How to Plan for AI That Is Guaranteed to Get Better (You Just Don't Know How)

In 2020, researchers at OpenAI published a paper called “Scaling Laws for Neural Language Models.” Several of its authors, including Jared Kaplan, Sam McCandlish, Tom Brown, and Dario Amodei, went on to co-found Anthropic, the company behind Claude. The paper made a discovery that explains most of what has happened in AI since: performance improves on a smooth, predictable curve as you add computing power, data, and model size.

Think of it like fuel economy testing. Once you know the curve, you can predict how far a tank will take you without driving the route. The researchers found that these curves held across enormous ranges, which is like a rule about walking speed that still works at jet speed. This is why AI investment exploded. The labs stopped gambling and started budgeting. They could forecast how capable the next model would be before spending a dollar to build it.

Here’s the part that matters for your planning. The curves predict one thing with precision: the model’s overall error rate. What they predict only roughly is which specific skills show up, and when. A model can slide smoothly down the error curve and then suddenly cross a threshold where it can do something new, like write working code or follow a complicated multi-step instruction. Even the labs get surprised by the details.

So you’re planning around a strange combination: guaranteed improvement, uncertain specifics. That sounds exotic, but operators have handled this before. Nobody running a business in 1995 understood transistor physics, but smart ones planned around the reliable fact that computing got twice as cheap every couple of years. That planning input built entire industries. AI’s scaling curves are the same kind of input. Here’s the playbook.

1. Budget like it’s Moore’s Law

Assume the cost of a given level of AI capability keeps falling. What the most expensive model does today, a cheap model will do in a year or two. This changes how you evaluate use cases. If automating a workflow pencils out at ten times the current price, the idea isn’t dead. It’s early. Put it on a shelf with a revisit date, and re-run the numbers annually. Any pricing assumption about AI that’s more than a year old is stale.

2. Give failed pilots expiration dates

If AI botched a task for you last year, that result describes last year’s error rate and nothing more. The threshold for “good enough” gets crossed quietly, without a press release for your specific use case. So keep a simple list: every AI experiment that didn’t work, what it failed at, and a re-test date about six months out. Most businesses run a pilot once, write off the whole category, and never look again. A re-test calendar turns that write-off into a pipeline. Some of the best AI wins come from ideas that failed eighteen months earlier.

3. Build light, and plan to simplify

When a tool has a weakness, the instinct is to engineer around it: elaborate prompt templates, multi-step review chains, custom middleware. Some of that is worth it. A lot of it becomes obsolete before it pays back, because the weakness it compensates for disappears in the next model release. My rule of thumb: build the lightweight version of any workaround, note which limitation it exists to solve, and check in each quarter on whether that limitation still exists. Treat scaffolding as temporary by default. The goal is a workflow that gets simpler over time, and a workflow that only ever accumulates duct tape is a sign you’re solving temporary problems with permanent solutions.

4. Hold the one constant: verification

Here’s what the curves do not predict, at any point, at any scale: trustworthiness. The error rate falls, but these models are trained in ways that reward confident, articulate answers, and that stays true as they improve. A more capable model is often a more convincing model, including when it’s wrong. So while everything else in your AI plan should be provisional, verification is permanent. Check the numbers, check the claims, keep a human at the checkpoints that matter. Every other line in this playbook has a revisit date. This one doesn’t.

The bottom line

The research gives you one certainty and one uncertainty, and both are useful. The certainty: capability keeps improving on a schedule set by compute budgets, so patience is a strategy and stale conclusions are a liability. The uncertainty: specific skills arrive unannounced, so regular re-testing beats prediction. Plan around the trend, verify the output, and let the labs worry about the physics.