Noah Smith: AI can transform productivity even if intelligence has diminishing returns

In a Noahpinion essay, Noah Smith asks why rapidly improving AI has not yet produced an economic singularity. He examines possible limits to intelligence, then describes three computer-native sources of productivity: replicable machine cognition, captured tacit knowledge, and “cloud laws.”

AI can solve difficult mathematical problems, while jobs and daily routines look much as they did before the latest capability gains. Smith calls AGI or ASI already present, but says its arrival has been incremental.

Ruxandra Teslo points to governance and other frictions. Clifford Sosin argues that current systems are powerful tools whose intelligence did not unlock immediate control over physical and economic reality.

François Chollet describes intelligence as a bounded conversion ratio. Once a reasoner is near the optimum, each improvement adds less.

Smith gives two possible limits. Observations are finite, so a model cannot extract information that was never measured. Complex systems can also be chaotic: small measurement errors grow until reliable prediction becomes impossible. Arvind Narayanan, Sayash Kapoor, and Sosin make related arguments about irreducible error and the boundary between reasoning and new facts.

Machine cognition can be copied. More data centers can run more agents in parallel, while the supply of human minds depends on population. Robots extend that capacity into physical work, and remote compute can control machines at many scales.

Smith expects a higher capital-to-labor ratio. Each person could direct many intelligent machines without any one machine becoming incomparably smarter than a human.

Zeiss mirror production and Chinese rare-earth refining depend on thousands of small techniques spread across workers and organizations. Much of that knowledge is not written down or consciously understood by the people using it.

Smith proposes recording work through glasses, gloves, and plant sensors. AI could combine those observations, suggest process changes, and learn from the results. If the loop works, lagging firms could adopt better methods faster and frontier firms could shorten production experiments.

LLMs reproduce human language without giving researchers a simple theory of how language works. Smith calls similarly useful but hard-to-explain regularities “cloud laws.”

He suggests that AI might find them in social systems, plasma, materials, or turbulence. A machine can hold patterns that human teams must split into smaller pieces for communication. If those regularities exist, AI could make parts of nature controllable without first turning them into human-readable laws.