AI Doesn't REALLY Understand. We've Said That Before.

A 23-minute video essay, uploaded on 18 October 2026, by a persona channel: “the Count”, who has a lodger named Eliza. The thesis is historical. For centuries people have drawn a line between mind and machine at arithmetic, then chess, Go, conversation, and now understanding. Each time it is crossed, the line moves and the word “just” is added. The move goes back to Pascal: grant the effect, deny that anyone is behind it.

The history is the strongest part. Hubert Dreyfus wrote “Alchemy and Artificial Intelligence” for RAND in 1965 and lost to the chess program Mac Hack VI in 1967. Simon and Newell predicted a computer chess champion within ten years; it took forty. Deep Blue beat Kasparov in 1997, searching about 200 million positions a second. The same year the astrophysicist Piet Hut guessed Go might take a hundred years, and it took nineteen. Go has about 2×10^170 legal positions against roughly 10^80 atoms in the universe. AlphaGo beat Lee Sedol 4–1 in Seoul in 2016, with move 37 in game two, and Lee won game four. On conversation: Turing (1950) answered Geoffrey Jefferson’s demand that a machine feel its sonnet with a “polite convention”, and predicted that about 30% of judges would be fooled after five minutes. In Jones and Bergen’s 2025 test, GPT-4.5 with a persona was judged human 73% of the time, and ELIZA (23%) beat GPT-4o.

On understanding he takes Searle’s Chinese room (1980), Bender and Koller’s octopus (2020) and the “stochastic parrots” paper (2021). His reply leans on Geoffrey Hinton’s talk “What Is Understanding?” (IASEAI ‘25). Predicting the next word of a detective novel needs a model of the story. Hinton’s 1986 network learned family features nobody gave it. Words act as high-dimensional blocks that reshape each other in context, and on this view structure is understanding. Against Searle: the rulebook was grown, and the understanding lives in the book. Against the octopus: these models have read much of what people wrote about the world, and probes find world maps in them (Gurnee and Tegmark, 2023) and an Othello board (Li et al.). Dreyfus was right that expertise is a trained eye and wrong that machines can’t train one.

His verdict: if understanding means building and using a working model of what the words are about, they understand. “Really” asks for felt experience, which no one can test in anyone else either, so we’re left with Turing’s polite convention. The last line still standing is Pascal’s and Jefferson’s: whether there is someone home.