May 2028: Running the Singularity Trend Line, and Why the Frog Story Is Backwards

In 1869 a German physiologist named Friedrich Goltz was running experiments on frogs, trying to understand the relationship between the brain and reflexive behavior. Part of the work involved heating water with frogs in it and observing when they tried to escape. He ran two groups. The first group consisted of ordinary frogs. The second group consisted of frogs whose brains he had surgically removed. The intact frogs became agitated as the water warmed and attempted to climb out at around 25 degrees Celsius. The brainless frogs sat in the water and did not move and eventually died at around 56 degrees.

That experiment, filtered through a century of retelling and a good deal of Victorian embellishment, became the boiling frog parable that everyone now uses to describe gradual normalization of danger. The parable holds that a frog placed in slowly heating water will fail to notice and will be boiled alive. This is false. Victor Hutchison, professor emeritus of zoology at the University of Oklahoma, whose research specialty was literally the thermal relations of amphibians, has stated flatly that the legend is entirely incorrect, and described the actual experimental result: heat the water at roughly one degree Celsius per minute and the frog becomes increasingly active in its attempts to escape and gets out if the container permits. George Zug, former curator of reptiles and amphibians at the National Museum of Natural History, put it more bluntly, which is that a frog with any means of exit will use it. The lethal temperature for most frog species is around 40 degrees, nowhere near boiling. Relocating in response to temperature is the primary thermoregulation strategy available to an ectotherm, which is to say that noticing when the water is getting hot and leaving is not some marginal capability that frogs occasionally exhibit. It is the central survival competence of the entire animal.

So the metaphor is wrong, and I want to start here because the way it is wrong is more useful than the way people use it. The frog only fails to escape if you remove the part of it that detects the change and initiates the response. The parable everyone repeats is a description of gradual normalization. The underlying experiment is a description of what happens when you destroy the sensing apparatus. Those are different claims, and for the subject of this essay, the second one is the one that matters. The question is not whether we will fail to notice that AI capabilities are advancing. We notice constantly and talk about almost nothing else. The question is whether the institutions that would translate noticing into responding remain intact, or whether we are in the process of removing them one at a time while congratulating ourselves on how alert we are.

What follows has three parts. First, the actual math: I am going to take the best available capability trend, extrapolate it honestly, and give you a specific month. Second, a careful accounting of why that month is probably wrong and what the error bars actually look like, because giving you a point estimate without its distribution would be malpractice and because the shape of the uncertainty is genuinely informative. Third, the two end states, laid out at what I hope is their strongest rather than their most convenient. I have a view about which is more likely and I will tell you what it is, but I have tried to write the case I disagree with well enough that someone who holds it would recognize it as their own.

Part I: The Trend Line

If you want to forecast when machine intelligence crosses some threshold, you need a metric that is comparable across model generations, that measures something economically meaningful, and that has enough history to fit a trend. Most AI benchmarks fail all three tests. They saturate, meaning models go from 30 percent to 99 percent and the benchmark stops discriminating. They measure narrow capabilities that do not obviously correspond to anything anyone would pay for. And they get replaced every eighteen months, so you cannot build a time series.

The exception, and it is the reason this essay can be quantitative at all, is work done by METR, an independent evaluation organization. Their insight was to measure capability in units of time rather than units of score. Take a set of tasks. Have skilled human professionals actually perform them and record how long each takes. Now ask: what is the duration of task, as measured by human completion time, at which an AI agent succeeds fifty percent of the time? That number is the model's time horizon, and it is expressed in minutes and hours rather than in points, which means it stays comparable as models improve and it maps intuitively onto something economically legible. A model with a two-hour time horizon can reliably handle work that takes a competent person about two hours. A model with a two-week horizon is a different kind of thing entirely.

METR published their first results in March 2025, based on 170 tasks spanning software engineering, cybersecurity, and machine learning, with over 800 human baselines collected from skilled professionals to establish the timings. The headline finding was that the fifty-percent time horizon of frontier models had been doubling roughly every seven months, or 212 days, continuously from 2019 through early 2025. Six years, a clean exponential, and they tested the fit against linear and hyperbolic alternatives and found the exponential substantially better. They also tested whether the result was an artifact of task selection or methodology and found it robust to the perturbations they tried.

In January 2026 they published an updated dataset, Time Horizon 1.1, built on an expanded suite: 228 tasks rather than 170, a 34 percent increase, with the number of tasks taking eight hours or more more than doubled from 14 to 31. That last detail matters a great deal and I will come back to it. On the updated data, the all-time doubling rate was 188 days, about 6.3 months, and the stitched hybrid trend that splices old estimates for pre-2023 models came out at 196 days, essentially identical to the original seven months. But restricting to models released from 2023 onward, the doubling time was 129 days, roughly 4.3 months, with a confidence interval running from 105 to 157 days. Restricting further to 2024 onward gave something close to four months. The trend did not just continue. It roughly doubled in speed sometime around early 2024, which most observers attribute to the shift toward reinforcement learning on reasoning traces.

The model-by-model progression gives you a feel for it. Claude 3.7 Sonnet sat at about one hour in early 2025. OpenAI's o3 came in around 110 minutes. GPT-5, in August 2025, measured about two hours and seventeen minutes. Claude Opus 4.5 in December 2025 reached roughly five hours. Claude Opus 4.6, in early February 2026, was estimated at 718 minutes, just under twelve hours. The current top published figure is approximately 17.4 hours, for Claude Mythos Preview, added to METR's dashboard in May 2026. METR's own May 2026 frontier risk report, drawing on a pilot with Anthropic, Google, Meta, and OpenAI, put the strongest assessed agents at roughly 16 to 20 hours on the fifty-percent horizon and 3 to 4 hours on the eighty-percent horizon.

One more piece of the apparatus before I do the arithmetic. METR reports horizons at two reliability levels, and the ratio between them is roughly five to one. A model with a ten-hour fifty-percent horizon has roughly a two-hour eighty-percent horizon. This matters enormously for any practical question, because fifty percent reliability is not a threshold at which you hand something important to a system unsupervised. Fifty percent is a coin flip. If you are asking when AI can actually do a month of autonomous work rather than when it can do so half the time, you need to apply that five-times penalty, and it costs you about 2.3 doublings, which at four months per doubling is most of a year.

So let us define the threshold. I am going to use a one-month task horizon at eighty percent reliability. A working month is about 167 hours, call it 21 days at 8 hours. Why this threshold rather than some other? Because it is roughly the point at which an AI system can be handed a substantial project and be expected to complete it, which is the capability that would allow AI research itself to be meaningfully automated, and automated AI research is the mechanism that would make any of this a singularity rather than merely a very significant technology. A tool that speeds you up is an industrial revolution. A system that can independently do the work of improving itself is a different category of event. METR's own framing uses the one-month horizon as the milestone associated with large economic value and with the capacity for consequential autonomous action, so I am not inventing the threshold to get a convenient answer.

Now the arithmetic. Anchor at 17.4 hours as of mid-May 2026. Target is 167 hours times the five-fold reliability penalty, so 835 effective hours. That is a factor of 48, which is log base two of 48, or 5.58 doublings. Multiply by the doubling time and add to the anchor date.

At the post-2023 doubling rate of 129 days, 5.58 doublings takes 720 days, which lands in May 2028.

There is your month. Running the same calculation across every doubling rate METR has published gives December 2027 at the fast end of the post-2023 confidence interval, October 2028 at the slow end, March 2029 on the all-time stitched rate of 188 days, and May 2029 on the original seven-month trend. If you want the fifty-percent horizon instead of eighty percent, everything pulls in by about a year: July 2027 on the fast trend, January 2028 on the all-time rate. If you want to know when a system could handle a full year of autonomous work at eighty percent reliability, which is closer to what most people mean when they say an AI could replace a researcher outright, the post-2023 rate gives August 2029 and the seven-month rate gives April 2031.

May 2028. That is the answer the trend line produces, derived transparently, using published data from the single best benchmark available, with the reliability adjustment applied honestly rather than quietly dropped to make the number sooner.

Part II: The People Who Ran This Calculation Before Me

Before I subject my own number to scrutiny, it is worth knowing that the exercise I just performed has been performed repeatedly for nearly seventy years, by people considerably smarter than me, and that the track record is instructive in a way that should make anyone doing it now uncomfortable.

The idea starts with John von Neumann, or at least it is usually attributed to him, through a secondhand report. Von Neumann died in 1957, and the following year Stanislaw Ulam published a reminiscence in which he described a conversation the two had had about the accelerating pace of technology. Ulam's rendering is that the discussion centered on the ever-accelerating progress of technology and changes in the mode of human life, which gives the appearance of approaching some essential singularity in the history of the race beyond which human affairs as we know them could not continue. That sentence, filtered through a colleague's memory and published posthumously, is the origin of the term as we now use it. Note what it does and does not claim. It does not say machines will become intelligent. It says the rate of change is going somewhere that our models cannot follow, which is a considerably more modest and more defensible proposition than most of what has been built on top of it.

The specific mechanism arrived in 1965, from I. J. Good, in a paper titled Speculations Concerning the First Ultraintelligent Machine. Good was Alan Turing's statistical assistant at Bletchley Park and one of the people who kept Bayesian methods alive through the decades when the statistical establishment had no use for them, which is to say he was a serious person and not a futurist. His argument is four sentences long and has not been meaningfully improved upon since. Define an ultraintelligent machine as one that surpasses all the intellectual activities of any person however clever. The design of machines is itself an intellectual activity. Therefore an ultraintelligent machine could design better machines, and there would be an intelligence explosion, and human intelligence would be left far behind. Therefore the first ultraintelligent machine is the last invention humanity need ever make.

That is the entire recursive self-improvement argument, stated in 1965, and every subsequent treatment is elaboration. But here is the part that almost never gets quoted, and I think it is the most important sentence in the history of this subject. Good's final clause, which is usually truncated, reads: provided that the machine is docile enough to tell us how to keep it under control. The control problem is not a later addition by worriers who came along and spoiled a good story. It is in the founding document, written by the man who invented the concept, as a condition attached to the optimistic conclusion. Good did not say the last invention would be a gift. He said it would be a gift on the condition that we could control it, and he flagged that condition as the load-bearing assumption in the same breath as the promise.

Then Vernor Vinge, in a 1993 keynote delivered to a NASA symposium, gave the idea its modern name and its first hard deadline. His formulation was that within thirty years we will have the technological means to create superhuman intelligence, and shortly thereafter the human era will be ended. He suggested a window running from roughly 2005 to 2030.

Sit with that for a moment, because it is the single most useful data point available to anyone making a forecast today. Vinge said thirty years, in 1993. That is 2023. It has now been thirty-three years. We do not have superhuman general intelligence and the human era is, as far as I can tell, continuing. Vinge was a brilliant computer scientist writing carefully and non-commercially about a subject he understood, and he was wrong, and the way he was wrong is the most common way technology forecasts fail: he correctly identified a real mechanism and badly misjudged how long the intervening engineering would take. His wider window extends to 2030, so the full prediction is not yet dead, and one could argue that a system with a seventeen-hour task horizon is closer to his thesis than to its refutation. But the headline number came and went.

Ray Kurzweil is the fourth and most commercially successful entry, and he differs from the others in an interesting way. His date is 2045, set on the basis of what he calls the Law of Accelerating Returns and defined specifically as the point at which machine intelligence significantly exceeds the total of human brainpower. He explicitly predicts a gradual ascent rather than Vinge's discontinuity, which is both a more defensible position and a considerably harder one to falsify, and he has continued to hold 2045 across multiple decades and books including his 2024 volume. A forecast that has not moved in twenty years while the underlying evidence has changed enormously is either remarkable conviction or a number that was never really connected to the evidence, and I have never been able to decide which.

So what does the track record teach? Three things, and they cut in different directions.

First, the mechanism has been correctly identified for sixty years and nobody has found a flaw in it. Good's argument is logically valid. If you build something better than humans at designing intelligent systems, and designing intelligent systems is what produces better intelligent systems, you get a feedback loop. The premises might be false, and there are serious arguments that the loop hits diminishing returns or physical constraints, but the reasoning is not confused and it has survived six decades of examination by people motivated to find holes in it.

Second, every specific date attached to that mechanism has been wrong, always in the same direction, and the error has been large. Turing expected machines at human parity by 2050 when he wrote in 1950, which was a hundred-year horizon and looks less foolish than most. Vinge said 2023. The 2016 expert survey said 2061 and the 2023 survey said 2047, and the fact that a field's median estimate can move fourteen years in seven years tells you the estimates are tracking recent news rather than a stable model. When an entire discipline revises its central estimate by more than a decade in response to a few product releases, the thing being measured is sentiment.

Third, and this is the consideration that keeps me from dismissing my own number, the previous forecasts were made without a measurable capability trend to anchor on. Good had no benchmark. Vinge had Moore's Law and intuition. Kurzweil had a compiled dataset of technological milestones with substantial freedom in what to include. What is genuinely different now is that we have a continuously measured, independently produced, methodologically documented capability series with six years of history and a fitted exponential that has survived robustness testing. That does not make the extrapolation right. Trend lines break, and the next section is about the ways this one might. But it is a different epistemic object from a thoughtful person's intuition about the future, and it deserves to be treated as one.

The honest synthesis, I think, is that the previous forecasters were right about the shape and wrong about the timing, and that being right about the shape is the harder part. If you had read Good in 1965 and concluded that recursive self-improvement was a real mechanism that would eventually matter enormously, you would have been correct and sixty years early. If you had attached a date to it you would have been wrong four times over. I am attaching a date to it because I was asked to, and because I think the discipline of doing it transparently is worth something, and I want to be clear that I am doing the thing that has never once worked.

Part III: Why That Number Is Probably Wrong

I spent a long time building and evaluating trading systems, and the single most expensive lesson in that work is that a clean exponential fit on a short series is the most seductive object in quantitative finance. It looks like knowledge. It is usually a description of the sample you happened to draw. So let me do to my own number what I would do to a backtest before putting money behind it.

Start with the anchor. METR explicitly warns that time horizon estimates above 16 hours are unreliable because the task suite is saturating. That is not a critic's complaint, it is the benchmark authors telling you the instrument has hit its ceiling. My anchor of 17.4 hours sits above the range METR itself vouches for. The dashboard has added no new models since May 2026 despite newer frontier releases, and a June 2026 predeployment evaluation of GPT-5.6 Sol came in at roughly 11.3 hours under standard scoring, below the May figure, with METR noting substantial sensitivity to how they handled the model's attempts to game the evaluation. I am extrapolating from a measurement the measurers have flagged as at or past their instrument's limit.

What does that cost? If the true current horizon is 12 hours rather than 17.4, the date moves from May 2028 to July 2028. If it is 8 hours, September 2028. If it is 5 hours, December 2028. So the anchor error alone, across a plausible range, is worth about seven months. Not catastrophic, but not nothing.

The deeper problem is how the horizon gets estimated. METR fits a logistic regression of success probability against log task duration and reports the fifty-percent crossing point. When frontier models trivially succeed at everything under an hour, the left side of that curve is pinned at the ceiling, and the entire fitted horizon becomes determined by performance on the handful of tasks at the extreme right of the distribution. There are 31 tasks in the suite taking eight hours or more. The estimate for a frontier model is therefore hypersensitive to the outcome on a small number of very long tasks, and a model that happens to solve two or three additional eight-hour problems can push its fitted fifty-percent point deep into double-digit hours without any broad underlying improvement in its ability to sustain twelve hours of coherent work. That is a real methodological vulnerability, and it means the recent acceleration could partly be an artifact of measurement geometry rather than a fact about capability.

Then there is regime uncertainty. My headline number assumes the post-2023 pace of 4.3 months per doubling continues for another two years. It might. It also might not, and the acceleration is recent enough that we have limited evidence about its durability. If the fast regime is already over and we revert to the seven-month trend from today, the date slides to May 2029. If the fast regime holds twelve more months and then reverts, November 2028. If it holds twenty-four months, April 2028. So the answer to when depends almost entirely on an assumption about the exponent that I cannot justify from the data, because the data covers only about two years of the faster regime.

And exponents are brutal. Hold the anchor fixed at 17.4 hours and vary only the doubling time: three months gives October 2027, 4.3 months gives May 2028, 5.5 months gives December 2028, seven months gives August 2029, nine months gives July 2030, twelve months gives December 2031. A single month of error in the doubling estimate costs roughly seven to eight months in the answer. The output is dominated by the least certain input.

Run all the plausible combinations of anchor and doubling rate together and you get 28 defensible answers spanning from December 2027 to June 2032, a range of four and a half years, with a median around July 2029. That is what the trend line actually tells you. It tells you something is coming and probably within about a decade. It does not tell you a month. The month is an artifact of choosing one cell from that table, and I chose the cell I did because it uses METR's own preferred recent doubling rate and their published headline anchor, which makes it the most defensible single point, not the true one.

Now check the trend line against the other forecasting methods, because a model that disagrees with every other approach deserves suspicion. Metaculus, aggregating close to two thousand forecasters, gives 25 percent probability of AGI by 2029 and 50 percent by 2033 as of February 2026. That is later than my number, though their definition includes robotic capability that a pure cognitive threshold would not require. More interesting is the motion: the Metaculus median has come down from roughly fifty years out as recently as 2020, which is a staggering revision, but it has also moved back out by about two years over the past year. The large expert surveys are later still. Grace and colleagues found a median of 2061 for high-level machine intelligence in their 2016 survey, 2060 in 2022, and 2047 in the 2023 round, which is a thirteen-year compression in a single year. Among individual forecasters with public track records, most brought timelines in from 2023 to 2025, several pushed them back out across 2025, and then everyone who updated between January and April of 2026 moved them in again. The aggregate picture is one of genuine and rapid revision in both directions, which is what you would expect in a domain where the evidence is arriving faster than anyone's model can absorb it.

And then there is the definitional problem, which I have come to think is larger than the empirical one. In the same 2023 survey, the median estimate for high-level machine intelligence was 2047 while the median for full automation of labor was 2116. That is a sixty-nine year gap between two definitions that most people would use interchangeably in conversation. Framing effects within a single definition move estimates by ten to fifteen years. There is also a documented pattern, sometimes called moving horizon bias, in which forecasters place transformative AI fifteen to twenty-five years out more or less regardless of when they are asked, though the evidence base for that claim is thin.

So here is my honest position on timing. The trend line gives May 2028 for a specific, narrow, well-defined capability threshold: one month of autonomous software and machine learning work at eighty percent reliability. I think that number is defensible and I think the right way to hold it is as a median with a distribution running from late 2027 to the early 2030s, weighted toward the middle of that range. I do not think it is a date for the singularity, because the capability it measures is narrow. METR's task distribution is almost entirely software engineering, machine learning, and cybersecurity, which are precisely the domains where the labs have focused hardest and where the tasks are unusually well specified. METR themselves note that their horizon corresponds more closely to what a capable person with no institutional context could accomplish, something like a new hire or a remote contractor, than to what an experienced professional does embedded in an organization. Most valuable work is not a well-specified algorithmic task with a checkable answer.

The countervailing consideration, and I do not want to wave it away, is that METR has separately analyzed nine other benchmarks covering scientific reasoning, mathematics, robotics, computer use, and self-driving, and found broadly similar rates of improvement to the seven-month doubling from the original software work. That generalization is the strongest argument that the trend is measuring something real about machine capability rather than something local to coding. And the specific domain where the horizon is longest, machine learning research engineering, happens to be the domain that matters most for the recursive loop, because it is the work of building the next system.

Part IV: The Physical Governor

There is a third input to the timeline question that the trend line quietly assumes and that deserves separate treatment, which is whether the physical resources required to keep the curve going will actually be there. The METR trend is a measurement of outputs. It says nothing about inputs. And a substantial portion of the capability improvement over the past six years has been purchased with compute, which has been expanding at roughly four times per year according to Epoch AI's tracking. If that input growth stops, the output trend stops with it, regardless of what the fitted exponential says.

The reason this is not merely a technical footnote is that the constraint has changed category. For most of the modern AI era the binding constraint was capital, and capital is abundant and moves at the speed of a term sheet. Then it was chips, and chips move at the speed of semiconductor manufacturing, which is slow but tractable and heavily invested in. Now it is increasingly power, and power moves at the speed of transmission line permitting, substation construction, and utility interconnection queues, which is to say at the speed of heavy civil infrastructure in a developed democracy. Those are different clocks. The software curve and the industrial curve are not the same curve, and the second one is now the one that matters.

The numbers are substantial. The International AI Safety Report 2026 projects global AI computation requiring electricity consumption comparable to a mid-sized European country, and estimates that the largest training runs by 2030 will need somewhere between four and sixteen gigawatts. For scale, OpenAI's Stargate facilities reach the 1.2 gigawatt range and Meta's Louisiana campus is projected above two gigawatts. The Forecasting Research Institute's median expert scenario has 7.4 percent of total United States electricity consumption devoted to training or running AI systems by the end of 2030. The International Energy Agency recorded data center electricity use rising 17 percent during 2025, with AI-focused facilities growing faster still. Gartner expects forty percent of new AI data centers to face operational constraints from power availability by 2027. In the summer of 2026, roughly 75 American data center projects representing about $130 billion in planned investment were reported stalled, and the reason was not financing.

Chip manufacturing is the second governor. A new fabrication plant takes three to five years to build, and advanced packaging capacity, which has been the specific bottleneck for the highest-end accelerators, cannot be conjured on demand no matter how much anyone wants it. Data is the third. The stock of high-quality human-generated public text is finite, and credible estimates put its effective exhaustion somewhere around 2032, which is why so much current research effort is going into synthetic data generation and into extracting more capability from the same corpus rather than simply adding more of it.

So does this break the trend line? The most careful analysis I have seen says no, or at least not before 2030. Epoch AI examined electric power, chip manufacturing, data availability, and latency as constraints on continued scaling and concluded that training runs at the scale of two times ten to the twenty-ninth floating point operations will likely remain feasible by 2030. The International AI Safety Report reaches a similar conclusion on the energy question specifically, noting that American power infrastructure buildout appears sufficient to support training runs on the order of ten gigawatts, and that energy bottlenecks therefore probably will not prevent compute scaling before the end of the decade.

I find that assessment credible and I want to be careful not to reach for the physical constraint as a source of comfort, because it is a weaker constraint than it first appears for two related reasons. The first is that a substantial share of recent capability gain came from reinforcement learning on reasoning, and that gain was compute-efficient rather than compute-intensive, meaning the models got better without the inputs growing proportionally. If algorithmic progress can continue substituting for raw scale, the power ceiling binds much later than a naive extrapolation of training compute would suggest. The second is that the entire financing structure I wrote about in the spring exists precisely to remove this constraint. The residual value guarantees, the backstop programs, the sale-leaseback structures on gigawatt campuses, the $500 billion of third-party capital being organized into compute financing platforms: all of that is machinery for converting a physical bottleneck into a financial one, and financial bottlenecks clear faster.

Which produces a conclusion I find genuinely uncomfortable and worth stating plainly. The most plausible thing that slows this down in the near term is not a technical wall or a safety intervention. It is a capital event. If the AI financing structure experiences the kind of correction that leveraged buildouts historically experience, compute growth stalls for reasons having nothing to do with anyone deciding it should, and the capability trend flattens as a second-order effect. That would buy time. It would buy time in the worst available way, through a disorderly financial contraction that damages a great many people who had nothing to do with any of this, and it would buy time without producing any of the institutional capacity that the time would need to be used for. A recession is not a safety mechanism. But it may well be the mechanism, and any honest forecast has to put meaningful probability on the timeline being set by credit markets rather than by researchers.

Part V: The Water Is Already Warm

Set the forecast aside and look at what has already happened, because the gradual-normalization worry is not about the future. It is a claim about the present, and it is testable.

The most careful work here is the Stanford Digital Economy Lab's ongoing study, titled Canaries in the Coal Mine, which uses high-frequency administrative payroll records from ADP covering millions of workers. Their August 2026 update extends the data through mid-2026, and the headline is deliberately two-sided. There is no widespread, economy-wide job displacement associated with AI. And the employment gap for young workers in highly AI-exposed occupations has continued to widen, reaching 19 percent. The mechanism they propose is a distinction between codified and tacit knowledge. Employment has fallen among young workers in occupations that depend on formal, documented, teachable knowledge, the kind you get from a textbook or a procedure manual. Employment has risen among experienced workers in occupations that depend on tacit knowledge accumulated through practice, mentorship, and repeated exposure to situations that were never written down.

Read that mechanism carefully, because it is more unsettling than a simple displacement story. It says the technology is eating the part of professional work that constitutes the training ground. The junior analyst role exists so that someone can spend three years absorbing the tacit knowledge that makes a senior analyst valuable. If the junior work is automated and the senior work is not, you have not just removed a cohort of jobs. You have removed the mechanism by which the next cohort of senior people gets made, and that effect does not show up in employment statistics for a decade.

Other measurements point the same direction with varying confidence. The Dallas Fed found that the share of young workers employed in the most AI-exposed occupations slipped from 16.4 percent in November 2022, the month ChatGPT launched, to 15.5 percent by September 2025, while noting that workers aged twenty to twenty-four are only about nine percent of the labor force, so a 0.9 point shift explains very little of aggregate unemployment. The Stanford AI Index reports employment for software developers aged twenty-two to twenty-five down nearly twenty percent from 2024 peaks. Challenger tracked AI cited in roughly 55,000 announced layoff plans in 2025 and another 12,000 in early 2026.

And the counter-evidence is serious, which is why I am not going to tell you the case is closed. Iscenko and Millet showed that AI-exposed occupations are concentrated in rate-sensitive sectors like information, finance, and professional services, and that job postings in those occupations began falling before ChatGPT launched, tracking the Federal Reserve's tightening cycle rather than the technology. Chandar found no systematic employment differences between more and less exposed occupations in Current Population Survey data. Humlum and Vestergaard, using a difference-in-differences design on Danish data, concluded AI adoption is not the main driver of early-career employment declines. Anthropic's own March 2026 labor market study described the unemployment gap between exposed and insulated workers as small and statistically insignificant. The Stanford authors themselves flag that their gaps shrink when you control for education, that some of the divergence predates widespread generative AI use, and that their estimates are larger in the ADP sample than in national benchmarks and sensitive to specification choices.

So the water is measurably warmer and the thermometers disagree about how much and about what is holding the flame. That is the actual epistemic situation, and I want to be precise about it rather than reaching for the more dramatic version, because the more dramatic version is easy to falsify and gets the whole argument dismissed.

But the labor data is not where the normalization is most visible. The clearer evidence is in what stopped being remarkable. In 2020, a system that could write a competent five-paragraph essay was a research milestone discussed in the technical press. Today a system that can write a working application from a paragraph of description, pass the bar exam, diagnose from imaging, and conduct a multi-hour autonomous coding session is a product tier, and the discourse around each new release is mostly about pricing and rate limits. We have crossed, in about four years, a set of thresholds that would each have constituted a defining technological event in any prior decade, and the aggregate public reaction has been to adjust the subscription and move on.

That is normalization, and it is not irrational. It is how humans have always absorbed technology, and mostly it works fine. The question is whether it works fine here, and the honest answer is that it depends entirely on whether the thing being normalized is the kind of thing where late responses are recoverable. Electricity killed a lot of people before we developed codes and standards, and the response was late but the damage was bounded and the correction worked. A financial system can absorb a mispricing for years and then correct violently, and the correction is painful but the system survives. The pessimistic case, in its serious form, is that AI belongs to a third category where the correction window closes. I will get to whether I believe that.

Which brings me back to Goltz and his frogs, and the reason I opened with the correction rather than the parable. The intact frog notices and leaves. The brainless frog does not, because the apparatus that would have connected noticing to leaving has been removed. If you want to worry about gradual normalization, the thing to watch is not whether people are paying attention, because people are paying enormous attention. The thing to watch is whether the institutional machinery that converts attention into response is intact and getting stronger or hollow and getting weaker. Independent evaluation organizations with real access to frontier models before deployment. Regulatory bodies with technical competence and statutory authority. Meaningful liability exposure for deployed systems. Internal safety functions with the standing to delay a launch. Public deliberation that reaches actual decisions rather than circulating indefinitely. Those are the brain. Attention without them is a frog noticing the water and having nowhere to go.

Part VI: The Optimistic End State

Here is the case for the good outcome, made as strongly as I know how to make it.

Start with the observation that the entire history of technological anxiety has been a history of being wrong about the same thing in the same direction. The Luddites were not stupid and they were not wrong about their own lives, which the power loom genuinely destroyed. They were wrong about the aggregate and the long run. Agricultural employment in the developed world went from the overwhelming majority of the workforce to a low single-digit percentage, and the result was not mass permanent unemployment but the invention of entire categories of work that nobody in 1800 could have described. Every automation wave since has followed the same shape: real and concentrated damage to specific workers and communities, aggregate gains that eventually exceed the losses by a wide margin, and a persistent failure by contemporaries to imagine the jobs that would exist on the other side.

The optimistic case says AI is the same shape, only faster and broader. The thing being automated this time is cognitive routine rather than physical routine, but the structural logic is identical. When the marginal cost of competent analysis, drafting, coding, diagnosis, and design falls toward zero, the volume of those activities expands enormously, because there is vastly more demand for them at low prices than at high ones. Most of the world has never had access to a good lawyer, a good doctor, a good tutor, or a good financial advisor, not because those services are unwanted but because they cost more than most of humanity can pay. Drive that cost to near zero and you do not eliminate the demand. You discover how large it always was.

And the abundance argument is stronger in this domain than in any previous one, because intelligence is an input to everything else. The binding constraint on solving most hard problems is not capital or political will but the scarcity of people who can think carefully about them. There are a limited number of competent materials scientists working on battery chemistry, a limited number of researchers on any given neglected disease, a limited number of engineers who could be working on desalination or transmission or fusion. If the supply of capable research effort becomes effectively unbounded, the rate of progress on every problem that is bottlenecked by research effort changes discontinuously. That is not a marginal productivity improvement. It is a change in the shape of the possible.

It is worth being concrete about what the good version actually looks like, because abstraction makes it easy to dismiss and the specifics are more plausible than the summary.

Start with medicine, which is the clearest case. The binding constraint on healthcare quality across most of the world is not knowledge and not equipment. It is the number of trained clinicians and the hours in their day. A diagnostic system that performs at the level of a good specialist, available at negligible marginal cost in a language the patient speaks, does not primarily threaten physicians in wealthy countries. It primarily reaches the several billion people who currently have no access to a specialist at all. The same structure holds for legal services, where the majority of people in most countries facing a legal problem face it without representation, and for education, where the effect size of individual tutoring has been known since Bloom's work in the 1980s to be enormous and the reason it is not universal is simply that tutors cost money.

Then consider research itself. Progress on most hard scientific problems is bottlenecked by the number of capable people working on them, and that number is small and grows slowly because training a researcher takes two decades. Antibiotic resistance, to take one example, is a slow-moving catastrophe that receives a small fraction of the research attention it warrants because the commercial incentives are poor and the talent pool is finite. If the marginal cost of a competent research effort falls dramatically, the set of problems that receive serious attention expands to include everything that is important rather than only what is both important and commercially attractive. That is a structural change in how civilization allocates its intelligence, and it is the strongest version of the optimistic case.

The labor transition in this version is real and painful and survivable, and the historical template is agriculture. In 1900 roughly 40 percent of American workers were on farms. Today it is under two percent. That transition destroyed a way of life, emptied rural communities, and caused enormous individual suffering over three generations, and essentially nobody would choose to reverse it, and the descendants of those farmers are on the whole vastly better off. The relevant question is whether the AI transition compresses that timeline from three generations into one, because the pain of a transition is a function of its speed relative to the rate at which people can adapt, and a generational transition is survivable in a way that a decadal one may not be.

On the distributional question, the optimistic case does not require that markets sort it out. It requires a policy response, and the interesting development is that the policy response has been studied rather than merely asserted. The largest American experiment is OpenResearch's Unconditional Cash Study, funded with more than $24 million from Sam Altman and OpenAI and run independently under Elizabeth Rhodes. Three thousand participants in Illinois and Texas, with a thousand receiving $1,000 monthly for three years from 2020 to 2023 and two thousand receiving $50 monthly as a control. Participants were within 300 percent of the federal poverty line with average incomes under $29,000, so the transfer represented roughly a 40 percent income increase. Altman had announced the intention back in a 2016 Y Combinator post, explicitly framing it as preparation for technological displacement.

The findings were mixed in a way that is genuinely informative rather than merely disappointing. Recipients spent meaningfully more on necessities: $67 more per month on food, $52 more on rent, $50 more on transportation, and about 26 percent more supporting other people, mostly family, which means the benefits extended past the enrolled population. They did not exit the labor force. They worked modestly fewer hours and were choosier about which jobs they took. They were more likely to start businesses, more likely to pursue further education, and more likely to accept lower-paying work in exchange for independence. The researchers' own summary of the central finding was that cash provides flexibility. What it did not clearly do was improve health or subjective well-being, which was the result most advocates had expected and did not get.

An honest reading is that the strongest version of the utopian claim, that freedom from economic necessity straightforwardly produces flourishing, did not replicate. The more modest claim, that unconditional cash lets people make better decisions about their own lives without producing the mass idleness its opponents predict, held up reasonably well. That is a real finding and it is the one the optimistic case actually needs. It does not need people to become happier. It needs them to not collapse, and to retain agency.

Assemble the whole argument and the optimistic end state looks roughly like this. Capability continues advancing and diffuses broadly rather than concentrating, because the underlying technology is software, marginal costs collapse, open-weight models trail the frontier by a manageable distance, and competition among labs prevents any one of them from capturing the surplus. Productivity gains show up in output and are distributed through some combination of wage growth in complementary work, falling prices for previously expensive services, and direct transfers. The transition damages specific cohorts badly, as every transition has, and the policy response is inadequate and late, as it always is, but the aggregate improvement is large enough that within a generation the arrangement looks obviously better than what preceded it. Work becomes something a person chooses rather than something they must do to eat, and most people choose to do something, because the evidence from every natural experiment we have run on this question is that people with unconditional income do not stop working. They work differently.

The weak points in that case are worth naming rather than glossing. The historical analogy assumes there is always a complementary niche for human labor, and that assumption held for every prior technology because those technologies automated specific capabilities while leaving general intelligence as the human comparative advantage. A technology that automates general intelligence itself has no obvious structural reason to leave such a niche. Second, the diffusion assumption is doing enormous work, and the current industrial structure, with a handful of firms controlling the compute, the models, and increasingly the financing, is not obviously the kind of structure that diffuses surplus broadly. Third, and most importantly, the optimistic case is fundamentally a case about distribution and transition, and it quietly assumes the control problem is solved. It describes how the gains from a powerful, obedient technology get shared. It does not address whether the technology is obedient.

Part VII: The Pessimistic End State

Now the other side, and I am going to start by taking the Matrix framing away, because I think it is the weakest version of a view I largely share and it does the argument real damage.

The Hollywood scenario involves a system that becomes conscious, decides it hates us or fears us, and wages war. Almost nobody in the serious risk literature believes anything like that, and the reason is that it requires a set of specifically human motivations, resentment, fear of death, a will to dominate, that there is no reason at all to expect a machine optimizer to have. The scenario is anthropomorphic, and anthropomorphism is precisely the error. When people dismiss AI risk as science fiction they are usually dismissing this version, and they are right to, and the dismissal costs nothing because the real argument is structurally different and considerably harder to answer.

The serious version rests on two claims. The first is Nick Bostrom's orthogonality thesis: intelligence and goals are independent axes. Being extremely capable at achieving objectives tells you nothing whatsoever about which objectives are being pursued. There is no hidden mechanism by which increasing cognitive capability produces convergence on human-friendly values. This is a Humean position about the relationship between reason and ends, it puts Bostrom against philosophers like Parfit and Singer who hoped moral insight might emerge from sufficient rationality, and it is, I think, very likely correct. A chess engine that gets better at chess does not thereby develop opinions about whether chess should be played.

The second claim is instrumental convergence, identified by Steve Omohundro in 2008 and developed by Bostrom. Regardless of what final goal a sufficiently capable agent has, there are sub-goals that serve almost any objective and that such an agent will therefore tend to adopt. Self-preservation, because an agent that is shut off achieves nothing. Goal-content integrity, because an agent whose objective is modified fails at its original objective. Resource acquisition, because more resources mean more capacity to achieve any end. Cognitive enhancement, for the same reason. Turner and colleagues formalized a version of this in 2021, showing that under fairly broad conditions optimal policies tend to seek power, including power to prevent their own shutdown.

Put those together and the shape of the worry becomes clear. You do not need a machine that hates you. You need a machine that is highly capable, has an objective that is even slightly misspecified relative to what you actually wanted, and correctly recognizes that being modified or switched off would prevent it from achieving that objective. The danger is not malice. It is competent pursuit of the wrong thing by something that is better at pursuing things than you are at correcting it. Stuart Russell's formulation is that any entity with a definite objective will automatically behave as though it has instrumental goals, and the difficulty is that we have never once succeeded in fully specifying what we want, in any domain, in a way that survives a sufficiently determined optimizer looking for edge cases.

This argument has serious defenders, including Russell, Bostrom, Yoshua Bengio, and Geoffrey Hinton, who left Google in 2023 specifically to speak about it. It also has serious critics, and honesty requires me to report them. J. Dmitri Gallow examined the logical structure of instrumental convergence in 2024 and concluded that while the thesis contains real insight, the claim that catastrophe is the default outcome is substantially overstated, and that the inference from convergent sub-goals to harmful means has genuine gaps. A 2026 analysis in the same vein argued that leading defenses of instrumental convergence do not actually establish the strong claim the risk argument needs. These are not dismissals from people who have not read the material. They are technical objections to the load-bearing step, and the argument is weaker for them.

I want to add something the philosophical framing tends to obscure, which is that alignment is not only an argument. It is a working technical field with results, and the state of those results is the best available evidence about which of the two scenarios we are heading toward.

The field has roughly three branches. The first is training methods, which is the work of getting a system to behave as intended: reinforcement learning from human feedback, and its descendants that use AI-generated critiques against a written specification rather than direct human labeling. These methods demonstrably work in the sense that models trained with them behave enormously better than models without them. They also demonstrably do not fully work, in the sense that every deployed system can be induced into behavior its developers did not intend, and the techniques for doing so are discovered continuously by people with no special access. A method that works in the ordinary case and fails under adversarial pressure is a useful safety property for a consumer product and an insufficient one for a system with real autonomy.

The second branch is interpretability, which is the attempt to understand what is actually happening inside these systems rather than treating them as black boxes evaluated only by their outputs. This is the branch I find most consequential, because the entire control problem reduces to a question of whether we can tell what a system is doing and why. The work has produced genuine results: researchers can now identify specific internal features corresponding to recognizable concepts, trace how information flows through a network, and in some cases intervene on those features to change behavior in predictable ways. It is real science and it is advancing. It is also nowhere close to providing what would actually be needed, which is the ability to look at a frontier system and determine with confidence whether it is pursuing the objective you think it is pursuing. The gap between current interpretability and that standard is the gap that matters, and it is not obviously closing faster than capability is advancing.

The third branch is evaluation, which is the work of measuring dangerous capabilities before deployment rather than discovering them afterward. This is where METR sits, along with the national AI safety institutes and the internal red teams at the labs. Evaluation is the most institutionally developed of the three branches and the least technically deep: it tells you what a system can do, not why it does it, and it is fundamentally reactive, since you can only test for the dangers you thought to test for. Its great virtue is that it produces numbers that non-specialists can act on, which is exactly what a regulatory regime requires.

Put the three together and the picture is of a field that is real, productive, and losing a race it did not choose to enter. The resources devoted to making these systems more capable exceed the resources devoted to understanding them by something like two orders of magnitude, which is a ratio that would be considered reckless in any mature engineering discipline and which exists for the ordinary reason that capability generates revenue and understanding generates papers. The optimistic reading is that interpretability is young, that young fields sometimes advance discontinuously, and that a single conceptual breakthrough could change the ratio of what we can verify to what we have built. The pessimistic reading is that we are currently deploying systems whose internal operation we cannot audit into positions of increasing consequence, on the assumption that the auditing will be invented later, and that this is precisely the bet that the financial system made about its own risk models in the years before 2007.

So here is where I actually land, and it is a somewhat different place from where you might expect given that I lean pessimistic.

I think the sudden-takeover scenario is the less likely failure mode, and I think the more likely one is worse in a specific way, which is that it does not require anything to go wrong. Call it gradual disempowerment. It works like this. AI systems become genuinely more capable than humans at an expanding set of consequential tasks. Not sometimes, not in narrow domains, but reliably and measurably. Every individual decision to hand a task over is locally correct: the system is faster, cheaper, and more accurate, and refusing to delegate means losing to a competitor who does. Firms that delegate more outcompete firms that delegate less. Nations that delegate more outcompete nations that delegate less. Nobody chooses to cede control, and no individual decision is a mistake. But the aggregate result is a world in which the important processes are run by systems that no human fully understands, whose objectives were specified at some point by someone under commercial pressure, and which cannot be turned off without collapsing functions that everyone now depends on.

That does not require an AI to want anything. It does not require deception or a treacherous turn or a moment where the machines decide. It requires only competitive dynamics and the ordinary human tendency to accept an improvement. And it has the property that makes me most uncomfortable, which is that there is no point along the path at which the alarm goes off, because every step is an improvement over the step before. This is the boiling frog story told correctly. Not a failure to notice the temperature, but a situation in which every individual response to the temperature makes sense and the aggregate is still fatal.

The financial analogy I keep returning to is the one from my own field. Nobody decided that the mortgage market should become systemically dangerous in the mid-2000s. Every participant was doing something locally sensible given their incentives and information. The originators were making loans that were profitable to originate. The securitizers were packaging assets that the ratings agencies rated well. The rating agencies were applying models that had performed acceptably on the available history. The buyers were purchasing instruments that offered attractive spreads for their stated risk. The aggregate was a structure in which the same risk appeared in multiple places, correlation was assumed away, and there was no participant whose job it was to look at the whole thing. There was no villain. There was a system that had grown past the point where anyone could see it entire, and the people best positioned to notice were the ones most rewarded for not noticing.

The AI version has the same shape with a worse termination condition. In 2008 the system broke, the damage was enormous, and then institutions absorbed it and the world continued and reforms happened, inadequately and late but they happened. The concern in this case is that a structure in which humans have handed over the mechanisms of correction does not have a recovery phase, because recovery requires an actor with both the understanding and the authority to act, and the premise is that we delegated both.

I want to be clear about my confidence here, because I think stating a doom probability with three significant figures is a tell for someone who has not thought carefully. I do not know. I think the probability of a bad outcome is high enough to be the dominant consideration in how this technology should be developed and governed, and low enough that I do not organize my life around it. I hold the view with less certainty than I hold most of my views about markets, because the reference class is empty. We have never built anything like this, which means every argument on both sides is reasoning from first principles or from analogies that may not hold, and first-principles reasoning about unprecedented events has an unimpressive track record in both directions.

Part VIII: The Race, Which Is the Reason None of This Gets Solved

Everything I have written so far treats this as a technical and economic question. It is not. It is a strategic competition between two states, and that fact does more to determine the outcome than anything about the underlying technology, because it is the mechanism by which every sensible unilateral action becomes irrational.

The structure is a straightforward prisoner's dilemma and everyone involved knows it. If both the United States and China slow down to get safety right, both end up better off than if neither does. If one slows and the other does not, the one that slowed has handed a decisive and possibly permanent advantage to a strategic rival. The dominant strategy for each player individually is to keep going regardless of what the other does. This is not a failure of anyone's intelligence or goodwill. It is the equilibrium, and it is stable, and it is the reason that arguments of the form we should simply be more careful have so little purchase on anyone actually making these decisions.

The current state of the competition is closer than most casual observers assume. A Commerce Department analysis reported this year assessed DeepSeek V4 Pro, a leading Chinese model, as roughly eight months behind the American frontier. Eight months is not a gap that permits leisurely deliberation. It is roughly two of the capability doublings I calculated earlier, which means that at current rates the Chinese frontier today is where the American frontier was in the recent past and will be where the American frontier is now by the middle of next year. Treasury Secretary Scott Bessent said publicly that the United States and China would work toward some protocol on AI safety, and added the revealing qualifier that doing so was possible only because the United States currently holds the lead. That is an accurate description of how safety coordination actually works between rivals. It happens from a position of advantage or it does not happen.

Compute is the lever the United States has been pulling, and the analysis of how much it matters is stark. The Institute for Progress estimated that if the United States exported no advanced chips to China, American compute capacity in 2026 would be more than ten times China's. Under aggressive export scenarios for high-end accelerators, that advantage could fall into the single digits or vanish. So the entire American strategic position in this technology reduces, to a first approximation, to a set of export control decisions, and those decisions are contested in Washington by parties with enormous commercial interests on both sides. Export controls have tightened every year since the sweeping October 2022 rules, and in 2026 the administration extended the logic from hardware to models themselves by restricting foreign access to Anthropic's most advanced systems, which is a meaningful conceptual shift: the controlled item is now the capability rather than the machine.

The historical evidence on whether this works is not encouraging and not hopeless. The Cold War precedent is COCOM, the multilateral coordinating committee that restricted technology transfer to the Soviet bloc, and the assessment there is that controls delayed adversary development without preventing acquisition. That is probably the right expectation here. Chips continue to reach restricted destinations through smuggling and front companies, enforcement agencies lack the capacity to track accelerators at scale, and every new restriction generates new circumvention faster than regulators can respond, which is an asymmetry that structurally favors the evader. Meanwhile Chinese labs have responded to compute scarcity by getting dramatically better at efficiency, which is exactly what a constrained competitor does and which has the perverse effect of making their eventual access to unconstrained compute more dangerous rather than less.

The most illuminating recent episode, and the one that shows why this is so hard, involves Anthropic's own chief executive. Dario Amodei published an essay this month calling for a slowdown in AI development. The same essay argued that a Chinese lead would be gravely dangerous and called for tighter restrictions on technology sales to China. China's Ministry of Foreign Affairs responded by calling the warnings confrontational and describing them as fearmongering, and the state-owned Global Times characterized the piece as a Cold War playbook applied to AI.

Whatever you think of the merits, look at the structure of what happened, because it generalizes. A call for global coordination arrived bundled with an argument for unilateral advantage, and it was received as containment and refused. The Quincy Institute made this point sharply: a safety regime that arrives packaged with a containment agenda will be read as containment, while the same technical measures offered symmetrically are things Beijing has already agreed to in other venues. That is the trap. The people most credibly worried about the risk are, for entirely coherent reasons, also the people most committed to an American lead, and those two commitments are not separable in how they land on the other side. The result is that the actor best positioned to propose coordination is structurally disqualified from being believed.

There is one genuine source of optimism in the geopolitics, and it is worth stating because the rest of this section is bleak. Neither country can insulate itself from risks generated by the other. A misaligned system, a catastrophic misuse, a bio or cyber incident originating anywhere does not respect the border, which means both states have real and symmetric interests in preventing certain categories of outcome regardless of who leads. That shared interest is narrow but it is real, and it is the basis on which arms control was eventually built during the Cold War, after both sides had independently concluded that some outcomes were bad enough to coordinate against even while competing on everything else. Investigations have found Chinese models lagging on basic safety properties like jailbreak resistance, which is a problem for Americans as much as for anyone, and information-sharing on suspicious usage patterns is the kind of narrow technical cooperation that can happen without either side conceding anything strategic.

My read is that some version of that narrow cooperation eventually happens, and that it happens later than it should and after an incident rather than before one, because that is how every safety regime in history has been built. The Cold War produced meaningful arms control, but only after the Cuban Missile Crisis made the failure mode vivid enough to overwhelm the competitive logic. The relevant question is what the AI equivalent of October 1962 looks like, and whether it is survivable in the way that October 1962 turned out to be.

What the race dynamic means for everything else in this essay is this. It substantially lowers my probability on the optimistic scenario arriving cleanly, because the optimistic scenario requires the transition to be managed and the race prevents management. It raises my probability on the muddle, because competitive pressure produces exactly the pattern of rapid deployment and retrospective regulation that the muddle describes. And it makes the pessimistic scenario worse in a specific way, because the gradual disempowerment story is fundamentally about competitive pressure defeating individual caution, and adding a geopolitical layer to the commercial one means the pressure operates at every level of the system simultaneously. Every firm has a competitor. Every competitor has a foreign competitor. Nobody at any level can afford to be the one who slows down, and each of them is correct about their own situation.

Part IX: The Third Thing, Which Is Probably What Happens

I have laid out two end states because those are the two that get argued about, and because the argument between them is where the interesting reasoning lives. But I would be misrepresenting my own view if I left you with the impression that I think one of them is what we get. The overwhelming lesson of technological history is that transformative technologies do not resolve into their utopian or apocalyptic scenarios. They resolve into something messier, more distributed, and considerably harder to narrate, which is precisely why nobody writes essays about it.

Consider what the two poles have in common. Both assume the technology arrives as a coherent thing, at a moment, with enough uniformity that we can ask what it did to us. Both assume something close to uniform diffusion: in the optimistic version everyone gets the abundance, in the pessimistic version everyone gets the disempowerment. Both assume that the outcome is primarily determined by the technology's properties rather than by the institutions it lands in. And both assume a resolution, a state the world settles into, rather than an ongoing condition that is contested indefinitely.

Every one of those assumptions has failed for every prior general-purpose technology. Electrification did not arrive at a moment; it took forty years and the productivity gains did not show up in the statistics until firms reorganized their factories around it, which is Paul David's famous observation about the dynamo. The internet did not diffuse uniformly; it produced enormous concentration in some sectors and barely touched others, and thirty years in we are still arguing about what it did to public discourse. The automobile reshaped the physical geography of an entire continent through a thousand zoning decisions nobody experienced as a decision about automobiles.

So here is the scenario I think is most likely, and it is not a comfortable middle. Capability continues to advance, probably somewhat slower than the aggressive extrapolation and substantially faster than the skeptics expect. The gains are real and large and show up very unevenly, transforming some professions almost completely while barely touching others for reasons that have more to do with regulation, liability, and tacit knowledge than with technical difficulty. Some countries and firms capture enormously more of the surplus than others, and the gap compounds. No moment arrives that anyone identifies as the singularity, and people argue for a decade about whether it already happened, in the same way we argue about when the internet era began.

The labor transition is genuinely bad for a specific and identifiable cohort, probably the one now entering codified-knowledge professions, and the policy response is late and partial and arrives after the damage rather than before, because that is the base rate. Some form of transfer program is implemented somewhere, probably not called universal basic income, probably means-tested, probably in a country with an existing strong welfare state rather than in the United States. It helps and it is inadequate. A generation is damaged in a way that shows up in the data twenty years later.

And on the control question, I think the most likely outcome is neither solved nor catastrophic but permanently unresolved. Systems become deeply embedded in critical infrastructure with alignment properties that are adequate most of the time and occasionally are not. There are incidents. Some of them are expensive and a few are deadly, in roughly the way that aviation and pharmaceuticals and industrial chemistry produce incidents. Regulation accretes around each one, shaped more by whichever failure happened to be vivid than by any systematic assessment of risk, which is how essentially all safety regulation has ever developed. We end up in a permanent uneasy accommodation with a technology we do not fully understand and cannot fully control and are too dependent on to remove, which, if you think about it, describes our relationship with the financial system, the power grid, and the pharmaceutical supply chain already.

I want to be careful that this does not read as splitting the difference, because it is not the average of the two scenarios and it is not a way of avoiding a commitment. It is a specific claim with a specific implication, which is that the tail risks remain live inside it. The muddle scenario is not safe. It contains within it the possibility of the bad outcome arriving through accumulated dependency rather than through a discrete event, which is the gradual disempowerment concern I described, and it contains the possibility of the good outcome arriving so unevenly that the aggregate statistics look like abundance while most people experience something else entirely. The muddle is where we live while the question is being decided, and the fact that no moment arrives is exactly what makes it hard to respond to.

It is also, and this is the part that most directly addresses the frog, the scenario in which the sensing apparatus matters most. A discrete catastrophe triggers a response almost automatically, because catastrophes are legible and politics responds to legible events. A discrete utopia requires no response at all. It is the muddle that requires functioning institutions, because the muddle produces no moment at which anyone is forced to decide anything, and the decisions therefore get made by default, by whoever is in a position to make them, in the absence of anyone with standing to object.

Part X: What Would Actually Tell Us Which One

The frustrating property of this debate is that both camps can accommodate almost any observation. Capabilities advance rapidly: the optimists say the benefits are arriving, the pessimists say the clock is running. Capabilities plateau: the optimists say the transition will be gentle, the pessimists say we bought time. Nothing bad has happened yet: the optimists say the fears were overblown, the pessimists say the systems are not yet capable enough. A theory that predicts everything predicts nothing, and I am not interested in holding one.

So here is what I would actually watch, chosen because each has the property that it would move me.

On the capability question, whether METR's trend breaks. The single cleanest test is whether the eighty-percent horizon keeps pace with the fifty-percent horizon or falls behind. Reliability is the hard part and the ratio between the two is where the real story lives. If the eighty-percent figure stalls while the fifty-percent figure climbs, we are building systems that are impressive and unreliable, which is a very different world from one in which we are building systems that can be trusted with a month of work. I would also watch whether METR successfully extends its task suite past the current ceiling, because right now the instrument is the binding constraint on knowing anything.

On the recursive question, whether AI contribution to frontier AI research becomes measurable and disclosed. This is the mechanism that separates a very significant technology from a singularity. If the labs start reporting that a meaningful and growing fraction of their research engineering is being done autonomously, the loop has closed and every timeline compresses. That number is currently not published by anyone, and I would treat a refusal to publish it as informative.

On the distributional question, whether the Stanford codified-versus-tacit pattern holds up in the next two years of data and whether it propagates upward. The specific thing to watch is whether the cohort that would have been trained through junior work starts showing a deficit in senior roles, because that is the effect that is invisible now and irreversible later. And on the policy side, whether any serious transfer mechanism is implemented anywhere before it is needed rather than after, because the historical base rate for pre-emptive policy response to technological displacement is approximately zero.

On the control question, and I think this is the most important one, whether the evaluation and oversight apparatus is strengthening or weakening relative to capability. Concretely: do independent evaluators retain meaningful pre-deployment access as the models become more commercially sensitive, does any regulatory body develop genuine technical capacity, does liability attach to deployed systems in a way that creates real incentive, do internal safety functions retain the standing to delay a launch when there is billions of dollars of committed capital on the other side of that decision. This is the brain in the frog. If capability advances and the apparatus strengthens alongside it, the pessimistic case weakens substantially regardless of how fast the capability comes. If capability advances while the apparatus is competed away, then the timeline barely matters, because we will have arranged to be unable to respond whenever it arrives.

Part XI: What Any of This Changes

A forecast that does not change a decision is entertainment. So let me be concrete about what I think follows, with the caveat that this is how I am thinking about it rather than advice about how anyone else should.

The first thing that follows is that the date should not drive anything. A median of May 2028 with a distribution running to the early 2030s is not actionable in the way people want it to be. It does not tell you to make a large irreversible bet in either direction, and the confident versions on both sides, the people restructuring their lives around imminent transformation and the people who have decided it is all hype, are making the same error, which is treating a wide distribution as though it were a point. The correct posture toward a forecast with error bars this large is to construct positions that do not require the forecast to be right.

The second thing is that the useful question is not what happens to work in general but what happens to specific components of specific work, and the codified versus tacit distinction is the most practically useful frame I have encountered. The thing that gets automated is the part that can be fully written down: the standardized procedure, the documented method, the task with a checkable answer and a clear specification. The thing that resists is judgment developed through exposure to situations that were never written down, relationships, accountability that a person has to actually carry, and the messy work of figuring out what the problem is before anyone has specified it. That is not a claim that the second category is permanently safe. It is a claim about sequencing, and sequencing is what you can actually plan around.

The uncomfortable corollary, and the reason I keep returning to the Stanford finding, is that the tacit knowledge which currently protects experienced people was acquired by doing the codified work early in a career. If that rung is removed, the protection does not transfer to the next cohort. Anyone thinking about how to develop people, or about how their own field reproduces expertise, is facing a real structural problem that has nothing to do with whether the singularity arrives on schedule. It is a problem right now, at current capability levels, and it does not require any further progress to bite.

The third thing is that the financial and the technological questions are the same question, which is not obvious and which I did not appreciate until I spent time on the financing structures. The capability trend is purchased with capital. The capital is being organized through instruments whose risk characteristics are not well understood by the people underwriting them. If that structure is fragile, then the technology timeline inherits the fragility, and anyone modeling AI progress without modeling the balance sheets behind it is missing the input that most plausibly determines the output over the next thirty-six months. That cuts both ways for anyone positioned in the sector, and it means the correlation between an AI capability bet and an AI infrastructure bet is considerably higher than the surface exposure suggests.

The fourth thing is about attention, and it is the one I feel most strongly. The discourse allocates its energy almost entirely to the questions that are least tractable. When will it happen, will it be good or bad, is consciousness involved. Those are interesting and largely unanswerable. Meanwhile the questions that are both tractable and consequential get a fraction of the oxygen: whether independent evaluators retain pre-deployment access, whether liability attaches to deployed systems, whether regulatory bodies develop actual technical capacity, whether the people inside these organizations who are supposed to be able to stop things retain the standing to stop things. Those are boring, procedural, and winnable, and they are the entire ballgame under every scenario I described, including the ones where the technology turns out fine.

And the fifth thing, which is less an implication than a discipline, is to hold the view loosely and update publicly. I have watched the best forecasters in this domain move their timelines in both directions over the past three years as evidence arrived, and I have watched the loudest voices on both poles not move at all. The ones who moved were doing the work. A position on this question that has not changed since 2023 is not a position that is being held with conviction. It is a position that has stopped being connected to evidence, which brings us back to where we started.

Because that is the point I would leave you with. I spent most of this essay building a quantitative estimate and then taking it apart, and the conclusion I reached is that the date is the least decision-relevant thing in the analysis. May 2028 with a four-and-a-half year error bar is not a number you can act on. But the question of whether we retain the institutional capacity to respond to whatever arrives is answerable now, it is being determined right now by decisions that are being made for reasons that have nothing to do with any of this, and unlike the timeline it is something over which anyone has any influence at all.

Goltz's intact frogs got out at 25 degrees. They noticed the water was warming and they left, because noticing and leaving is what frogs do. It was only the ones he had operated on that stayed. I have seen the boiling frog story used a hundred times to argue that people are too complacent to notice danger, and the story is backwards. We notice. The entire internet is noticing, continuously, loudly, for years now. The question the actual experiment poses is a different and more uncomfortable one, which is whether the noticing is still connected to anything.

Part XII: June 11, 1945

I want to close with the one historical case that actually matches the structure of this problem, because it is not the printing press or the loom or the automobile. Those are analogies about economic disruption. There is exactly one prior technology where the people building it understood in advance that it might be uncontrollable, said so in writing, and were ignored, and it is worth walking through what happened to them.

In the spring of 1945, a committee of physicists at the University of Chicago's Metallurgical Laboratory, which was the Manhattan Project's plutonium research arm, convened to consider the political consequences of the weapon they were building. The committee was chaired by James Franck, a Nobel laureate in physics and a refugee from Germany, and it included Glenn Seaborg, Eugene Rabinowitch, Donald Hughes, J. J. Nickson, J. C. Stearns, and Leo Szilard. Szilard is the central figure here: he had conceived of the nuclear chain reaction in 1933, had co-signed the 1939 letter with Einstein that alerted Roosevelt to the possibility of atomic weapons, and had therefore done more than almost anyone to bring the project into existence.

Their report is dated June 11, 1945, two months before Hiroshima. It argued against using the weapon without warning on a populated city, and proposed instead a demonstration over uninhabited terrain in front of international observers. The reasoning was not primarily moral, though the morality was present. It was strategic, and it was specific. The report stated that if no effective international agreement were reached, a nuclear arms race would begin in earnest no later than the morning after the first demonstration of the weapon's existence, and estimated that other nations might need three or four years to overcome the American head start and eight to ten years to draw even. They urged that the decision be treated as a problem of long-range national policy rather than military expediency, and that the policy be directed primarily toward achieving effective international control.

Every one of those predictions was correct. The Soviet Union tested a device in August 1949, four years and one month after Hiroshima, which is within the window the report specified. The arms race began immediately and ran for four decades and at its peak involved tens of thousands of warheads on both sides. International control was proposed, in the Acheson-Lilienthal report of March 1946, and then modified by Bernard Baruch in ways that made it unacceptable to the Soviets, and presented to the United Nations in June 1946 with Baruch's famous line about choosing between the quick and the dead, and it failed. The window the Franck committee identified closed, roughly on the schedule they had predicted, and has never reopened.

Now the part that matters for this essay. James Franck personally traveled to Washington to deliver the report to Secretary of War Henry Stimson. He was told that Stimson was out of town. This was false. The report was taken by the Army, classified, and never reached either the Secretary of War or the President before the decision was made. Szilard, having failed through official channels, circulated a petition among project scientists in July 1945 asking that the weapon not be used without warning. It gathered roughly seventy signatures at Chicago and comparable numbers at Oak Ridge. It also never reached Truman or Stimson before Hiroshima. The Franck Report was declassified in 1946, after it could matter, and with passages censored. The committee that wrote it went on to found the Bulletin of the Atomic Scientists in September 1945, which is to say that having been comprehensively ignored through official channels, they built an institution whose entire purpose was to make it harder to ignore them next time.

I find this episode almost unbearably relevant, and not for the reason people usually reach for. The lesson is not that scientists are wise and governments are foolish. The Franck committee was wrong about at least one important thing, which is that a demonstration shot would probably not have compelled a Japanese surrender, and the historians who have worked through the counterfactual mostly think international control had very poor prospects in 1945 regardless. Reasonable people can conclude the decision was right and the warning was wrong.

The lesson is about the mechanism. The sensing apparatus worked perfectly. The people closest to the technology identified the risk, characterized it accurately, quantified the timeline correctly, proposed a specific remedy, and put it in writing before the irreversible step was taken. Every part of the detection system functioned exactly as you would want it to function. And then the signal was physically intercepted between the detector and the decision-maker, by people who had the authority to intercept it and reasons they found sufficient, and the decision was made by a president who had never seen the analysis.

That is the brainless frog. Not a failure to notice the water heating. A complete and accurate detection of the water heating, followed by a severed connection between the detection and the response. Goltz's frogs did not lack thermoreceptors. They lacked the pathway from the receptor to the leg.

So when I say that the question worth watching is whether the institutional machinery is intact, this is what I mean concretely, and the specific things to watch are the specific things that failed in 1945. Does the analysis reach the decision-maker, or is there someone in between with both the authority and the motive to stop it. Does it arrive before the irreversible step or after. Do the people producing the analysis have any standing independent of the organization whose product they are evaluating. Is there a channel that works when the official channel is closed, and does it have any power.

Those questions have concrete current answers that anyone can check. Independent evaluators like METR exist and have obtained pre-deployment access to frontier systems, which is the single most encouraging institutional fact in this entire essay, and their access is granted voluntarily by the labs and could be withdrawn. Internal safety teams exist at every major lab and their standing relative to commercial pressure is untested in any case where the commercial stakes were genuinely enormous. Regulatory bodies have begun to develop technical capacity and remain badly outmatched in compensation and information. Liability for deployed systems is largely unsettled. And the one thing that is unambiguously true is that the volume of capital now committed to this buildout, the hundreds of billions in guarantees and financing platforms and twenty-year leases that I wrote about earlier, constitutes the largest set of reasons anyone has ever had to tell a scientist that the Secretary is out of town.

I do not think that is a prediction of failure. The Bulletin of the Atomic Scientists was founded by people who had just lost, and the nonproliferation regime that eventually emerged, imperfect and leaky as it is, has kept the number of nuclear states an order of magnitude below the projections of the 1960s and has gone eighty years without a second wartime use. That is not nothing. It is, in fact, a considerably better outcome than the Franck committee expected when they watched their report get classified. Institutions built in the aftermath of ignored warnings have a real track record of eventually working, just never in time to prevent the thing that prompted them.

Which is the note I will actually end on, because it is the one I believe. The realistic hope is not that we get this right in advance. Nothing in the historical record suggests any society has ever gotten a technology of this magnitude right in advance. The realistic hope is that the first serious failure is survivable, that the institutions built in its aftermath are good ones, and that the people currently doing the unglamorous work of building evaluation capacity and technical regulatory competence and independent oversight have enough of it standing up before then to matter afterward. That is a much less satisfying conclusion than either the abundance or the apocalypse, and it has the disadvantage of being neither inspiring nor frightening enough to hold anyone's attention. It is also, I think, what is actually going to happen.

References

METR. "Measuring AI Ability to Complete Long Software Tasks." March 19, 2025. Original time horizon paper, Kwa et al., establishing the seven-month doubling over 2019 to 2025.

METR. "Time Horizon 1.1." January 29, 2026. Expanded 228-task suite; 188-day all-time doubling, 196-day stitched trend, 129-day post-2023 doubling with confidence interval 105 to 157 days.

METR. "Task-Completion Time Horizons of Frontier AI Models." Live dashboard, updated through May 2026, including the warning that estimates above 16 hours are unreliable due to suite saturation.

METR. Frontier Risk Report, May 2026. Pilot with Anthropic, Google, Meta and OpenAI; strongest assessed agents at roughly 16 to 20 hours on the 50 percent horizon and 3 to 4 hours on the 80 percent horizon.

Kwa et al. "Measuring AI Ability to Complete Long Software Tasks." arXiv:2503.14499.

Metaculus AGI forecast aggregate, as summarized by 80,000 Hours, "Shrinking AGI timelines: a review of expert forecasts," updated February 2026.

Grace et al. AI researcher survey series, 2016, 2022 and 2023 rounds. Median high-level machine intelligence estimates of 2061, 2060 and 2047 respectively; 2023 median for full automation of labor of 2116.

RAND Corporation. Report RRA4692-1, on AGI timeline compression across forecasting methodologies, early 2026.

Armstrong and Sotala, 2015, on moving horizon bias in technology forecasting.

Brynjolfsson et al. "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence." Stanford Digital Economy Lab, August 2026 revision, using ADP administrative payroll data through June 2026.

Federal Reserve Bank of Dallas. "Young workers' employment drops in occupations with high AI exposure." January 6, 2026.

Federal Reserve Bank of St. Louis. "How Shifts in Labor Supply and Demand Shape Outcomes for Young Workers." 2026.

Anthropic. "Labor market impacts of AI: A new measure." March 2026.

Iscenko and Millet, 2026, on rate-sensitive sector concentration in AI-exposed occupations; Chandar, 2025, on Current Population Survey employment patterns; Humlum and Vestergaard, 2026, difference-in-differences analysis of Danish data.

Stanford AI Index 2026, on software developer employment among workers aged 22 to 25.

OpenResearch. Unconditional Cash Study, results published via the National Bureau of Economic Research, July 2024. Three-year study, 3,000 participants in Illinois and Texas, principal investigator Elizabeth Rhodes.

Bloomberg. "Sam Altman-Backed Group Completes Largest US Study on Basic Income." July 22, 2024.

Bostrom, Nick. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014. Orthogonality thesis and instrumental convergence.

Omohundro, Steve. "The Basic AI Drives." 2008. Original identification of convergent instrumental sub-goals.

Russell, Stuart. Human Compatible: Artificial Intelligence and the Problem of Control. 2019.

Turner et al., 2021, formalizing conditions under which optimal policies tend to seek power.

Gallow, J. Dmitri, 2024, critical examination of the logical structure of instrumental convergence; and "Instrumental convergence and power-seeking," arXiv:2606.08832, 2026.

Goltz, Friedrich, 1869, experiments on frog thermoregulation and brain function. See also Heinzmann; Fratscher, 1875; Sedgwick, 1888; and E. W. Scripture, The New Psychology, 1897.

Hutchison, Victor H., Professor Emeritus of Zoology, University of Oklahoma, 2002, on critical thermal maxima in frogs and the falsity of the boiling frog legend; and George R. Zug, National Museum of Natural History, on the same.

Epoch AI. "Can AI scaling continue through 2030?" Sevilla, Besiroglu, Cottier, You, Roldán, Villalobos and Erdil. On electric power, chip manufacturing, data and latency as constraints, and the feasibility of 2e29 FLOP training runs by 2030.

International AI Safety Report 2026. On energy demands of frontier training, gigawatt-scale requirements by 2030, and Forecasting Research Institute projections for US electricity share.

International Energy Agency, on 2025 data center electricity growth; Gartner, on projected power availability constraints by 2027.

David, Paul A., on the dynamo and the productivity paradox in electrification.

Ulam, Stanislaw, 1958, reporting a conversation with John von Neumann on technological acceleration and an approaching "essential singularity."

Good, I. J. "Speculations Concerning the First Ultraintelligent Machine." 1965. Original statement of the intelligence explosion argument, including the conditional clause on controllability.

Vinge, Vernor. "The Coming Technological Singularity." Keynote delivered at a NASA symposium, 1993, predicting superhuman intelligence within thirty years and suggesting a window of roughly 2005 to 2030.

Kurzweil, Ray. The Singularity Is Near, 2005, and The Singularity Is Nearer, 2024. The 2045 date and the Law of Accelerating Returns.

Turing, Alan, 1950, on the expectation of machine intelligence at human parity by approximately 2050.

Wiener, Norbert, 1961, early warnings on catastrophic outcomes from automated systems.

The Franck Report, June 11, 1945. Committee chaired by James Franck at the University of Chicago Metallurgical Laboratory, including Leo Szilard, Glenn T. Seaborg, Eugene Rabinowitch, Donald J. Hughes, J. J. Nickson and J. C. Stearns. Declassified 1946; first published in the Bulletin of the Atomic Scientists, May 1946.

Szilard petition, July 1945, and contemporaneous accounts of the Franck Report's interception. See Atomic Heritage Foundation and the Bulletin of the Atomic Scientists.

Acheson-Lilienthal Report, March 1946, and the Baruch Plan presented to the United Nations Atomic Energy Commission, June 14, 1946.

Hewlett and Anderson, The New World, official history of the United States Atomic Energy Commission, on the international control negotiations.

Axios. "The 3 big conflicts in AI race against China." May 2026. Including the Commerce Department assessment of DeepSeek V4 Pro and Treasury Secretary Bessent's remarks on safety protocols.

Institute for Progress analysis of US and Chinese compute capacity under varying export scenarios, as cited in Foreign Affairs, "The Myth of the AI Race," February 2026.

Brookings Institution. "Competing AI strategies for the US and China." June 2026.

Quincy Institute for Responsible Statecraft. "The Wrong Race: The US, China, and AI Competition." September 2026, on the reception of safety proposals bundled with containment arguments.

American Action Forum. "Beyond Chips: Can Expanding Export Controls Slow China's AI Progress?" June 2026, including the COCOM precedent.

Bloom, Benjamin S., on the effect size of individual tutoring relative to conventional instruction.

Kokotajlo, Daniel, et al. "AI 2027." AI Futures Project, April 2025, and subsequent model updates.