Historical Context

On Tap Intelligence

Every great technology has transformed work. Here’s what history tells us about what comes next.

AI is doing for cognitive capabilities what electricity did for physical power — turning a scarce, expensive resource into an on-demand utility available to anyone.

8 min read
01

Every GPT Follows the Same Arc

Every major general-purpose technology follows a predictable five-phase arc. The names change, but the shape is the same. Steam, electricity, computers — each transformed the labor market through the same sequence of emergence, diffusion, displacement, reorganization, and new equilibrium.

AI is here

Click or hover on a phase to learn more

10x faster

AI is diffusing at roughly 10x the speed of prior general-purpose technologies.

ChatGPT reached 100 million users in 2 months — the internet took 7 years, the PC took over a decade. By 2025, 54.6% of US working-age adults had used generative AI (St. Louis Fed), a penetration the internet didn’t reach for 5+ years and the PC didn’t reach for 15. Enterprise adoption surged from 33% to 88% of organizations in roughly two years (McKinsey, 2023–2025). Microsoft’s AI Diffusion Report calls it “twice the pace of smartphones, five times faster than the internet” — and those technologies were themselves orders of magnitude faster than steam or electricity. See live adoption data →

The compounding is staggering: electricity took 40 years from dynamo to productivity gains. The PC took 30+ years from mainframe to the Solow Paradox resolving. AI went from GPT-3 to 88% enterprise adoption in 5 years. This is not incremental acceleration — it is an order-of-magnitude compression of the diffusion phase. One reason: AI is built on top of two prior GPTs (computers and the internet) that already saturated the economy. Every smartphone is already an AI terminal. See enterprise adoption data →

The forecasts on this page extrapolate from this 10x adoption speed. If the diffusion phase that historically took 10–25 years is happening in 1–3, the displacement and reorganization phases will compress proportionally. The constraint: organizational restructuring, education systems, and policy still operate at human speed — but even these are being accelerated by AI itself.

02

The Four Revolutions

Four technologies. Four massive disruptions. All eventually created more jobs than they destroyed — but the path was never smooth or quick.

ElectrificationClosest AI Analog

1880–1930
The Innovation
Transformed power from a scarce, locationally-fixed resource into a ubiquitous, on-demand utility
What It Automated
Centralized shaft-and-belt power distribution; many domestic labor tasks
Jobs Destroyed
Millwrights, shaft-and-belt mechanics, specific factory roles tied to the old organizational form
Jobs Created
Electricians, electrical engineers, the entire consumer appliance industry (radio, refrigeration, washing machines), domestic electrification created conditions for women’s mass labor force entry
The Painful Part
Paul David’s “Productivity Paradox” — electric dynamos were introduced in the 1880s but didn’t show up in productivity statistics until the 1920s. 40 years.
The Lesson
The technology isn’t the bottleneck — the organizational, educational, and institutional ecosystem surrounding it is. On-tap power democratized access to energy in ways that shifted competitive advantage from those who owned power infrastructure to those who used it most intelligently.
AI Parallel

This is the on-tap intelligence moment. AI transforms cognitive capabilities from scarce expert resources into utilities. The productivity gains will arrive later than expected, and through organizational redesign more than simple substitution.

03

The On-Tap Intelligence Shift

Prior automation technologies had a consistent structure: they automated physical capabilities (steam, combustion, electricity) or rule-based cognitive tasks (computers). Each wave created a new protected domain — work the technology structurally couldn’t do — that workers could move toward. The task model of labor (Autor, Levy & Murnane, 2003) categorized this as the difference between routine tasks (codifiable, automatable) and non-routine tasks (requiring judgment, context, creativity).

AI breaks this pattern. Large language models and multimodal AI systems perform tasks that are simultaneously cognitive and ostensibly non-routine: legal analysis, medical reasoning, strategic synthesis, creative writing, code generation. This doesn’t mean AI equals human intelligence — it doesn’t — but it means the protected domain of prior automation waves is now being encroached on.

The most useful analogy is electrification. Before electricity, accessing significant mechanical power required physical proximity to a power source — a river, a steam boiler. Power was scarce, locationally fixed, and expensive. Electrification transformed power into a utility: standardized, reliable, available on demand anywhere on the grid, priced per unit of use. On-tap power.

AI performs this same transformation for cognitive capabilities. Legal analysis was previously accessible only to those who could pay $400/hour for a lawyer in a major city. Medical reasoning lived in academic medical centers. Strategic insight required expensive consulting firms. AI threatens to make these capabilities available to anyone with internet access, at marginal cost approaching zero. On-tap intelligence.

From Scarce to On-Tap

Physical Power
Access
Before
Near a water wheel or steam engine
After
Any electrical outlet
Cost
Before
High capital, fixed infrastructure
After
Per unit of use
Geography
Before
Mill towns, factory districts
After
Anywhere on the grid
Competitive advantage
Before
Owning the power infrastructure
After
Using power intelligently
Cognitive Capability
Access
Before
Expensive credentialed experts
After
Any smartphone
Cost
Before
$200–$700/hour
After
Near zero marginal cost
Geography
Before
Major metros with legal/medical/finance clusters
After
Anywhere with internet
Competitive advantage
Before
Accessing expert talent
After
Using AI intelligently
04

What the Pattern Predicts

History doesn’t tell us the outcome — it tells us the shape. Here is what the pattern predicts, offered not as certainties but as the most historically-grounded expectations.

Near Term

1–3 yearswas 5–15 years
  • Labor market polarization accelerates — already visible in 2024–2025 hiring data for content, coding, and customer service roles
  • Routine cognitive work faces displacement pressure first: document review, standard writing, basic code, data analysis, customer service
  • Wage compression and employment decline appear in these categories as AI-assisted workers handle dramatically more volume
  • New roles in AI oversight, training, and application emerge — but won’t immediately compensate for losses
  • With 88% of organizations already adopting AI (McKinsey, 2025), the Solow Paradox may resolve in years rather than the decades it took for electricity and computers

Medium Term

3–7 yearswas 15–30 years
  • The productivity paradox resolves — organizational complements to AI develop (new business models, new processes, new educational pathways, new regulatory frameworks)
  • Aggregate productivity growth becomes visible and accelerates — potentially arriving by the early 2030s rather than the 2050s a historical baseline would predict
  • Whether this growth translates to broadly shared wages depends entirely on institutional choices being made now
  • AI’s low infrastructure requirements could spread gains more geographically than prior GPTs, which concentrated in industrial centers

Long Term

7–15 yearswas 30–50 years
  • A new occupational equilibrium emerges, dominated by human-AI collaborative roles
  • Growth in human-specific services: caregiving, high-stakes physical tasks, relational work that requires accountability and presence
  • Entirely new industries emerge, enabled by democratized access to cognitive capabilities — analogous to how electrification created the consumer appliance economy
  • Average wages in the new equilibrium may be substantially higher — but the compressed transition means less time for workers and institutions to adapt

Historical calibration: In prior GPT transitions, the total timeline from meaningful diffusion to new equilibrium ranged from 40 to 70 years. AI’s diffusion phase is running at roughly 10x the speed of prior GPTs: enterprise adoption went from 33% to 88% in two years (McKinsey), and 54.6% of US working-age adults used generative AI within three years of launch (St. Louis Fed) — penetration rates the internet and PC took 5–15 years to achieve. The timelines above extrapolate from this 10x adoption speed. If the historical 40–70 year arc compresses proportionally, the full transition from diffusion to new equilibrium could play out in 7–15 years — a single generation rather than three.

05

Occupational Vulnerability Snapshot

Five vulnerability categories, grounded in the historical pattern of how general-purpose technologies reshape occupational structures.

Note: Vulnerability does not mean elimination — it means transformation. Most “displaced” occupations in prior GPT transitions were reorganized, not eliminated. The exception: occupations where the technology directly substitutes for the core input (handloom weaving → power loom; data entry → AI processing). For these, the historical record offers little reassurance.

06

What History Actually Proves

The technology doesn’t decide. We do.

Lesson 1: Invest in Complements, Not Preservation

Every successful institutional response to a GPT transition invested in workers’ capacity to participate in the new economy, not in protecting the old one. The Morrill Acts (public universities, 1862) equipped workers for the industrial era. The GI Bill (1944) prepared workers for the postwar economy. The community college system extended computing-era skills broadly. The equivalent for AI: radical investment in AI literacy, domain-expert + AI collaboration skills, and accessible retraining pathways.

Lesson 2: The Distribution Problem is Institutional, Not Technological

The computer era’s inequality — 40 years of wage stagnation for workers without degrees — was not technologically inevitable. It reflected specific choices: declining union density, wage policy, trade liberalization, corporate governance norms. The AI era’s distributional outcome will similarly reflect choices. Ford’s Five Dollar Day (1914) is the counterexample: he doubled wages deliberately because mass production requires mass consumers. If AI dramatically increases productivity, the economic stability of the outcome depends on how those gains are distributed.

Lesson 3: The Gains Are Real. The Timeline is Not What You Think.

Every GPT ultimately created more jobs than it destroyed and raised average wages. This is true and important. It is also true that the prior pattern’s timeline — 40 to 70 years — is politically and humanly unacceptable as a response to workers experiencing disruption today. “Eventually” is not a policy.

If on-tap intelligence enables the democratization of expertise — putting the equivalent of world-class legal, medical, educational, and financial guidance within reach of everyone rather than only the affluent — it could be among the most equalizing forces in human history. If it primarily displaces workers while concentrating gains among capital owners and a small elite of knowledge workers, it could be among the most destabilizing. The technology does not decide. We do.