Machines

The name is exact. What it is trained on is language, so the signal is syntactic: which symbol follows which. Meaning is never handed to it. That this works as well as it does is the genuine surprise of the decade.

Context in, a probability for every next token out, one sampled, appended, and round again. The answer becomes the next question.

It cost megawatts for weeks; you run on twenty watts. Neither figure changes what kind of object it is: arithmetic on bits, a C* program in principle, which mipster would run.

A model with a few hundred billion parameters is a point in a state space of the kind week 2 measured. Its behaviour — its output on every possible context — is one of week 7's infinite answer sheets.

So when a laboratory says a model was tested extensively, you know what that means. Whatever the tests exercised, the machine did. Whatever they did not, nobody knows.

Testing shows presence, not absence. Dijkstra, 1969, about programs. It was always about this.

Fluency is syntax. Correctness is semantics. We built a machine of extraordinary fluency, so the gap between the two — week 7's gap — is something you now meet before breakfast.

Not a defect to be patched away: it is the proof/truth distinction at consumer scale. Harnad, 1990: symbols defined only by other symbols never touch the world.

So the durable response is not "trust it more" or "trust it less" but check it against something with a semantics — a compiler, a test, an experiment, a source, a colleague.

Not more text but a model of the world — state, dynamics, consequence. Predict what happens, not what is said: an attempt at real semantics, and a bid for efficiency. On a fixed budget efficiency is capability, so world models may well outperform language models outright.

And still a finite notation, sitting inside the world it models. Counting: models are countable, behaviours are not. Rice: "is this model right?" is semantic. Gödel: no self-certificate from inside. Cost: the state space did not shrink. Only the map did.

Models are trained on text that models wrote — with measurable degradation when the loop tightens (model collapse, 2024). Models grade models. Models write the code that trains models, and agents invoke themselves.

"Is this system safe?" is a semantic property of a program. Rice says: no general decision procedure. Not hard — impossible. And by Gödel's second theorem, no system this expressive can certify itself. An explanation of its own output is more output.

Which is not despair. It says exactly where to spend effort: on external, independent, bounded checks.

Prompt in English, and what comes back is plausible notation. Unchecked, it is simply accepted — and then there is no semantics anywhere in the loop. Checked, it meets a compiler, a test, a proof, a measurement, and is kept or sent back round.

The fourth appearance of one theorem. Cantor's list needed an object built from outside it. Gödel's system needed a stronger one. Thompson's compiler needed a second compiler. The generator needs a check it cannot be.