The reasoning you can see is a setting, not a window

When a model prints "let me think step by step" it is easy to feel you are watching it think. Gary Marcus, on stage with Brian Greene, says the quiet part: the visible trace is a disclosure setting chosen by the lab, not a window into the computation, and lately labs show less of it so competitors cannot copy the steps.

• Apple's "Illusion of Thinking" is the clean demo: a model that solves Tower of Hanoi at seven disks and collapses at eight, where a child applies the rule.
• The progress on hard problems comes from symbolic scaffolding around the model (provers, interpreters, loops), which the industry rarely names because "scaling" is the fundraising story.

• Marcus has been predicting the wall for years; the wall keeps moving. Take the epistemics, discount the forecast.
• A benchmark score is a surface the optimiser can satisfy without doing what you meant, like the classifier that told wolves from huskies by the snow.

Source: Gary Marcus με τον Brian Greene, World Science Festival (YouTube)