in, I am still waiting for the renaissance. Crypto failed less as a technology than as an inspiration. It did not send a generation off to build something truly new - again, it just tried to replace the dollar in most ventures. Worse yet, it sent them off to trade - where the only motivation was to generate money.
Now hold that up against an AI industry whose dominant pitch is replacement. Gains concentrated with whoever owns the cheapest-looking substitute. Trust heading the wrong way: 52 percent of Americans are now more concerned than excited about AI, the highest share Pew has recorded, against 9 percent more excited. Worldwide, 54 percent of people say they are wary of trusting AI systems. And durable output we keep promising and keep struggling to find on the income statement.
You see the analogy? Sure - we solved some interesting math problems and science challenges. But has it changed an information worker into a superior information worker? Not universally. And durably solving that problem is essential to the success and trustability of AI.
AI is not crypto. The capability is real, and I use it every day. But the replacement thesis carries the same three weaknesses (concentration, distrust, and doubtful durability), and we ought to notice before the market notices for us.
Build a new muscle
The people who invented this field knew which game they were playing.
In 1960, J.C.R. Licklider imagined humans and computers “coupled together very tightly,” in a partnership that would “think as no human brain has ever thought.” Humans would “set the goals, formulate the hypotheses, determine the criteria, and perform the evaluations.” Two years later, Douglas Engelbart defined his life’s work as augmenting human intellect: “increasing the capability of a man to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems.” Steve Jobs remembered a Scientific American chart in which a human on a bicycle “blew the condor away” in efficiency of locomotion, and called the computer “the equivalent of a bicycle for our minds.”
Not a replacement for the rider. A bicycle. The velocity argument is interesting. But it is velocity of outcome (not intermediate work product) that most customers will measure.
Are information workers getting promoted and employed at a faster rate because of their use of AI? (I have seen the actual opposite, where folks are almost doubted due to their use of AI to accomplish tasks)
Is there a new job family they can truly get trained for and change careers into? (Please don’t harass me with the useless ML training programs from universities, folks)
Erik Brynjolfsson calls the alternative the Turing Trap. “When AI augments human capabilities, enabling people to do things they never could before, then humans and machines are complements.” When it merely mimics them, the value flows to whoever owns the mimic. And he argues we have built “excess incentives for automation rather than augmentation.”
The evidence on which path pays is not subtle. In a study of 5,179 customer-support agents, generative AI raised productivity 15 percent on average and about 30 percent for the least experienced and lowest-skilled workers, while the most experienced saw small gains. The tool was, in effect, handing the know-how of the best agents to everyone else. At BCG, consultants using AI completed 12.2 percent more tasks, 25.1 percent faster. But on a task deliberately placed outside the model’s competence, they were 19 percentage points less likely to get it right. Augmentation lifts the floor. Blind substitution walks off the jagged edge. David Autor names the prize:
But notice the striking contrast with my own world. Hand a coding agent to a seasoned distributed-systems engineer and you see something close to the opposite: the experts pull away.
A study of AI-written code on GitHub, published in Science this year, found that newer developers lean on AI for more of their code, yet the productivity gains come exclusively from experienced developers. “Beginners hardly benefit at all,” its lead author said. In a Fastly survey, 32 percent of senior developers said more than half the code they ship is AI-generated, against 13 percent of juniors.
The evidence is not unanimous: Microsoft’s study of its own early-2026 rollout of command-line coding agents found a C-shaped pattern, with the newest and the longest-tenured engineers gaining most. But nobody has yet shown that expert surge for information workers. In knowledge work, the measured gains still flow mostly to the bottom of the distribution; in one experiment with 453 professionals, ChatGPT cut writing time by 40 percent and narrowed the gap between workers rather than widening it.
Why the difference? Because the software engineer already works inside a complete, verifiable context.
The programming language is a composable grammar: small, precise pieces that snap together into larger ones, and that a machine can check. The comments are enrichment: the human explanation of why the code is the way it is, sitting right next to it. The repository is the ontology. The type system and the tests are the guardrails. The compiler and the CI pipeline are the receipts. An expert can hand an agent a well-described space and verify every step it takes.
The analyst, the planner, and the claims adjuster get none of that.
There is no grammar for a business question, so every one arrives as free text. There are no comments on the data, so its meaning lives in somebody’s head. Their context is scattered across forty systems and three decades of tribal knowledge, and there is no test suite for a business answer.
Give them the same scaffolding (a structured language with a composable grammar for asking questions of the business, enriched with what its people know) and I would bet the same expert surge shows up.
That is the bet we are making.
And next week in Las Vegas we are not going to describe it on a slide. We are going to show you an entire engine, and the language that drives it, running live.
The labor data is starting to show the fork in the road. Stanford’s Digital Economy Lab finds no evidence of widespread job displacement, but employment for 22- to 25-year-olds in AI-exposed jobs is 19 percent below where it would have been, and the declines are “concentrated in occupations where AI usage primarily substitutes for human tasks.” Where AI complements workers, employment is flat or rising. In Anthropic’s January data, roughly three-quarters of API traffic, where businesses build, showed automation patterns; its March report found that share falling sharply. We are choosing which economy we get, one deployment at a time.
So here is what I mean by leveling up. AI gets interesting when it makes mathematics that was out of reach for most people accessible to them. When high-performance computing becomes just another tool on a worker’s bench. When information scattered across an enterprise becomes meaningful to the entire workforce, not just the three people who know where the bodies are buried. That is not swapping a muscle for a cheaper one. It is growing a muscle the organization never had.
And the real superpower is the combination: the distinctiveness and creativity of human beings, encoded into the system, multiplied by the brute power of the machine. The machine brings the scale. The people bring the judgment worth scaling.
Context is a capability
Here is where I bring it home, because this is the problem I get paid to lose sleep over.
Data runs a business through decisions. Every price, every route, every maintenance window, every hire is a decision made against data that somebody had to find, trust, and interpret. Most organizations are worse at that than they will admit. McKinsey found that knowledge workers spend nearly 20 percent of their week looking for internal information or tracking down the colleague who has it.
Gartner predicts that through 2026, organizations will abandon 60 percent of AI projects unsupported by AI-ready data, and 63 percent of organizations either lack or are unsure they have the data management practices AI needs.
Read that again. The bottleneck isn’t the model. It is that the business’s own knowledge is scattered, unlabeled, and locked inside the heads of the people who happen to remember it.
You can respond to that in two ways. You can point a swarm of agents at the mess and let them rediscover your enterprise from scratch on every question: a brilliant new hire with amnesia, who re-learns the org chart every morning and bills by the token. Or you can build a real context layer: one that does not just tell a model what your world contains, but hands the model, and the person behind it, the instruments to act on it. In the terms of the mathematics above, it is an innovation on the search space itself. It does not make the model search harder. It gives the model a better space to search: one described in the concepts your business actually reasons in, joined across systems that never talked to each other, so questions that used to be unaskable become askable.
Let me be clear about what this is not, because the market is crowded with things that sound like it. It is not another manual ontology project, where a team spends a year drawing boxes and arrows that are stale the day they are approved. It is not a natural-language box that turns a question into SQL against a curated catalog; that is a better search bar, and a useful one. It is not a symbolic rules engine bolted onto a model to grade its homework after the fact. Each of those is about information: describing it, retrieving it, or policing it.
We are after capability: what the model, and the person behind it, can actually decide and do.
Here is what that takes.
A language, not just a vocabulary. A structured language with a composable grammar, so questions and actions snap together like code instead of being improvised from scratch in free text every time.
Tools that write themselves. Tools generated from the ontology and from how people actually use the data, described in the business’s own language and scoped to what each user and agent may touch. An agent never reads your data model. It reads your tool descriptions, so we treat those as the product.
Instruments, not just answers. Forecasting models for demand, Bayesian methods for risk, expert systems for eligibility, and high-performance compute for the heavy runs, orchestrated behind those tools so the right engine answers each question. The language model earns its keep on what the data means, and on knowing which instrument to pick up.
The power to act, not just to read. Writes and structured workflows, so the same layer that explains the business can change it, under the same governance.
Memory with a sense of time. Findings tied back to the data they came from, each with a freshness timestamp, and a timeline of how the enterprise itself has evolved, so an agent can reason about trajectory, not just a snapshot.
An ontology that generates and maintains itself. Generated and kept current from schemas, documents, and query logs, with humans in the loop where formality matters, instead of drawn once and left to rot.
Two commitments sit underneath all of it.
Context is amortized reasoning.
You pay once to understand the data, the relationships, and the workflows, and every question afterward inherits that understanding. That is the honest answer to the token problem: stop paying a model to rediscover your business every time someone asks it a question.
Every answer carries its receipts.
Citations trace in two directions: into the ontology, so you can see which business concept the agent thought it was answering, and into the data, so you can see the rows and documents it relied on.
All of it is permission and identity aware, because an agent that remembers what one user was allowed to see and repeats it to another has built a very polite data leak.
The data stays where it needs to be: in your lakehouse, your databases, and your applications, and, across organizational boundaries, in your partners’, suppliers’, and regulators’ systems too, under its owner’s keys and its owner’s rules. The computation travels to the data. Only the answer travels back. The meaning consolidates. That is the ambitious part: a supply chain, a clinical trial, or a consortium of banks can reason over one shared understanding of the business without anyone handing over a copy of their data. Data is most valuable where it meets someone else’s data, and today sharing it means copying it, losing control of it, and trusting a contract no system enforces. Sell answers, not files.
Why does this matter to the argument I have been making?
Because a context layer built this way is, at its core, an enablement technology.
Its job is not to eliminate the analyst. Its job is to take what your best analyst knows on her best day (which revenue table finance actually trusts, what that twenty-year-old status code really means) and put it in front of the planner, the field engineer, the nurse manager, and the CFO at the moment they decide. Then it hands them instruments they have never had: the forecast, the Bayesian risk estimate, the simulation on high-performance compute, the governed workflow that actually changes something. Information enablement is the floor. Capability enablement is the point. It encodes the distinctiveness of your people once, and multiplies it by the scale of the machine.
Now let me tell you how far I think this goes, without front-running anything. For half a century, enterprise software has been a system of record: it tells you what happened. The next decade belongs to software that reasons: it tells you what is true, what will happen, what to do about it, and what you can safely share. And the customer for that reasoning is increasingly an agent: Datadog reported MCP tool calls up 22x since the fourth quarter of 2025, while Atlassian saw them rise 400 percent in a single quarter with barely a blip in human traffic. Three shifts get us there.
From questions answered by queries to questions answered by computation. Today an agent reads a table. Tomorrow the same agent, on the same governed data, optimizes a route, simulates a part, or folds a protein, on behalf of a planner or a scientist who has never written a line of Python. That is high-performance computing as just another tool on the bench. Scientists get to focus on science, and accountants on finance.
From answers in a chat window to decisions inside the business. A business process does not want a paragraph. It wants a value it can branch on, a confidence, and a receipt. When an agent proves it can answer a question well, that answer can be hardened into a decision the business calls a thousand times a day. The model is probabilistic once, when the decision is authored and validated, and deterministic every time it runs. That is the honest answer to the non-determinism I complained about at the top.
From the data you have to the world you operate in. Documents, images, video, sensor streams, and the public record become typed facts with their evidence attached. Aircraft, towers, molecules, patients, and supplier networks join the model of the business, instead of living in a PDF nobody has time to read.
And underneath all three: never pay for the same insight twice. Today an expensive answer (a week of analyst time, a GPU experiment, a long investigation) dies in a chat history or a notebook. With the right context layer, it becomes a finding, and a finding is far more than a note. It links back to the citations it rests on, to the exact data it was computed from, and to the concepts in the ontology it is about. It carries its method and a receipt that can re-run it, and it links forward to whatever later confirms, extends, or contradicts it. Touch any of those threads (a source, a dataset, a customer, a molecule) and the finding surfaces, so the next person and the next agent start from it instead of re-deriving it. That is how you stop brute-forcing your own enterprise. Insight compounds instead of evaporating.
That is the lens for what you will see from us at AI World. I am not going to front-run the announcements. I will just give you the question to ask of everything you see there, ours included: Does this make my people capable of something they could not do before? Or does it do what they already do, slightly cheaper and considerably less reliably?
Three things, and then I’ll get off the stage
If you are building in this industry, sit with three things before you make your entire business about replacing a human workflow with an agent workflow.
One: the token economics are not settled. Per-token prices fall, tokens per task climb, and the biggest labs selling them have, until very recently, spent far more than they earned. A margin that depends on someone else’s subsidy is not a business model. It is a promotional period, and promotional periods end.
Two: distinctiveness is an asset, and the average machine erodes it. Models regress toward the most common answer; I have written about the average machine before. Swap your most distinctive people for the median of the internet and you do not just get an average answer. You get an average company. And if you cannot reconstruct why the agent did what it did, you get an untrustworthy one.
Three: you are leaving the superpower on the table. The prize is not a workforce that costs a little less. It is a workforce that achieves ten times more at a tenth of the latency, because the knowledge of the whole enterprise is finally at every person’s fingertips. Headcount reduction is a line item. A workforce like that is a strategy.
The railroad took us out of the fields and put us in classrooms. The computer and the internet took the classroom global. Every one of those revolutions made the same bet: give human beings better tools and they will do something more interesting with their time than you could have predicted. That bet has never lost.
We get to make it again. We can say the model will replace our intellectual achievement. Or we can say that, through the right context, the model will supercharge it, by making the data, the tools, and the conversations that were opaque transparent to everyone who needs them.
I know which one I am building. Facebook retired “move fast and break things” a dozen years ago. It is time we retired “move fast and replace people.”