AI adoption runs through four stages—and the slowest one has barely begun

"Normal" in AI as Normal Technology does not mean AI is a hammer or a toothbrush. They treat it as industrial-revolution-scale tech. The point is a causal frame: how capability becomes economic and social impact, step by step.

Past work on electricity-style tech often uses diffusion of innovation: invention → innovation (downstream products like appliances) → diffusion (gradual adoption). They stretch that into four stages, using software engineering as the running example:

Diffusion applied to AI: capability → product → early adoption → adaptive change. Source: talk slides

Stage four is the slow one. Even in software engineering—relative early adopters of coding agents—he says real org-level redesign has barely started. He allows a speculative beat: if agents can reliably ship multi-million-line codebases with few bugs and security holes, building one uniform product for a billion users gets less compelling. Software may tilt toward extreme per-team customization, and even "do we still need software companies as an org form?" becomes fair game. Those shifts are human and organizational, and history usually measures them in decades.

Electricity in factories: drop-in replacement never worked

Before electricity, factories ran on one giant steam engine, powering the plant through gears and belts. When electricity arrived, owners first swapped a generator for the steam boiler—hoping for a more efficient drop-in replacement. That path failed.

From steam drive to the assembly line: electricity's real payoff came from redesign, not swapping boilers. Source: talk slides

People often say agents will be drop-in replacements for human workers. Electricity's lesson: the real payoff came from reorganizing work, not swapping humans for machines one-for-one. And that was not the utility company's job—same for AI. Org redesign is not something AI labs finish alone. In their four-stage frame, it is the slowest stage, and today it has barely begun.

Measured: capability rose; reliability did not keep pace

Across industries, a huge gap sits between what AI could do and what people actually deploy. Slow adoption is part of it. They also suspect deployers hit hard walls that leaderboards never measure—before the rest of the industry does.

Reliability is the concern people name most often. SAGE collapses roughly a dozen reliability metrics into four dimensions:

AI agents: the capability–reliability gap. Source: talk slides · SAGE work

Software engineering: writing code was never the bottleneck

A 2019 paper already argued that writing code is not software engineering's bottleneck. The past year, blogs rediscovered that. Coding agents sped up the middle layer—the whole job did not shrink by the same factor.

The decide–execute–deliver "sandwich": AI mostly bites the middle. Source: talk slides

A useful metaphor: the crane operator. Source: talk slides

The lump-of-labor fallacy: automation rarely cuts jobs one-for-one

The lump-of-labor fallacy treats work as a fixed pie: if AI takes a slice, jobs vanish forever. History often says otherwise—when efficiency rises, demand and job structure shift with it.

The lump-of-labor fallacy and a few historical checks. Source: talk slides

Software itself: from machine code onward, successive ~10× productivity-tool leaps came with ~10,000× more jobs—because total code to write grew faster still. He is not saying "no one will ever lose work." He is saying "automation rate = unemployment rate" is a bad extrapolation.

If recursive self-improvement actually arrives: four dimensions people mash together

Many labs openly chase recursive self-improvement (RSI). He takes that path seriously—without equating a lab milestone with "humans immediately have nothing left to do."

The "Hawaii problem" in AGI talk: from far away it's one island; up close you need the names. Source: talk slides

Four dimensions of advanced AI: none automatically entails the others. Source: talk slides

Example: curing cancer is often bottlenecked by trials with thousands of subjects over 10–15 years—external constraints more FLOPs in the lab will not erase. Mapping compute or model milestones straight onto "society immediately out of work" misses those walls.

Why AI creativity lags—and how open-world evaluation measures it

Deep learning already excels at perceptual representations. Representations that support creativity and higher-order reasoning, he argues, still trail humans. A few hypotheses from cognitive science and practice:

On the AI creativity paradox. Source: talk slides

SAGE's open-world evaluation: give an agent a few thousand dollars of budget plus a real ML problem a human expert already spent months on and wrote up—but not yet on arXiv—then have those same experts grade the output. The team has already run evaluations like "have an agent independently build and ship an iOS app," and is recruiting senior researchers to expand.

How should individuals adapt in this wave?

Framework and evidence first. In the second half—"Personal reflections on adapting to AI"—Narayanan does not hand out one answer. He opens his research workflow: how he is riding this capability surge, for the audience to check against their own.

Personal reflections on adapting to AI. Source: talk slides

First, pick a stance: displacement vs amplification

Those opening stories become two life configurations at the personal level:

Floor vs ceiling: where saved time goes

He frames personal strategy with a pair of words:

His move: when AI clearly lifts productivity, reinvest the saved time in long-term growth—new skills and workflows that complement AI. He spends roughly 10 hours a week only on learning and trying new processes.

Three-legged stool: productivity · growth · control

Faster delivery with AI is not the only goal. Three legs have to stand at once:

Push only productivity and starve growth, and a rising floor flattens you. Chase only growth and ignore delivery, and reality ends you. Have both but surrender control, and long-term you become a button-pusher.

Two heuristics for keeping control

The job is redefined as operating and controlling the machine. Source: talk slides

Closing vision: human–machine co-superintelligence

In one sense, economically transformative AI has already started: not when some AGI switch flips, but as slow variables—reliability, integration, tacit knowledge, regulation—rewrite work. He rejects geopolitics reduced to "whoever hits a capability milestone first takes all the economic returns."

On superintelligence he stresses: tasks often have ceilings; human intelligence leans hard on learning and tools, and AI is another tool—so the contest looks more like AI-augmented people vs AI acting alone. If we default to a future where AI owns companies and decides hiring and firing, then lean only on alignment as a backstop, that safety posture "opposes safety more than it supports it."

Vision: co-superintelligence. Source: talk slides

He likens computers to "bicycles for the mind" and AI to "cranes for the mind"—lifting human potential to heights once unimaginable. The learning curve is steep, like a treadmill that never stops. He still treats co-superintelligence as a fight worth fighting: not abandoning work, but redefining it as a higher-ceiling dance with AI.

Talk takeaways slide. Source: talk slides

The capability floor rises on its own; you have to push the ceiling. Adaptation is not dumping all work on AI—it is reinvesting saved time in complementary skills, and staying in the control seat. Based on Arvind Narayanan · ICML 2026 keynote

In 24 months, capability jumped; reliability barely moved

SAGE folds a dozen-plus reliability metrics into four buckets—consistency, robustness, calibration, operational safety—then scores models released by three frontier labs over ~24 months on two complementary benchmarks.

Programmers aren't vanishing—writing code was never the bottleneck

A 2019 paper already said writing code is not software engineering's bottleneck. Narayanan splits knowledge work into three layers: decide (set direction), execute (code and debug), deliver (own the release).

Productivity gains rarely cut jobs one-for-one

Economists call it the lump-of-labor fallacy: treat total work as fixed, so AI taking a slice permanently deletes jobs. History often runs the other way.

RSI does not automatically snowball into superintelligence

Many labs say they are chasing recursive self-improvement (RSI: AI helping improve the next AI). Narayanan's "Hawaii problem": once the ship nears the archipelago, you need island names—or you get lost on where to sail. He splits RSI, human-level AI, economically transformative AI, and superintelligence into four non-entailing dimensions.

• RSI — the spectrum is huge: upgraded AutoML on one end, replacing the creative sum of the world's AI researchers on the other. Which end companies mean is often unclear.
• Human-level AI — strong on verifiable tasks (speed, efficiency); still far behind on unverifiable creative work.
• Economically transformative AI — bottlenecks are reliability, system integration, domain tacit knowledge, and regulation; decades of gradual adoption chew them down.
• Superintelligence — many tasks have ceilings (weather chaos limits; 10–15-year drug-trial protocols). Humans already sit near those limits.