The Sigmoid Comfort: Why 'AI Will Plateau' Isn't the Reassurance You Think

There’s a phrase circulating in AI discourse that functions as a kind of emotional sedative: “Every technology follows a sigmoid. It looks exponential now, but it’ll flatten soon.” Scott Alexander of Astral Codex Ten recently took this comfort apart, calmly, and the argument lands harder than you’d expect.

The Reassurance Built Into an S-Curve

A sigmoid is the familiar S-shape: slow start, explosive middle, gradual flattening. Moore’s Law followed one. So did aircraft altitude, car top speeds, and most major technologies once you zoom out far enough.

The implication people draw is obvious. GPT-3 to GPT-4 was a leap, sure, but data runs out. Compute hits walls. Algorithmic gains saturate. The message, delivered with a shrug, is “don’t worry, it’ll plateau.”

You can hear the psychological need underneath it. AGI anxiety, job displacement fears, alignment dread — sigmoid talk whispers that nature itself will hit the brakes before things get weird. It’s less an argument than a coping mechanism dressed in a graph.

Alexander’s Counter: The Plateau Height Is What Matters

The problem, Alexander points out, is that nobody knows where the curve flattens. That’s not a small footnote. It’s the entire question.

Aircraft speed is the cleanest illustration. Yes, it followed a sigmoid, and yes, it eventually plateaued near the sound barrier. But imagine if the plateau had landed at 10 km/h instead. Civilization would look unrecognizable. Now imagine it had flattened near the speed of light. Also unrecognizable, in the opposite direction.

AI is the same shape problem. The sigmoid might top out at “competent grad student” — annoying but manageable. Or it might top out at “every genius in history combined and running in parallel.” Knowing the curve is S-shaped tells you nothing about which world you live in. It’s like buying airline stock because you’ve confirmed planes follow a sigmoid, without checking whether they max out at jogging speed or escape velocity.

Where Are We on the Curve, Exactly?

Alexander’s second jab is sharper. Even granting the sigmoid framing, we have no reliable way to locate ourselves on it. Are we at the foot of the steep climb? Mid-ascent? Already past the inflection?

Benchmarks keep getting saturated, which suggests momentum, but benchmark saturation isn’t a position on a curve — it’s a position on a benchmark. Some researchers see clear walls approaching. Others say the field is barely warmed up. Both camps point at the same data and read it differently.

This is where the comfort collapses. “It will eventually flatten” can be true and useless at the same time. The sentence that matters — when, and at what capability level — gets no answer from the shape alone. Sigmoids promise a silhouette, not a finish line.

Sigmoids Stacked on Sigmoids

There’s a more uncomfortable framing worth sitting with. Technological progress isn’t usually one sigmoid. It’s a chain of overlapping ones, each new paradigm starting its climb roughly when the previous one stalls.

Steam plateaued, internal combustion took over. Vacuum tubes flattened, transistors began their own curve. There’s no obvious reason AI gets a single sigmoid either. Transformers may saturate, and something else — different architecture, different training paradigm, different substrate — picks up where they leave off. In that world, “it’ll plateau” isn’t even wrong; it’s just irrelevant. The plateau of one curve is the launchpad of the next.

This is why Alexander’s piece bites. To draw any genuine comfort from the sigmoid story, you’d need to answer three things: what capability level it tops out at, how long it stays flat, and whether the next paradigm starts before the old one cools. Nobody has those answers. Nobody is close.

What You’re Actually Left With

The sigmoid argument turns out to be a poor shield in AI safety debates. It’s an observation about shape, not a prediction about altitude. The only question that matters — where does this curve stop, if it stops — sits exactly where it did before anyone invoked the S.

So it’s worth asking yourself, honestly: when you reach for the sigmoid argument, are you doing analysis, or are you doing therapy? They look similar from the outside. They feel very different from the inside.