In January 2023, two months after ChatGPT launched, I started building annexr. I thought I was late. Three years later, I realize I was early.

Feeling late while being early is the normal state of building in AI right now. It isn’t foresight. Progress here compounds along several tracks at once, each on its own clock, and that makes your own timing almost impossible to read while you’re in it.

One curve versus many

Moore’s Law was one curve. Transistor density doubled on a predictable clock, and you could plan around it. Even that was less a law than a promise Intel kept working to honor. When the physics pushed back, they found another way to hit the number.

AI has no equivalent. No single company owns a single metric. What we have instead is a bundle of loosely linked curves: training compute, algorithmic efficiency, inference cost, hardware, and how long an AI system can run before a human has to step in.

Two different engines drive them at the same time.

  • Moore’s Law runs on the calendar. Time passes, chips get better.
  • Wright’s Law runs on volume. The more units you build, the cheaper each one gets.

In AI both run at once, which is why progress never seems to pause. When they come apart, you get a wall.

The curves, briefly

Four numbers worth carrying around:

  • Training compute has grown about 4.5x a year since 2010.
  • Algorithmic efficiency improves on its own track: the same capability costs roughly 3x less compute each year.
  • Inference cost has dropped by orders of magnitude, somewhere between 9x and 900x a year depending on the task.
  • Agent task horizon, the length of a job an AI can finish before it needs help, has been doubling every four months or so.

They feed each other. Cheaper inference means you can spend more compute at the moment of use, which buys better reasoning, which stretches the task horizon further, until something gives.

Where the curves break

Benchmark scores rise with the logarithm of compute. Each doubling in score costs far more than double the training budget. The industry’s current bet is to shift spending out of training and into reasoning at the moment of use. Whether that arithmetic survives at scale is still open.

High-quality human text is nearly used up at frontier training scale. Synthetic data and self-play have kept things moving in narrow areas like math and code. Whether that carries over to everything else, nobody knows yet.

Task horizon measures how long, not how reliably. An agent that runs for a hundred hours but fails about once an hour is no use for anything that matters. Reliability numbers trail capability numbers badly, which tells you what the field has been optimizing for.

Then there is electricity. Moore’s Law was about doing more in the same space. AI scaling is about raw size: bigger clusters, more gigawatts. Grid capacity moves on civil engineering timelines, not chip ones.

The slowest curve is the institutional one. Compliance, procurement, security review, change management. None of it scales exponentially. The capability curves are steep; the institutional curve barely tilts.

What this means if you build things

The line of what’s worth building keeps moving. Something too expensive or too flaky last quarter can make sense this quarter. Something that barely works today can be boring infrastructure in a year.

But what stands between a working demo and actual adoption has changed. It’s rarely capability now. It’s reliability, power, data quality, and the years it takes a big organization to trust a new system.

Most of the worthwhile work left this decade sits in that gap, between a capability curve that climbs fast and an institutional curve that barely moves. That’s slower and less interesting than pushing the frontier, and it’s where I’d spend my time.

If you want the backstory on how these three threads came together, I wrote about that separately: The AI moment was seventy years in the making.

Sources and further reading