What will automated firms look like?

Dwarkesh Patel opens by relocating the debate about advanced AI from model capability to organizational form. His target is a specific failure of imagination: people picture AGI as a very capable assistant bolted onto existing institutions, and therefore ask only how smart the assistant will be. He argues the more important question is what happens when every worker and every manager in a company is an AI, because the consequences of that are not a scaled-up version of today's firms.

The mechanism he points to is not intelligence but digitization. The advantages he has in mind are collective and structural: AIs can be copied, distilled, merged, scaled and evolved in ways people cannot. This framing matters for how the rest of the essay should be read. If the argument were "AIs will be smarter than people," it would stand or fall with benchmark progress; because the argument is about reproduction and recombination of competence, it stands or falls with the engineering of model duplication, weight surgery and organizational design.

He is explicit about what the essay is not. It is not a forecast of a particular model release, and it is not about emulations of specific humans. The subject is AGI as the descendant of LLMs that can do basically anything a human can do, and the imagined object is the firm that such systems would staff. That scoping move lets him use familiar names (Sundar Pichai, Jeff Dean, Elon Musk) as stand-ins for capability profiles rather than as claims about actual digital twins.

The first of the five mechanisms is copying, and Patel's entry point is the hiring bottleneck. Firms today are constrained by how many excellent people they can find and train. If the talent is an AI, the constraint becomes the number of copies you choose to run. His illustration is a Google with a million AI software engineers who are not generic workers but the functional equivalents of its best engineers, carrying their skills, judgment and tacit knowledge.