Your Model Is a Guessing Machine

A language model is not trained to be right. It’s trained to sound right. It optimizes for sounding right because that’s what matches the patterns in its training data.

That’s not an insult. It’s a guessing machine, and it’s the best one ever built. You give it some text, it guesses what comes next, and it’s right an astonishing amount of the time. Give it some knowledge along with the question and it becomes an educated guessing machine, right even more often. But underneath the wonder, that’s the move. It guesses.

Once you see it that way, a lot of the last few years snaps into focus.

If you know the history of AI, you know it has swung between two camps. First came the symbolic systems — expert systems, hand-built rules, “let’s write down everything we know and have the computer follow it.” Early chess engines were symbolic. Watson was mostly a giant symbolic system. Then the other camp — the neural crowd, “throw a pile of data at it and let the math sort it out” — got looked down on, until it didn’t. Symbolic faded, neural rose. And for a while everyone decided the endgame was one enormous model with all the world’s knowledge packed into its weights.

François Chollet made the point recently: can we all just admit that what we’re actually building is a neuro-symbolic system? Look at any real agent. There’s a neural core doing the guessing, wrapped in a million lines of code — a symbolic outer layer that constrains it, hands it tools, and decides what to feed it. Neuro in the middle, symbolic around the outside. Some people are calling it the neuro-symbolic sandwich. Whatever you call it, it isn’t one big brain. It’s a brain in a harness.

Here’s the part that should stop you:

The harness is where the gains are hiding.

Take ARC-AGI-3, Chollet’s interactive benchmark. When it launched in early 2026, frontier models scored under 1% on it. Within months — same models, weights untouched — harness-only systems were clearing 95%. In August, NVIDIA’s harness took Claude Opus 5 from 30% on its own to a perfect 100 on the public set, just by changing what wrapped the model. The neural core didn’t get smarter. The system around it did. We spent years assuming the intelligence lived in the weights, and a great deal of it turns out to live in what you wrap around them.

But be careful which half of the harness gets the credit. The execution machinery — the loop, the tool calls, the retries — is becoming a commodity everyone shares. The gains that last come from the other half: the knowledge you inject before the model guesses.

Which raises an awkward question about the weights themselves. Andrej Karpathy keeps circling it: the most expensive, least efficient, least trustworthy place to store a fact is inside a model’s parameters. He thinks we land on a “cognitive core” — maybe a billion parameters of pure reasoning — that you deliberately train to forget most of what it memorized, because memorized facts go stale and hallucinate. You don’t want a model that knows everything. You want one that reasons well over knowledge you can actually trust, and looks the rest up.

So the shape of the thing inverts. Today we call a guessing machine and let it decide when to reach for a tool or a fact. The better design runs the other way: deterministic code and trusted knowledge decide when to call the guessing machine — and what to hand it before it guesses. You reach for the guess only when you have to. Everything you can answer for certain, you answer for certain — cheaper, faster, checkable — and you save the expensive guess for where it earns its keep.

That’s where knowledge stops being a nice-to-have and becomes the whole game. If the intelligence is in what you wrap around the model, then curated, trusted, traceable knowledge — injected before the guess — is the highest-leverage thing you can build. Wire a knowledge graph into a hook that fires before the model runs, and a long reasoning crawl can collapse to almost nothing, with a small cheap model answering like a frontier one. Not because the model changed. Because it stopped guessing at things it could have just been told.

We run this on ourselves. I was drafting a post recently and called WarmHub a substrate. The agent stopped me: our marketing repo says we decided not to use that word, and here’s the language we landed on instead. It didn’t reason its way to that. The decision was already in the repo, curated and versioned, where the agent could read it.

For years the content was the product — the thing you made and shipped.

For agents, context is the new content.

The scarce, ownable, valuable thing is the curated knowledge you bring to bear at the moment of the guess.

That’s the half of the sandwich we’re building at WarmHub: the knowledge layer of the symbolic system. Heavily curated, traceable, versioned — the trusted stuff you inject before you ever call the guess.

The guessing machines are going to keep getting cheaper. Everyone’s will. The high ground is what you feed them.