Why I don’t think AGI is right around the corner

This episode is a narration of Patel's own June 2025 essay, and it opens by locating the disagreement precisely. His guests argue about timelines to AGI with numbers ranging from two years to twenty, and he wants to explain where he sits and why. He borrows Dornbusch's line that things take longer than you think and then happen faster than you thought, which sets up the structure of the whole argument: a near-term delay caused by a specific missing capability, followed by a very fast takeoff once that capability arrives.

His first move is to reject a common claim among accelerationists: that today's systems would be economically transformative even if progress stopped entirely. He disagrees, and his evidence is deliberately personal. Fortune 500 firms are not failing to adopt LLMs out of institutional inertia; they are failing because getting ordinary human-like labor out of an LLM is genuinely hard. His own testbed is roughly a hundred hours of building small LLM tools for his post-production workflow, and he says the experience lengthened his timelines rather than shortening them.

What makes the example pointed is that the tasks are exactly the kind the technology should own: short-horizon, self-contained, language-in-language-out work such as rewriting auto-generated transcripts for readability or pulling quotable clips out of a transcript. He rates current models at five out of ten on these. The failure is not that they are bad at language, but that they do not improve the way a person would. That is the hinge of the essay: the missing capability is continual learning, and the cost of its absence shows up as a ceiling on what any firm can delegate.

The core of the argument is an analogy about skill acquisition. Human value is not mainly raw intelligence but the ability to accumulate context, interrogate one's own failures, and pick up small efficiencies through practice. Patel then asks how one would teach a saxophone if the only available channel were written instructions. A student makes one attempt, is sent away, and receives a detailed note about what went wrong; the next student reads the note and tries to play Charlie Parker cold.

This just wouldn't work.