How fast will AI show up in growth?

20 min read Original article ↗

An interactive reference

AI is moving fast. The economy is slower.

For 150 years, US living standards rose ~2% a year, through electricity, the transistor, and the internet. AI may be another general-purpose technology. The question is how fast task-level progress becomes measured growth, and how much depends on the weak links still handled by people.

I built this on Charles I. Jones's weak-links view of AI and growth, especially Jones & Tonetti's automation forecast, and set it against the wider recent evidence on how fast AI is actually moving: the capability curve, the 2% line, computers' shrinking GDP share, weak-link business chains, the spectrum of expert forecasts, the objections I find hardest to answer, and role horizons I update as the evidence comes in.

~100M

ChatGPT users

in ~2 months, fastest ever at the time

150 yrs

of ~2% growth

through every prior transformative tech

4.3% → 3.0%

computer share of GDP

falling since 2000, despite more chips

20+ yrs

self-driving diffusion

from “solved in 5” to still-rare

The puzzle

Transformative technologies, and still 2% a year.

Electricity, internal combustion, antibiotics, semiconductors, and the internet all changed everyday life. Measured growth still stayed close to its long-run path: each wave kept the 2% line going as the previous one ran out of steam.

The chart that surprised me

Computers are everywhere, so I assumed their slice of GDP had grown. It has fallen by about a third since 2000. Prices dropped faster than we bought more, so the abundant input got cheap while the scarce inputs, people, trust, coordination, kept most of the value.

That is the weak-link pattern in one chart: making an input cheap is not the same as automating the work it sits inside.

The capability curve

The models really are getting faster.

None of this means the capability is standing still. METR measures the length of task an AI agent can finish on its own, and that horizon has roughly doubled every six to seven months since 2019. Capability is on a genuine exponential. The weak-link claim is not that AI is slow, it is that fast capability meets slow diffusion, and the gap between the two curves is where the remaining human links live.

Frontier models (50% reliability) ≈6-month doubling, dashed forward

ModelsGPT-23sdavinci-0029sGPT-3.536sGPT-44mGPT-4o7mClaude 3.5 Sonnet11mo1-preview20mo139mClaude 3.7 Sonnet1ho32hGPT-53.4hGPT-5.25.9hClaude Opus 4.612h

Source: Task-completion horizons use METR's public Time Horizon 1.1 YAML (p50 estimates for SOTA frontier points). METR notes that measurements above 16 hours are unreliable with the current task suite, so the dashed forward trend is illustrative. Compute and cost context from Epoch AI and the Stanford AI Index. · METR Time Horizon 1.1 dashboard · Epoch AI (2024), Can AI Scaling Continue Through 2030? · Stanford HAI, 2025 AI Index

≈6 mo

capability doubling

METR all-time p50 fit

4×/yr

training compute growth

with room to run to ~2030 (Epoch)

280×

cheaper inference

GPT-3.5-level in ~2 years (AI Index)

~1 yr

to add 2% to income

the economy's long-run pace

Three of these clocks run in months and one runs in years. That mismatch is the whole argument: the inputs are sprinting while measured growth keeps near its 2% pace, because output still waits on the links that have not been automated.

Where we are today

Dates, predictions, and bottlenecks.

The page tracks claims that can be checked: adoption milestones, failed forecasts, and places where the remaining bottleneck is still human. The pattern so far: adoption can be very fast while economic transformation is slower.

Milestone timeline

  1. Mar 2004Verified

    DARPA Grand Challenge: zero finishers

    Not a single autonomous vehicle completed the desert course. The starting gun for self-driving, and a reminder of how hard the physical world is.

  2. Oct 2005Verified

    Stanford's "Stanley" wins the DARPA Grand Challenge

    One year after zero finishers, Sebastian Thrun's team completed the 132-mile course. Twenty years later, robotaxis are still rare outside a few cities.

  3. Sep 2012Verified

    AlexNet ignites the deep-learning era

    A deep neural network crushed the ImageNet benchmark, kicking off the modern wave of AI capability gains.

  4. Mar 2016Verified

    AlphaGo defeats Lee Sedol

    DeepMind's system beat a top human Go player 4–1, years ahead of expert expectations for the game.

  5. Jun 2020Verified

    GPT-3 released as a developer API

    A 175B-parameter model showed broad few-shot ability, but reached developers, not a mass consumer audience.

  6. Nov 2020Verified

    AlphaFold cracks protein structure prediction

    At CASP14, AlphaFold2 reached near-experimental accuracy, transforming structural biology. The work earned a 2024 Nobel Prize in Chemistry.

  7. Nov 2022Verified

    ChatGPT launches

    OpenAI released a conversational interface over GPT-3.5. Adoption was almost immediate.

  8. Jan 2023Verified

    ChatGPT reaches ~100M monthly users in ~2 months

    The fastest-growing consumer app at the time. NOTE: this was ChatGPT (GPT-3.5), not GPT-3. Adoption can be blindingly fast even when economic transformation is slow.

  9. Mar 2023Verified

    GPT-4 released

    A large multimodal model with markedly stronger reasoning and coding, the workhorse behind the first wave of AI copilots.

  10. Aug 2024Verified

    Waymo scales paid robotaxi rides

    Driverless rides became a daily reality in San Francisco and Phoenix, yet remained rare nationally. Diffusion measured in decades, not years.

  11. Jun 2026Projection

    Today: you are here (mid-2026)

    Adoption is nearly universal; the economy-wide productivity jump is still small. Lightning-fast in a few lanes, slow across the rest, exactly the split the weak-link view predicts.

    How fast is AI?, present-day marker

Signals to watch, not yet verified

Forward-looking markers, kept out of the verified record. Treat each as a claim until a primary source confirms it.

  1. Nov 2025Claim

    A frontier model reportedly tops an engineering take-home exam

    A leading lab's multi-hour take-home hiring exam was reportedly completed by its newest model at a score higher than any human on record. Treat as a claim pending public confirmation.

    Reported, pending verification

  2. Dec 2025Claim

    Coding agents start holding context across multi-hour tasks

    The frontier shifts from quick answers to agents that run for hours, the first credible long-horizon autonomy in software.

    Unverified signal

  3. Feb 2026Claim

    Enterprises move AI agents into real workflows

    Pilots turn into production: support, coding, and back-office agents handle a slice of real volume while humans still own the judgment calls. Adoption races ahead of measured productivity.

    Unverified signal

Prediction scoreboard

Geoffrey HintonOver-predicted

“"We should stop training radiologists", none will have jobs within five years.”

By 2026 there are MORE radiologists than in 2016, and they are paid more. AI reads scans; humans handle the weak links: consults, hard cases, sign-off.

Lesson: weak linkssource

Elon MuskOver-predicted

“A Tesla will drive itself coast-to-coast by 2017, with full self-driving roughly two years away.”

A decade later, supervised "Full Self-Driving" still needs a human behind the wheel, and driverless robotaxis run in only a handful of cities. The physical world is full of weak links.

Lesson: diffusion takes decadessource

Almost everyoneUnder-predicted

“A chatbot won't reach hundreds of millions of users quickly.”

ChatGPT hit ~100M monthly users in ~2 months, the fastest consumer-app adoption on record at the time.

Lesson: adoption can be fast even when transformation is slowsource

AI-2027 / Situational AwarenessStill open

“Explosive, economy-wide acceleration arrives within ~3–4 years.”

Still open. The weak-link view expects real transformation measured in decades, with the downside risks arriving sooner.

Lesson: horizon is the cruxsource

Find the bottleneck, by domain

For each field: what AI has automated, and what remains hard to automate. Automating 75% of the tasks doesn't finish the job. The rest set the pace. Each card carries a real-usage signal from the Anthropic Economic Index, which maps millions of actual AI conversations onto real-world tasks.

Software engineering

The first thing being automated, and still bottlenecked by judgment and accountability.

Automated

  • Autocomplete & boilerplate
  • Test generation
  • Bug discovery in mature code
  • Routine refactors

Weak link

  • System design & architecture
  • Ambiguous requirements
  • Integrating with messy production data
  • Owning the outcome when it breaks

Real usage · Computer and mathematical tasks are the largest slice of current Claude usage (~19.8%), with software developers leading individual job-title usage.

Even if AI writes most code, integrating it into every business is a long, engineer-heavy process.

Radiology

Hinton's 2016 test case. AI got better at reading scans; the job grew anyway.

Automated

  • Scan triage
  • Cancer-detection assist
  • Measurement & flagging

Weak link

  • Surgical & treatment consults
  • The hardest, ambiguous scans
  • Liability and sign-off
  • Talking to patients and clinicians

Real usage · Healthcare-practitioner usage is tiny in the current index; radiologic technicians register at 0.00% of mapped Claude conversations.

More radiologists in 2026 than 2016, and better paid. Weak links win.

Driving

Seemingly simple, actually decades-long. The canonical slow-diffusion case.

Automated

  • Highway driving
  • Mapped-city autonomy (SF, Phoenix)
  • Sensing & lane control

Weak link

  • Long-tail edge cases
  • Unmapped regions
  • Adverse weather
  • Nationwide scale & trust

Real usage · Transportation and material-moving roles barely register in software-based AI usage data; many physical occupations show 0.00% usage.

20+ years from "solved in 5" to still-rare. The physical world bottlenecks.

Early-childhood teaching

A 'someday' automation gated almost entirely by trust and safety, not capability.

Automated

  • Content delivery (potential)
  • Practice & feedback (potential)

Weak link

  • Trust & safety with children
  • Care and supervision
  • Parental acceptance
  • Liability

Real usage · Educational instruction and library tasks are a visible usage category (~7.1%), but the index shows tutoring and content work, not classroom supervision.

We could build a world-class teaching robot before we'd let it run a kindergarten unsupervised.

Source: Domain bottlenecks are site synthesis from the weak-link framework. The self-driving cars slow-diffusion example follows Jones's discussion; real-usage signals are read from the Anthropic Economic Index; item-level timeline and prediction cards carry their own factual sources. · Jones, “A.I. and Our Economic Future” (forthcoming, JEP) · Jones & Tonetti (2026), “Past Automation and Future A.I.” · Anthropic Economic Index

Field evidence: real gains, unevenly shared

Controlled studies find large average productivity gains from generative AI, concentrated where skill was scarce. The strongest workers, the hardest links, move least. That is the weak-link pattern showing up in the labor data.

+14%

issues / hour

support agents, on average

+34%

for novices

the gains land where skill was scarce

−40%

writing time

mid-level tasks, quality up 18%

minimal

for experts

the hardest links barely move

Source: Customer-support field experiment from Brynjolfsson, Li & Raymond (2025); writing experiment from Noy & Zhang (2023). Figures are the studies' reported effects. · Brynjolfsson, Li & Raymond (2025), Generative AI at Work · Noy & Zhang (2023), Science

The forecast

Calibrate the model, then run it forward.

The Jones & Tonetti weak-link automation model is calibrated to history and run forward. Even the aggressive “Moore's Law everywhere” case takes about 30 years to play out, because output depends on the links that have not been automated yet.

1.04×

richer by 2050

Continuing-the-past baseline

1.38×

richer by 2100

slow compounding

3.2×

richer by 2040

Moore’s-Law-everywhere case

~2050

fast case matures

after many bottlenecks move

Where this model sits among the experts

Jones & Tonetti are one view on a wide dial. Here is the same question, how fast AI changes growth, answered from the skeptical end to the explosive end. Hover any view to read its claim.

Slower changeFaster change

Daron AcemogluJones & TonettiTrammell & KorinekErdil & Besiroglu

Skeptical<0.53% TFP

Daron Acemoglu

The conservative anchor: predicted TFP gains over the next 10 years are below 0.53%, with even the initial task-based estimate no more than 0.66%. Real, but small.

macro estimateAcemoglu (2024)

Gradualweak-link delay

Jones & Tonetti

Weak links cap the gains. In the extreme Moore's-Law-everywhere calibration, growth exceeds 7% by 2030 and 13% by 2040, but the path still depends on whether essential human-only tasks remain.

model calibrationJones & Tonetti (2026)

Conditionalwide range

Trammell & Korinek

Surveying the field: fully automating production can break the old growth regularities and raise the growth rate, but the range of plausible outcomes is genuinely wide.

survey paperTrammell & Korinek (2023)

Explosive~10x growth

Erdil & Besiroglu

The rigorous explosive-growth case: broad automation could make growth roughly an order of magnitude faster, then the paper tests that against nine named counterarguments.

review paperErdil & Besiroglu (2023)

What it means for you

Human bottlenecks move last.

Even in the aggressive scenario, the last tasks to automate are the links with accountability, trust, physical presence, or judgment. Pick a role or business and see where this model puts those links on the timeline.

When does AI stop needing you?

Routine work goes first. What keeps you in the loop is the link AI is worst at: owning the call.

Needed until

~2046

That's about 20 more yearsof being the weak link AI can't replace.

AI already does~78%

of the routine work to close the books and sign off on the numbers.

What keeps you needed

  • Edge-case judgment
  • Audit accountability
  • Client trust

The routine ledger work goes first. Putting your name on the numbers, and answering for them, stays human longer than the spreadsheets do.

A thought experiment, not a forecast. The years are round anchors on the aggressive scenario: the model's fast case matures around 2050, and links such as accountability, care, physical presence, and the human handshake are placed later, stretching toward 2060. Some may never fully disappear, which is the “if ever” in the question.

Source: Horizon years are site-created illustrative translations of the weak-link automation model, not Jones/Tonetti job forecasts. · Jones & Tonetti (2026), “Past Automation and Future A.I.” · “A.I. and Our Economic Future,” Professor Chad Jones (Stanford GSB)

Note: a “safe until” year is when the remaining human bottlenecks start to matter in the model, not a cliff where the role vanishes overnight. Automation erodes tasks gradually, and can even grow a field: there are more radiologists now than in 2016, not fewer. These are round, illustrative anchors, not job forecasts. Some links, accountability, care, physical presence, may never fully disappear.

The inputs will move.

I track capabilities, adoption, and the bottlenecks that still do not move much. Follow along on LinkedIn.

Subscribe on LinkedIn

The fast downside

The downside doesn't wait for the upside.

The same weak-link structure that delays the benefits can make some risks arrive sooner. Strengthening a chain is slow, link by link. Breaking one link is faster.

The bad actor with a jailbroken oracle

Near-term

Frontier models keep getting jailbroken, often within days of release. Hand a bad actor a model that can do what the smartest humans can, and ask it to design a pathogen more lethal than Ebola with a three-month latent period. We survived nuclear weapons because only a handful of people held the button. What happens when billions do?

Open-source bug-hunters turned on the grid

1–3 years

Models are already finding bugs humans missed in decades-old, battle-tested software. It is not hard to imagine a capable open-source version in many more hands within a year or two. How sure are we no one points it at the electric grid, the banking system, or a bio lab? Not an existential problem, but a plausible one, soon.

Retaining power over smarter entities

Speculative

When more advanced species have met less advanced ones in history, it hasn't gone well for the less advanced. Stuart Russell's question is the uncomfortable one: how do we retain power over entities more capable than us, forever?

Slow to improve, fast to fail

Structural

The same weak-link structure that makes the upside slow makes the downside fast. A chain takes enormous effort to strengthen link by link, but breaking a single link destroys the value instantly. The Space Shuttle Challenger was lost to a $25 O-ring. That asymmetry is the core warning.

Source: The “one broken link destroys the value” fragility draws on Kremer’s (1993) O-Ring theory; the slow-to-strengthen, fast-to-break asymmetry follows the weak-link view in Jones & Tonetti (2026). · Kremer (1993), O-Ring Theory · Jones & Tonetti (2026), “Past Automation and Future A.I.”

Strongest objections

The best arguments against this view.

This page is useful only if its assumptions are visible. These are the objections I think matter most, with responses and what would change the forecast.

This is the crux, and it lands. The forecast's slowness comes from calibrating to a past where the hard tasks never moved much. If AI keeps climbing at exactly the work we've labeled “human,” the gradual story collapses and the upside arrives far sooner. The honest position: the speed depends entirely on how strong those weak links really are, which is the one number we're least sure of.

The gradualism is an assumption, not a law. The cognitive weak links may not hold.

What remains scarce?

Scarce links capture value.

When one input becomes abundant, value often moves to the inputs that remain scarce: judgment, accountability, trust, physical presence, and capital ownership. This is the practical implication of the model.

Be the manager

Someone still decides which output to trust and what tradeoff to make. That judgment matters more when the cost of analysis falls.

Own the capital

If capital captures more of the gains, broad ownership matters more. A slice of the market is one way households share in that shift.

Redistribute deliberately

Abundance makes good outcomes possible, not automatic. Tax-and-transfer is a policy choice, not a law of the model.

Protect non-work meaning

If some work becomes optional, status and purpose still have to come from somewhere: craft, community, family, learning, and taste.

For change leaders: the same story at company scale

The weak links that bottleneck a whole economy have the same shape as weak links inside an organization: handoffs, identity, trust, incentives, and accountability.

Macro storyOrganizational analog
Normal tech vs FOOMTheory of change vs theory of changing
Continuation vs breakFirst-order vs second-order change
Weak links bottleneck gainsSensemaking, identity, coordination cost, OCC
Diffusion takes decadesTsoukas & Chia: organizational becoming
Fragile downsidePerforming → reinforcing → breakdown → reflecting
Scarce links earn returnsResistance as signal, not obstacle

The practical move: sequence your automation across the process chain and treat employee resistance as a signal that locates your weakest link, not as friction to overcome.

The library

The research behind the model.

The claims here trace back to published work, and where a number is an estimate, digitization, or projection, the page says so. The core weak-link framing, forecast model, computer-share figure, 150-year growth frame, infinite-automation result, and self-driving example draw directly on the work of Charles I. Jones and Jones & Tonettiwhile Daron Acemoglu's estimates anchor the skeptical end. Alongside them sits a wider, recent literature on how fast AI is actually moving: capability and compute trends, broader growth models, and field studies of productivity.

Economic growth

  • Charles I. Jonesin prep.

    A.I. and Our Economic Future

    Working paper, in prep. for the Journal of Economic Perspectives

    Direct non-technical source for the infinite-automation 1/(1-s) result, the self-driving-cars example, and the weak-link view of AI growth.

  • Charles I. Jones2026

    "A.I. and Our Economic Future," Professor Chad Jones

    Stanford Graduate School of Business (YouTube)

    Stanford GSB talk presenting the weak-link growth view and figures adapted on this page, including the computer-share and forecast charts.

  • Paul M. Romer1990

    Endogenous Technological Change

    Journal of Political Economy 98(5): S71–S102

    Ideas as the engine of long-run growth (Nobel 2018). The flywheel that 'wants to explode'.

  • Bloom, Jones, Van Reenen & Webb2020

    Are Ideas Getting Harder to Find?

    American Economic Review 110(4): 1104–1144

    Within any technology, ideas get harder to find (the steam engine runs out of steam), so each wave buys ~50 years of 2% growth.

  • Aghion, Jones & Jones2019

    Artificial Intelligence and Economic Growth

    in The Economics of Artificial Intelligence, Univ. of Chicago Press (NBER WP 23928)

    The earlier automation + weak-links growth model behind the baseline forecast scenarios.

  • Daron Acemoglu2024

    The Simple Macroeconomics of AI

    Economic Policy 40(121): 13–58 (NBER WP 32487)

    The conservative end of the dial: predicted TFP gains below 0.53% over 10 years, with the initial task-based estimate no more than 0.66%.

  • Philip Trammell & Anton Korinek2023

    Economic Growth under Transformative AI

    NBER Working Paper 31815

    A survey synthesizing the AI-and-growth literature: fully automating production can break the Kaldor facts, raise the growth rate, and lower the labor share. The wider map the weak-link view sits inside.

  • Ege Erdil & Tamay Besiroglu2023

    Explosive Growth from AI Automation: A Review of the Arguments

    arXiv:2309.11690

    The case that broad automation could accelerate growth by roughly tenfold, weighed against nine counterarguments. The most rigorous steelman for the fast end of the dial.

Weak links & fragility

Measuring AI's pace

  • Kwa, West, Becker et al. (METR)2025

    Measuring AI Ability to Complete Long Tasks

    METR; arXiv:2503.14499

    The length of task an AI agent can finish with 50% reliability roughly doubled every 7 months in the original paper; METR's current dashboard estimates about 6.2 months all-time. A direct, empirical answer to 'how fast,' independent of the growth model.

  • Sevilla, Besiroglu, Cottier, You et al. (Epoch AI)2024

    Can AI Scaling Continue Through 2030?

    Epoch AI

    Training compute has grown about 4x per year; power, chips, data, and latency still leave room for runs roughly 10,000x larger by 2030. The cheap input keeps getting cheaper; that is exactly why it is not the scarce one.

  • Maslej et al. (Stanford HAI)2025

    The 2025 AI Index Report

    Stanford Institute for Human-Centered AI

    Benchmark scores jumped in a single year (GPQA +48.9 points), training compute doubles about every five months, and the inference cost of GPT-3.5-level output fell roughly 280x in two years.

  • Anthropic2025

    Anthropic Economic Index

    Anthropic (ongoing)

    Claude.ai conversations mapped onto O*NET tasks show current usage led by computer/mathematical work and education/library tasks, with explicit augmentation-vs-automation views. A live read on which links are actually moving.

Labor & distribution

  • Karabarbounis & Neiman2014

    The Global Decline of the Labor Share

    Quarterly Journal of Economics 129(1): 61–103

    Context for the capital-vs-labor split that the scenarios track to 100% / 0%.

  • Brynjolfsson, Rock & Syverson2021

    The Productivity J-Curve

    American Economic Journal: Macroeconomics 13(1): 333–372

    Why measured productivity lags transformative tech: adoption races up the S-curve while output sits in the J-curve trough.

  • Brynjolfsson, Li & Raymond2025

    Generative AI at Work

    Quarterly Journal of Economics (2025); NBER WP 31161

    A field study of 5,179 support agents: AI raised issues resolved per hour 14% on average and 34% for novices, with little effect on experts. The gains land first where skill was scarce.

  • Shakked Noy & Whitney Zhang2023

    Experimental Evidence on the Productivity Effects of Generative AI

    Science 381(6654): 187–192

    In a writing experiment, ChatGPT cut time 40% and raised quality 18%, narrowing the gap between weaker and stronger writers. Task-level speed-ups are real even where measured GDP is slow to move.

Risk & the fast-takeoff case

  • Dario Amodei2024

    Machines of Loving Grace

    Essay (darioamodei.com)

    The optimistic 'country of geniuses in a datacenter' vision, useful context for the aggressive side of the forecast debate.

Organizational change