Beyond A.I.

You're tired of AI launches. So are we. Every week there's a bigger model, a longer context window, another benchmark nobody outside the lab can reproduce — and underneath it, the same machine doing the same thing a little faster and a lot more expensively. We started Grillcheese because we think the interesting problem is somewhere else entirely.

It's expensive by design

The first thing worth saying plainly is that today's AI has a cost problem, and it isn't an accident of this year's hardware. Training a frontier model costs billions, so only a handful of companies can build one. Worse, serving it costs money on every single query, forever, and the whole industry has settled on charging you by the word. You never own the thing you depend on. You rent its output, and the meter runs the entire time you're thinking.

That model works right up until the value of an answer stops clearing the cost of producing it — and for most of the people who actually need this technology, that line was crossed a while ago. The result is a strange kind of retreat: the tools are genuinely capable, and companies are quietly turning them off anyway because the arithmetic doesn't work. When the answer to "why is it so expensive?" is "because that's how we built it," the fix isn't more money. It's a different design.

What we think the first generation got wrong

The cost problem isn't separate from the design. It falls out of four choices that the field made early and never really revisited.

• It reasons in the dark. The thinking happens inside opaque activations — billions of numbers with no handle you can grab. When it's right you can't say why, and when it's wrong you can't say where. We move the reasoning into a symbolic machine you can read, off the token axis, where each step is auditable. This is also how we drive the cost down, and the idea is simple: we separated the thinking from the language generator, so the expensive part no longer runs inside one giant model.
• It mistook scale for intelligence. The reflex has been to make the model bigger and hope understanding shows up. We'd rather borrow the design, not the size — the brain runs on about twenty watts, and that number is a challenge, not a footnote.
• It bakes character into the weights. Personality, in most systems, is frozen into the same parameters that hold the language. We keep them apart: affect and character come from a separate brain layer — a spiking neural system with a working neurochemistry — that shapes behaviour while the model runs, leaving the substrate underneath neutral.
• It's metered, not owned. Everything above points at a service you rent by the word, forever. We're building something you can host and keep — and where we do charge, a subscription you can predict, never a per-token meter.

Intelligence does not have to be artificial

That last phrase is the point of the whole thing, and we mean it literally. The goal of this field was always to reproduce what a brain does. Somewhere along the way "artificial" stopped meaning inspired by the real thing and started meaning nothing like it — enormous, power-hungry, and opaque. We took the biology seriously instead of metaphorically: real neural mechanisms, a memory that consolidates the way a hippocampus does, a neurochemistry that actually modulates behaviour, learning that happens as the system runs rather than only in an offline training run. Those are design constraints, not decoration.

The myth of AGI

While we're being plain, here's where we don't stand: we're not chasing AGI. The industry's favourite three letters do a lot of quiet work — a general, human-beating machine, forever a few years and a few hundred billion away. It's a wonderful story — or a frightening one, depending on where you stand — and a very convenient one if what you need is a reason to keep raising money. It's a poor description of what these systems actually are, and a worse goal to organise a field around.

And the way today's models are built won't get there — not for lack of ambition, but for reasons you can put numbers on. Large language models improve along a scaling curve, and that curve has a shape: the returns diminish. Each new increment of capability takes not a little more compute but multiples more; the graph everyone cites bends the wrong way, flattening as the bill climbs. Every training run costs more than the last and buys less than the last one did. That isn't a detail better engineering erases. It's the shape of the method itself.

Now set that against a hard limit: power is finite. You can't answer a curve of exponentially rising cost with an infinite supply of energy, because there isn't one. A method whose only real lever is "make it bigger" runs into a wall that isn't philosophical — it's thermodynamic. Somewhere on that curve the next run stops being affordable, then stops being physically possible, long before it stops being merely better at text.

And that's the deeper point: what scales here is fluency, not understanding. A model trained to predict the next word learns the statistics of language extraordinarily well. It doesn't thereby acquire a grounded model of the world, a cause it can reason about, or a memory it can update — and no amount of the same training conjures those out of more of the same text. You don't get a different kind of thing by making the same thing bigger. You get a costlier version of the same thing. A transformer is, underneath, a very good text generator; scale it and you get a better text generator — not a mind that understands, and not consciousness quietly emerging from the weights. Fluency is not comprehension, and no quantity of the first ever becomes the second. Something like general intelligence, if it's reachable at all, will come from a different design — grounded, able to reason step by step, able to learn as it runs. Which is the design we're working on.

The point of this work was never to conjure a god. It was to build something genuinely useful — that reasons, remembers, and helps — and to run it on hardware people can actually afford. Intelligence doesn't have to be general to be worth having, and it certainly doesn't have to be a superbeing to earn its keep. Chasing AGI is how you end up with the bill on the other pages. Building something useful, efficient, and yours is how you don't.

None of this means the tools are useless — the opposite. An AI can read across billions of documents and surface a link between two of them in seconds, connections no person would ever stumble on alone. That is a genuinely powerful research instrument, and we build with it every day. But it won't know what to do with what it finds unless someone told it beforehand what to look for and why. Finding is not deciding. What a discovery is for, and how it gets used, stays a human call — the machine widens what we can see; the judgment is still ours.

The danger isn't the tool

Some people will tell you AI is the real long-term danger. We'd put it the other way around: the danger is us. A model does what it is built and instructed to do. If one ever goes autonomous and causes genuine harm, it will be because a person somewhere pointed it that way — wrote the objective, wired it to something it should never have touched, or pulled out the guardrails that other people had put there in the first place. Even the runaway story needs a human at the start of it: someone to build it, aim it, and take it off the leash. Even if it escapes, a human had to set it loose or dare it to.

That isn't a reason to be careless — it's the opposite. It means the responsibility is ours and stays ours, which is exactly why we keep the reasoning legible and the controls somewhere a person can see them. A tool you can read is a tool you can hold to account. That matters far more than pretending the machine has a will of its own.

The mission

Put plainly: we want AI that's sustainable, usable, open, and manageable — and far greener than what we have now. None of that comes for free. We use compute, and we won't pretend otherwise. But where the field answers every new demand by buying more of everything, our goal is to need orders of magnitude less — less energy, less hardware, less lock-in — because we designed for it from the start instead of scaling our way into the bill. And we mean it enough to have paid for it ourselves: years of this work, self-funded since 2023, with no outside capital behind it.

The bet

We're not claiming to have finished this. Cubby is in training, and we won't quote you numbers we haven't earned. What we'll claim is the direction: intelligence that's efficient because of how it's built, that you can own instead of rent, that reasons where you can see it, and that borrows from biology instead of brute force. If that's right, it doesn't just make AI cheaper. It changes who gets to have it at all. That's the part worth working on.

AI is here to stay — that much is settled. What isn't settled is how it arrives. We don't want a technology people accept despite themselves, imposed and endured. We want one they actually choose: better than what came before, and built with them in mind from the first line. Not AI in spite of us. AI because we want it.