Today’s AI Is the Worst It Will Ever Be. That’s the Reason to Start Now.

It’s getting smarter in two directions at once — in the big data centers, and on the computers people already own.

In April 2024, on a stage at Stanford University, the head of OpenAI said something that will stay true for the rest of the decade: “GPT-4 is the dumbest model any of you will ever have to use again.” He wasn’t insulting his own product. He was pointing at the floor. Whatever AI you use today is the worst one you will ever use — and next year, that sentence will still be true.

That’s easy to nod along to and harder to act on. Most people treat AI like a new appliance that isn’t quite finished yet — wait for it to mature, wait for the price to settle, buy in once it’s stable. That instinct misreads what’s happening. AI isn’t a product with a finish date. It’s a moving curve, and the curve is climbing in two directions at the same time.

Direction one: the big systems keep getting bigger

The most powerful AI systems live in enormous data centers, and the amount of computing power used to build them has been growing four to five times every single year, according to the researchers who track this most carefully. The system behind the original ChatGPT era was built with roughly seventy times less computing power than the one that followed it just three years later.

Part of that is smarter research. But a bigger part than most people realize is simpler: the chips got roomier. An AI system has to logically fit inside a computer’s memory, the way furniture has to fit inside a room. Every time chipmakers build a roomier chip, engineers can build a bigger, more capable AI. The specialized chips in data centers have more than doubled their memory in the past three years. Bigger rooms, bigger furniture. That trend has no end in sight.

Direction two: real AI now fits in a home computer

Here’s the part almost nobody talks about. The same progress is pushing serious AI down — onto ordinary computers that people already own.

A few years ago, a good home computer could only run a small, fairly limited AI — a pocket-calculator cousin of the real thing. Today, the graphics chips sold in regular desktop computers have grown their memory enough that a genuinely capable AI fits on a single machine in a home office. No data center. No monthly bill. No internet connection required, which also means whatever you type into it never leaves the house.

And these home-sized systems are not toys. One of the leading freely available models, released this spring called Qwen3.6-27B, fits on one desktop computer and scores within two points of the original GPT-4 on a standard test of knowledge and reasoning — the same GPT-4 that, three years ago, needed a warehouse full of equipment and made front-page news. Just, once more: the breakthrough that stunned the world in 2023 now fits on a countertop.

The trend keeps rolling downhill. Microsoft has an AI that runs entirely on an iPhone, answering faster than most people read. Apple, Google, and Meta all now ship small AI systems that live right on devices in your pocket. Today’s best-in-the-world models will run on home computers in a few years, and yesterday’s best are already headed for phones.

Why this matters for the price

Right now, most businesses that use AI pay for it the way you pay for electricity or long-distance calls used to work: by the amount used. Every question, every document processed, adds to the meter. That meter quietly decides what’s affordable — plenty of good ideas never happen because running them through a pay-per-use AI would cost too much.

But that price is collapsing. Stanford University’s annual AI report found that getting a fixed level of AI performance cost about $20 per unit of work in late 2022 — was about $0.07 by late 2024. The same capability, nearly three hundred times cheaper, in two years.

Now put the two trends together. Capable AI fits on machines people already own, and a machine you already own is already paid for. Whether it sits idle overnight or works all night, the cost is the same. When using AI stops costing anything extra, whole categories of ideas that never made financial sense suddenly do — including plenty nobody has thought up yet.

Regular software wears out. AI improves on its own.

One thing that makes AI different from every technology investment that came before it, and it’s the reason waiting is the expensive choice, is how they change, or don’t.

Traditional software is like a house: the day it’s finished is the best day of its life, and then the maintenance begins. Operating systems change, security holes need patching, the programs it connects to get redesigned. The software sits still while the world moves, so you pay real money every year just to keep it working the way it did on day one. Anyone who’s owned a computer for a decade has felt this.

AI runs in the opposite direction. A tool built around AI today gets better as the AI underneath it improves — and swapping in a newer, smarter version is usually a small job, not a rebuild. It’s less like a house and more like an orchard: a little tending, and the asset improves on its own. It’s the rare purchase where doing nothing means it gets better, not worse.

Plant the tree

Any AI tool a business puts in place today is the worst version of itself it will ever be — and that’s true even if nobody lifts a finger to improve it. The systems get smarter. The machines get cheaper. The floor keeps rising under whatever you started.

You cannot harvest an asset you never planted. The companies reaping AI’s rewards today are the ones who already put the seed in the ground. To harvest your own rewards tomorrow, the time to plant the seed is today. The curve isn’t going to wait for anyone to feel ready, and every year spent waiting for AI to “mature” is a year spent standing on a floor that was already higher than you thought.