The explainer
Models can look thoughtful while failing brittle tests, and still persuade people they "get it." This explainer separates benchmark fluency, chain-of-thought theatre, and the harder claim that something is understood from the inside.
Fluency is a performance
A long, tidy answer can be a pattern that survived the prompt. Change names, invert the structure, or move the same logic into an unfamiliar wrapper, and some systems fall off a cliff. That is useful evidence about robustness. It is not, by itself, a window into phenomenal understanding.
Functional understanding is about what a system can do: transfer a concept, explain a result, catch an error, act effectively in a new case. Phenomenal understanding is whether there is a felt grasp. A model can pass some functional tests and leave the phenomenal question untouched. People also rationalize, confabulate, and fail under unfamiliar representations. The job is to design discriminating tests, not to protect a flattering story about either kind of mind.
Chain-of-thought is not a confession
Intermediate tokens that look like reasoning can be useful scaffolding, a style the training distribution rewards, or a way to spend test-time compute. Extra “thinking” can even make some tasks worse. Treat the trace as behaviour to evaluate, not as a diary of inner life.
Session history
Deep Dive #4 (July 24, 2025, SFU Vancouver) used Apple’s GSM-Symbolic work and inverse-scaling notes as receipts, not as a verdict that machines are empty. The room treated strong performance as important evidence while resisting the jump from successful output to a claim about experience. This public page removes the historical participant roster and chat excerpts.
Nearby questions in the series
If the word “understand” itself is crowded, continue at what it means for AI to understand. If competence can exist with the lights off, see whether intelligence can exist without consciousness. If fluent personas start to feel like they have a life of their own, the later dossier is whether AI personas are information parasites.
Further reading
The curated route is the Can Machines Understand? shelf on the MAC Library. Sibling surfaces: the Community Source Index for receipts, and mac.bc-ai.ca for the group home. Do not treat any one of those as the single reading-resource winner.