The Hard Part Was Never the Instructions

Shyamal Anadkat published a long thread arguing that the economy is made of three things — atoms, energy, and instructions — and that AGI is collapsing the cost of the instruction layer. The framing is good. Firms exist because coordination is expensive. AI makes coordination cheap. Therefore the shape of firms changes. Classic Coase, applied cleanly.

Where it goes wrong is where these threads always go wrong: by treating everything that isn't atoms or energy as "instructions" and then declaring instructions solved.

The thread says management is "a workaround" and organisational hierarchy is "instruction debt." Some of it is. But a lot of what looks like coordination overhead — the meetings, the check-ins, the approval chains — is actually accountability wearing a lanyard. It's not that nobody knew what to do. It's that someone needed to be on the hook for the decision. AI doesn't delete that. AI makes it more visible.

To his credit, Anadkat names this. He calls it the "accountability premium" — the idea that when drafting and research get cheap, responsibility for outcomes becomes the scarce resource. That's the sharpest observation in the piece. But then he keeps going, past the insight and into a speculative vision of "labour operating systems" and "trust graphs" where jobs become acceptance tests verified by IoT sensors. A plumber's work stops being "fix the pipe" and becomes "holds pressure above 60 psi for 24 hours, verified by a sensor reading."

Anyone who has ever hired a contractor to do anything knows the gap between that sentence and reality. Most physical work outcomes are ambiguous, context-dependent, and argued about by humans standing in the room looking at the thing. That's not a temporary inefficiency waiting for better sensors. That's the nature of physical work.

The piece also has a tell: six company tags. Harvey, Cursor, Mercor, World Labs, Simile, and a cc to the Indian Prime Minister's office. When analysis starts reading like a portfolio map, adjust your priors accordingly.

The atoms-versus-bits distinction is the part worth keeping. Intelligence-bottlenecked work (law, software, finance) transforms at AI speed. Physics-bottlenecked work (construction, logistics, healthcare delivery) transforms at physics speed. If you're trying to figure out where your industry sits, that's the most useful two-sentence framework I've seen this month.

The rest is the standard move: observe what's happening in software, assume it generalises to everything, and skip the part where you explain why decades of failed attempts to systematise physical work and institutional trust will suddenly succeed now. The instruction layer is compressing. But the reason the hard parts are hard was never that we lacked good instructions.