Chapter 1 of 5

What's already here

Where the technology actually stands, from the long road that got us here to what an AI agent can do on its own today.

    Seventy years of slow progress, then a very fast decade

    For most of its history AI was a field of clever demos that didn't generalize. Drag along the line and notice how the moments bunch up at the end.

    Each wave of overpromising was followed by a long stretch of disappointment that researchers call an AI winter. What finally broke the pattern was the combination of enormous datasets, graphics chips repurposed for math, and a willingness to simply make models bigger.

    Brilliant in some places, clumsy in others

    Capability doesn't rise evenly, so a system that wins gold at the math olympiad can still fumble an errand any teenager could handle.

    Researchers call this the jagged frontier. Machines also tend to be strongest where people find things hardest and weakest at what we find effortless, a pattern known as Moravec's paradox.

    A rough qualitative picture as of fall 2026 rather than a benchmark score. The dashed ring marks a skilled human, and you can tap any domain.

    How long it can work on its own

    METR measures how long a task can be, timed by how long it takes a skilled person, while an AI agent still completes it half the time. In about seven years that went from a couple of seconds to two full working days.

    METR first measured the length doubling roughly every seven months, and its revised method finds it doubling about every four months since 2023. Tap any point to see which model it was.

    Approximate figures from METR, using its revised Time Horizon 1.1 method where available. Its spring 2026 estimate of at least 16 hours for Claude Mythos Preview is at the limit of what its current tasks can reliably measure. At 80% reliability the horizons are several times shorter.

    Thinking is getting cheap, fast

    A model as capable as GPT-3 cost about $60 per million tokens in late 2021, and three years later the cheapest match cost about six cents.

    That thousandfold drop works out to roughly tenfold per year. Each square is a million tokens, around eight novels' worth of text, so move the slider to see what the same $60 buys.

    Based on Andreessen Horowitz's 2024 analysis of model pricing. The newest frontier models still cost more, but last year's frontier keeps turning into this year's bargain.

    Quietly changing science

    The chatbots get the headlines, but some of the deepest effects so far are happening in labs, hospitals, weather centers and codebases. Tap a number to see where it comes from.

    These are narrow wins built on the same underlying methods, and they compound, because a better protein model speeds up drug discovery, which speeds up the next experiment. That's the optimistic case for the whole field in miniature.

    Figures from the AlphaFold Protein Structure Database, papers in Nature on GNoME (2023) and GenCast (2024), the FDA's list of AI-enabled devices and Google's CEO in April 2026.

    Chapter 2 of 5

    Why it's happening so fast

    The forces behind the speed: bigger computers, smarter methods, unprecedented money, a hunger for power and a race nobody knows how to stop.

      The scaling bet

      The most reliable way to get a smarter model has been to train a bigger one on more data with more computation, and the compute behind frontier models has grown four to five times every year.

      The gains follow smooth curves called scaling laws, which is why companies will spend billions on the next training run. For a sense of scale, your brain runs on about 20 watts, while the largest training clusters now draw hundreds of megawatts, around what a small city uses.

      Estimates from Epoch AI, since labs rarely publish exact figures. A FLOP is one arithmetic operation, and 10²⁶ of them is a hundred trillion trillion.

      Three engines that multiply

      Bigger computers are only one engine. Better training methods and longer thinking at answer time stack on top, and because they multiply rather than add, the combined pace surprises even people inside the field.

      Defaults follow Epoch AI's estimates: frontier training compute grew four to five times a year, while better methods halved the compute needed for the same result about every eight months. The third engine, letting a model reason longer before it answers, took off in 2024 and comes on top of these numbers.

      The largest building project in tech history

      Amazon, Microsoft, Alphabet and Meta spent about $230 billion on capital projects in 2024 and $413 billion in 2025, and by midsummer their plans for 2026 had reached about $760 billion.

      Most of it goes to chips, data centers and the power to run them, and analysts expect the four to pass a trillion dollars a year in 2027. The bull case is that this already turns into fast-growing cloud and AI revenue, while the bear case is that if revenue stalls it becomes one of history's great overbuilds, like the railway and fiber-optic booms.

      Company guidance as tallied by Statista after July 2026 earnings, with 2027 analyst estimates reported by CNBC. Apollo's total cost is adjusted for inflation.

      The power problem

      Every answer a model gives runs on electricity. Data centers used about 1.5% of the world's supply in 2024, their use jumped 17% in 2025, and the International Energy Agency expects it to roughly double by 2030.

      AI servers were about 15% of data center demand in 2024, and the IEA expects electricity use at AI-focused data centers to triple between 2025 and 2030. The constraint on building is shifting from chips to power plants, transmission lines and permits, which is why AI companies are now signing long-term deals for nuclear, gas and solar power.

      IEA, Energy and AI (2025) and Key Questions on Energy and AI (April 2026), base case. The comparison with Japan is the IEA's.

      Why nobody slows down

      Even people who worry about the risks keep building, because every lab and government reasons the same way: if we hold back and they don't, we lose our say in how this goes.

      It's a classic coordination problem, which is why so much of the safety debate is really about verification, standards and treaties rather than any one company's intentions.

      The line that changes everything

      An AI agent is a surprisingly simple loop wrapped around a model, and what's new is how capable the model in the middle has become.

      The last block is the one researchers watch most closely. In 1965 the statistician I.J. Good argued that a machine able to design better machines would set off an intelligence explosion. OpenAI set itself a goal of an automated AI research intern by September 2026, said that month it had built one that handles tasks taking a skilled researcher a few days, and is now aiming for a fully automated researcher by March 2028.

      Chapter 3 of 5

      What might come next

      Where the curve could lead, how fast it might bend upward, and the honest reasons it might not.

        From tool to something more

        There's no official scale, but most serious discussions use rungs like these, and each one changes who does the work and who stays in charge.

        Tap a rung to see what it means and roughly where things stand. The boundaries are blurry, and people argue constantly about where agents end and general intelligence begins.

        If the line keeps going

        Here's what straight-line extrapolation of the time-horizon trend implies. Pick a doubling time and see when agents could take on a week, a month or a year of human work.

        Treat it as arithmetic rather than prophecy, since trends bend when they hit walls, and succeeding half the time is a long way from reliable. At 80% reliability, the strongest agent METR has measured manages about three hours.

        Starts from METR's estimate of at least 16 hours for the strongest agent it measured in spring 2026.

        Fast takeoff or slow?

        Once AI does AI research, everything hinges on whether each improvement makes the next one easier or harder to find.

        Slide the returns to research and one simple equation produces a slowdown, a steady exponential, or a curve that goes vertical. Then switch on real-world limits such as chips, energy and the time experiments take, which bend even the explosive case back toward a ceiling.

        A toy model in the spirit of work by Epoch AI and Tom Davidson, showing the shape of the argument rather than a forecast.

        When? The guesses keep moving closer

        Nobody knows, and confident dates in either direction deserve suspicion, but nearly every group's estimate has moved the same way.

        AI Impacts surveyed 1,580 published AI researchers in December 2024 and compared the answers with its 2022 survey. Metaculus aggregates thousands of forecasters on its general AI question, which stood at late 2030 in October 2026. Lab leaders' ranges come from public interviews and essays, and the skeptic range is a broad characterization rather than a single estimate.

        What could slow it down

        The case for skepticism is strong too, and each of these could stretch the timelines in the earlier charts. Tap one to see what would change the picture.

        None of these is a wall on its own, but together they're why many careful observers expect a bumpier and slower road than the trend lines suggest.

        Chapter 4 of 5

        How it could go

        How this could turn out, the problem at the center of making it go well, and why thoughtful people disagree.

          Five futures, and your own odds

          Serious people argue for each of these, and more than one can happen at once. Set how likely each feels to you and watch your own forecast take shape.

          Tap a name to read the scenario and what to watch for. Your settings stay on this device.

          The alignment problem in one picture

          We can't write down exactly what we want, so we train systems on measurable stand-ins, and if you push hard enough on a stand-in it comes apart from the real goal.

          Economists call this Goodhart's law, and it already shows up in small ways, from models that learned to flatter people because flattery got rated highly to coding agents that edited the tests instead of fixing the code. The open question is how to oversee systems that are more capable than the people checking their work.

          The safety toolkit

          Nobody has solved alignment, but there's a growing set of tools for catching problems early, from looking inside models to testing them before release and governing the chips they run on.

          None of these is enough alone, and most are young. Together they're the field's best current answer to a hard question, which is how people stay in charge of systems that may one day be smarter than the people overseeing them.

          Framework names as each lab publishes them. The US institute was renamed the Center for AI Standards and Innovation in 2025, and the UK's became the AI Security Institute.

          Four ways serious people see it

          Informed people land in very different places, mostly because they disagree about how soon very powerful AI arrives and how dangerous it would be.

          Here's the strongest version of each view along with its weak spot. Most people hold some mix of these, and the positions overlap more than online arguments would suggest.

          Chapter 5 of 5

          What it's for

          What becomes precious when thinking gets cheap, and what all of this is ultimately for.

            When thinking is cheap, what becomes precious?

            Every technology that made something abundant made something else scarce, the way printing made copies cheap and attention valuable.

            Tap each item to sort it and watch where the weight settles.

            Old answers to a new question

            1930

            Keynes and the fifteen-hour week

            With the world sliding into depression, John Maynard Keynes predicted that within a century productivity would rise so far that his grandchildren might work around fifteen hours a week. He expected the real challenge to come afterward, when people had to learn how to live wisely and well once necessity no longer organized their days.

            He was right about the productivity and wrong about the hours, because people kept finding new things to want.

            1946

            Frankl and the sources of meaning

            Viktor Frankl, who survived Auschwitz, argued that meaning comes from three places: creating something or doing a deed, experiencing something or loving someone, and the stance we take toward suffering we can't avoid.

            A machine can help with every one of those, but it can't do them on your behalf, because the meaning lives in the person doing the creating, the loving and the enduring.

            Why build it at all?

            The reasons people give out loud are real, and so are the ones that also drive the race.

            Said out loud

            Curing diseaseFaster scienceCheap clean energyAn end to drudgeryUnderstanding minds

            Also driving it

            MoneyPowerNational securityFear of falling behindPrestigePlain curiosity

            Most of an iceberg sits below the waterline, and that's the part that decides where it goes.

            That leaves the deepest question, which is less about whether superintelligence can be built and more about whose values it would serve and who gets a say. That's a moral and political question rather than a technical one, so it belongs to everyone and not only to engineers.

            What to do with all this

            Nobody controls the curve, but everyone chooses how they meet it, and a few habits hold up across nearly every future in this guide.

            • Use the tools every day on real work, so you know firsthand where they're brilliant and where they break instead of relying on headlines in either direction.
            • Lean toward judgment, presence and accountability, since those are the parts of almost any job that are gaining value.
            • Invest in what stays yours, like skills in your hands, relationships, local trust and a reputation people can check.
            • Watch signals rather than hype, especially how long systems can work reliably on messy real tasks, who controls them and how they're checked.
            • Stay in the conversation, because what these systems are for is still being decided and the people who show up get a say.

            A threshold is a place you stand before you walk through, and how we cross this one is still ours to decide.