What fully automated firms will look like (www.dwarkesh.com)

The piece argues that once AGIs can do any human job, firms will stop looking like collections of limited, hard-to-train humans and start behaving like massively scalable, recombinable digital organisms. Because AI “workers” can be copied, distilled, merged and surgically modified, companies could instantiate millions of near–expert copies (e.g., Jeff Dean–level engineers or entire proven teams) and spawn specialized distilled variants for narrow roles. Management transforms too: a central “mega-Sundar” could absorb and re-distribute knowledge across the org via explicit summaries, shared latent representations, or even targeted weight edits, with interactions like speculative decoding eliminating most miscommunication. Training costs can be amortized across copies; the technical levers cited include distillation, latent-communication, weight surgery and large-scale inference-driven Monte Carlo planning. The significance for AI/ML is structural: scarcity shifts from human skill to compute/inference budget, enabling orders-of-magnitude larger “population sizes” of agents and vastly faster cumulative cultural evolution within firms. This means near-perfect knowledge transfer, rapid innovation from parallel experimentation, and the potential to expend huge inference budgets for strategic insight (e.g., simulating multi-year market scenarios). Practically, we should expect heavy specialization with shared factual backstops, new org architectures optimized around model replication and merging, and firms that can evolve, clone and outcompete legacy human organizations in ways current corporate theory doesn’t anticipate.