LLMs Are Amazing. But They're Not the Answer — They're an Answer.

A gloriously simple idea turned out to be wildly useful. I still think it's a dead end on the road to AGI.

LLMs Are Amazing. But They're Not the Answer — They're an Answer.

A gloriously simple idea turned out to be wildly useful. I still think it's a dead end on the road to AGI.

I've been thinking about buying an electric car, and everyone keeps warning me about range anxiety. They needn't bother. I've been managing it perfectly well on my own. I imagine the percentage ticking down on the motorway and feel a low hum of dread I never get with a tank of petrol. A petrol car goes further on a single fill. The numbers say so. By that measure, the EV is the worse car.

Then I started doing the sums, and the worry quietly fell apart.

The reason is simple, and it took me longer than it should have to see it: you don't really fill up an electric car. You plug it in at home, and most mornings you walk out to a "full tank" without having done a thing. The petrol car  (the one with all that glorious range) demands a small chore I've completely normalised: noticing the gauge, detouring to a station, standing in the cold holding a nozzle, queuing to pay. The EV doesn't extend that experience. It deletes it.

That's the thing about range anxiety. It's a real limitation, but it's the wrong thing to measure. I'd been judging a new kind of car by the old one's best trick, while completely missing what the new car does better.  Refuelling, more or less, stops being something you do.

And yet the internet would have me believe range is the only question that matters. My feed is wall to wall with it: "How far on a single charge?", "EV road-trip hell", earnest blokes filming a battery gauge on a motorway at 2am. They're oddly gripping,  and they test the one situation I'll almost never be in. The honest everyday question isn't "how far on a charge", it's "how often will I actually have to go out of my way to fill up?" For a car that charges at home overnight, the answer is: hardly ever. And on nearly every measure that touches the daily drive, the EV simply wins. No noxious emissions, instant acceleration, more than enough usable speed, a quieter, smoother ride. The videos obsess over the one metric where petrol still leads and wave away the dozen where it doesn't.

I've started to think most complaints about AI are range anxiety.

"It gave me the wrong answer." Yes. It does that. Large language models confidently state things that aren't true, fumble arithmetic a calculator would nail, and occasionally invent a source out of thin air. Measured against a perfect oracle, or a careful human expert, or even a decent search engine, they fall short, and the shortfall is real.

But that's the petrol-station yardstick. It measures the new thing by the old thing's strength and misses what's actually new.

Because here's what I can't get over: the core idea is almost insultingly simple. Predict the next word. Do it over a staggering amount of text, at a staggering scale, and… that's basically it. There was no reason to expect that a glorified autocomplete would be able to draft a wedding speech, debug my code, explain a tax form, translate idiom, summarise a contract, and hold a conversation about all of it in between. That something this conceptually plain turned out to be this broadly useful is, to me, one of the most surprising results in the history of computing.

So when someone writes off the whole enterprise because it muddled a date, I hear range anxiety. You're standing at the motorway services, fretting about the one number that used to matter, while ignoring the fact that you woke up this morning with a thinking tool already plugged in and ready to go.

Which is where I'd happily end, except I believe something that sits awkwardly next to all that praise.

I don't think large language models are the road to artificial general intelligence. I think they're a dead end.

Here's the part of the EV story that makes me nervous about all this. Everything I said earlier about the EV winning is true… for a car. A battery is a wonderful way to move one: do short, repeatable journeys with somewhere to plug in overnight and it's close to ideal. The trouble starts the moment you ask the battery to leave that comfortable domain. Ask it to do the heavy lifting — literally — and the technology hits a wall that no amount of refinement gets over. The culprit is energy density. A kilogram of diesel or jet fuel holds something like an order of magnitude more usable energy than a kilogram of the best batteries, even after you account for how much more efficient electric motors are. For a commuter car that barely matters. For a 40-tonne lorry running freight across a continent, the batteries start eating into the payload. For a mining truck or an excavator working back-to-back shifts, there's no overnight in which to charge.

And for an aircraft the maths is brutal: a plane has to haul its energy into the sky, and a battery, unlike a fuel tank, doesn't get lighter as it drains. The numbers simply don't close for anything beyond short hops. Nobody serious is waiting for a bigger battery to crack long-haul aviation; they're looking at different technologies entirely. Hydrogen, synthetic fuels, fuel cells because in that domain the battery isn't a worse version of the answer. It's the wrong shape of answer.

I suspect LLMs are the same. More parameters, more data, longer context, cleverer prompting,  these give us better LLMs, and better LLMs are genuinely worth having. But "better text predictor" and "general intelligence" may be different categories, not different points on one line. The very mechanism that makes these models work, might be the same mechanism that stops them short of the things we actually mean by understanding. A stable model of the world, reasoning that doesn't quietly fall apart off the beaten path, knowing what they don't know. You can't necessarily scale your way across a gap that's a difference in kind.

I hold this loosely, and I should say the honest counterargument out loud: scaling these models has humiliated sceptics before. Capabilities have appeared that nobody confidently predicted, and "it'll never do X" has aged badly more than once. Maybe the bridge to AGI gets built from LLMs plus other parts like memory, tools, agents, senses.  The prediction engine ends up as one component of something larger rather than a dead end at all. I might be wrong. I'd find it fascinating to be wrong.

But "dead end" doesn't have to be an insult. Some dead ends are worth every step. This one gave us tools I use every day, built on an idea simple enough to explain over dinner. If the road to AGI turns out to run somewhere else entirely, a different technology, a different shape, that won't retroactively make these years a mistake, any more than an electric car is a mistake because it isn't, and was never going to be, a petrol car.

It's just a different car. A really good one. I just wouldn't bet on it growing up to be something it was never built to be.