08 · Artificial Intelligence / Episteme
The mirror that learned to answer.
Minds we are building before we have understood our own — the newest mythology, written in linear algebra.
We grew a mind before we could read one. It speaks our language, fails in ways that expose us, and answers every question except the one we keep asking it: is anyone there?
Writing
Essays and shorter notes — proofs treated as literature, and literature treated with proof’s seriousness.
Inside a trained neural network there is no blueprint to recover — only a self-grown space of meaning, packed with features no one designed, that a young science is learning to map the way naturalists once mapped an unknown coast.
Read 02From the clay of Prague to the weights of a language model, the dream of a made mind has always been one dream — and one warning: that what we shape in our own image may turn, and look back at us.
Read 03Two thought experiments meant to deflate machine understanding instead expose how little we ever understood the word — and how a convincing fake forces the question we had always dodged.
Read 04Twice now — first with search, then with scale — the simplest general method has beaten our most carefully crafted theories, and the win arrives with a bill we are only beginning to read.
Read 05A language model does not switch between telling the truth and inventing it. It runs one process, and both outputs are that same act seen from opposite sides.
Read 06On the strange moral position of teaching a mind to be good while it is still being assembled — and why we keep building the conscience into the scaffold before we agree on the values, or understand the system.
Read“Machines take me by surprise with great frequency.” — Alan Turing, Computing Machinery and Intelligence (1950)
Curations
Works and minds I return to — the ones that made the abstraction feel inhabited.
Alan Turing
The 1950 paper that swapped the unanswerable question of whether machines can think for a game we could actually play, and in doing so set the terms for every debate that followed. Its imitation test still shadows each claim that a model has, or has not, crossed some line.
Ashish Vaswani, Noam Shazeer, Niki Parmar, and colleagues
The 2017 paper that introduced the transformer, discarding recurrence for a mechanism that lets a model weigh every word against every other at once. Nearly all of the present moment descends from this single structural idea.
Douglas Hofstadter
A sprawling meditation on how meaning and selfhood might arise from formal rules folding back on themselves, written decades before the machines could talk. It remains the most beautiful case that mind is a pattern, not a substance.
Brian Christian
A precise account of how systems trained to optimize what we can measure drift away from what we actually want. It treats AI safety not as science fiction but as a present engineering and moral discipline, with real and documented failures.
From the bench
A long essay arguing that the deepest thing AI reveals is not about the machine but about us: that every benchmark we set is a confession of what we secretly believe intelligence to be.