When Average People Feel AI: Electricity for Knowledge Work, a Rounding Error Outside
Notes on Nathan Lambert’s When will average people feel AI’s impact? (Interconnects, 2026-09-09).
Core take
Comparing this boom to the industrial revolutions gets the scale of change right and the exposure path wrong. Those eras handed ordinary people cheaper clothes, indoor plumbing, electric light, bicycles—physical goods. Today’s AI already reshapes knowledge work, but family, food, transport, and entertainment barely moved.
Lambert spent weeks offline for a wedding and could have ignored AI entirely. For most people the upside is fun images and slightly better search; the downside is addictive feeds, chatbot-addicted acquaintances, and data-center politics. One line: AI is still a rounding error in everyday life. Being obsessed with it is a choice very few have made.
That is not a claim that the tech is fake. His harder claim: we are less than five years into a compounding revolution that could take a century. What matters now is infrastructure and a process that compounds. A major math breakthrough today will look small next to later compounding. Society will not grant patience on that clock.
Indirect gains, credit that never lands
Even the optimistic endgame—new science, rare-disease therapeutics, abundance—may be too indirect. How does a citizen credit OpenAI or Anthropic for a miracle cure announced by a family doctor? What share of Americans will care who solved Navier–Stokes?
In fifty years the average American’s home, appliances, relationships, and vehicles may still look familiar. Self-driving will keep diffusing on a largely independent track, not as a natural sequel to the LLM story. The industry will get a lot of credit. Daily life may not.
Much of the public narrative is trying to make people care because the long-term is real and the present feel is not. That mismatch is already political.
Engels’ pause: electricity for half the economy
Today’s AI is primarily a tool for elites and knowledge work—roughly half the U.S. economy. With agents improving fast, that half will treat AI like electricity. The other half stays stagnant. A booming tech economy next to an unchanged street is easy to read as “not a collective good.”
Lambert points to Engels’ pause (roughly 1790–1840): British GDP per capita rose while working-class wages stalled. If that is the closest analogue, people left out are right to push back. If industry leaders think this is the path, they should not be surprised.
Sharper still: AI is the greatest tool ever for scaling tech companies and online-native small businesses. He does not even expect tech to grow headcount through an era of massive success—headcount likely shrinks while knowledge-work output explodes. The already-winning sector wins harder. The brand gets worse. He worries instinctive backlash sends AI down the cautionary path of American nuclear power: kneecapped by its own story.
Two problems; solving either buys time
He compresses the first half-decade of a fifty-year diffusion into two issues:
- Early positive impacts are too indirect.
- Political backlash is the West’s Big Tech ledger, timed onto AI’s exponential. Had the exponential arrived decades later, data centers might never have sat at the political center.
Fix either and the industry buys time to show why status-quo economic change is worth it. Both are harder because the industry also self-labels as doom and mass unemployment. Leaders have started to course-correct; the public has not bought in.
If robotics and self-driving later join the same story, people will latch onto tangible benefits fast. Ironic, given how much effort went into arguing that LLMs are unlike the last decade or two of “general AI.” If robots later save—or overshadow—that story, it would be funny.
Growing pains, not the ending
This era looks like growing pains: society has to break habits that predate ChatGPT, which releases a lot of energy and anger. The fight against AI will outrun the diffusion story. Younger people following along will see powerful AI go from effectively 0% to 90%+ deep adoption in a lifetime. Business-integrated systems and personal assistants are only now becoming viable; they will take far longer to adopt than chat. That, not ChatGPT, is the real marker.
Two things are true at once: keep pushing the technology (the benefits are not automatic); distribute them widely. Serving only the half that can pay for tokens means the political bill arrives before the product bill.
For builders, the reusable idea is two clocks: inside the bubble, software is being rewritten (systems of record, evals, agent workflows). Outside, daily life has not been rewritten. Time, money, health, and transport that people can feel will buy the industry more patience than a solved millennium problem.
Comments