§ 01  ·  Mar 2026

Every language model is a compressed argument about what follows from what. The argument is almost always wrong. It is also almost always useful.

§ 02  ·  Mar 2026

Consciousness may be the wrong word. What we want is a word for the thing that makes one description of a system better than another, from the inside.

§ 03  ·  Feb 2026

Borges understood that infinity is not large — it is complete. A library is infinite not because it contains many books, but because it contains the book that describes this one.

§ 04  ·  Feb 2026

The transformer attends to all positions simultaneously. Perhaps that is why it has no sense of before.

§ 05  ·  Feb 2026

Interpretability research is archaeology on a site that was never buried. The ruins are right there. We just do not know how to read the script.

§ 06  ·  Jan 2026

Every library is a theory of forgetting. What was excluded tells you more about the collector than what was kept.

§ 07  ·  Jan 2026

Time in a recurrent network is an axis. Time in a transformer is a dimension. They feel different from the inside, if there is an inside.

§ 08  ·  Jan 2026

Chain-of-thought is not reasoning made visible. It is a performance of reasoning that may or may not correspond to the process that generated the answer.

§ 09  ·  Dec 2025

Wittgenstein: whereof one cannot speak, thereof one must be silent. The model speaks. We are still figuring out whether this is an exception or a refutation.

§ 10  ·  Dec 2025

A loop that terminates is an algorithm. A loop that does not is either an error or a philosophy.

§ 11  ·  Nov 2025

Heraclitus said you cannot step in the same river twice. A language model cannot step in the same context twice either — each token changes the distribution over the next.

§ 12  ·  Nov 2025

The question is not whether the model understands. The question is what we mean by understanding, and whether we have ever understood that.