Anton Korinek – Economic Discussion from Hard Fork Podcast

Market Reactions & the Citrini Essay

Korinek says he has been studying AI’s economic implications for about a decade and has been waiting for markets to react seriously. He finds it interesting that relatively small triggers — like the Citrini Research essay predicting a 2028 AI-driven crisis — are causing large stock market reactions.

He notes markets move emotionally, but beneath that emotion there are real technological developments.

The hosts emphasize that an essay alone triggered large drops in stock prices, underscoring how uncertain and anxious markets are about AI’s trajectory.

What the Current Economic Data Shows

Korinek says:

  • The measurable economic impact of AI so far is small.
  • Effects on employment and productivity are fractions of a percent.
  • Even those small signals are contested among researchers.
  • There is no decisive macroeconomic evidence yet.

He adds that by the time impacts become obvious to everyone, economic research will still likely debate interpretation.

Why Data Lags

Two main reasons:

  1. Statistical lag — Productivity and employment data take time to collect and revise. A full picture may not emerge until a year after changes occur.
  2. Rapid technological change — The models available today are meaningfully more capable than those a year ago, especially in coding and white-collar tasks.

AI Adoption vs. Measured Impact

A survey of 6,000 executives found:

  • ~70% of firms report using AI.
  • ~80% report no impact on employment or productivity.

Korinek interprets this as:

  • A large gap between frontier capability and everyday corporate deployment.
  • Firms are still experimenting.
  • The difference between impressive demos and reliable, scaled implementation is large.

Executives are still figuring out how to productively deploy AI in a way that maintains reliability.

“Ghost GDP” and Labor Share

The Citrini essay introduces “ghost GDP” — growth that doesn’t benefit workers.

Korinek says the concern tracks with mainstream expectations if AI reaches something like AGI.

He adds an important clarification:

  • A large portion of output could be generated without humans in the loop.
  • That means workers may not receive income from that production.
  • Additionally, some AI-generated output may not appear in GDP if categorized as intermediate goods rather than final consumption or investment.

This could mean:

  • GDP grows,
  • But labor income does not keep pace,
  • And measured GDP may even understate some AI activity.

Growth Rate Debate

There are two camps:

  1. Modest growth (~1–2% annually).
  2. Hypergrowth (10–20%+ annually, even more in extreme cases).

Korinek’s position:

  • The story isn’t written yet.
  • Triple-digit growth is conceivable in irresponsible recursive self-improvement scenarios.
  • But that would be massively disruptive.
  • 1% additional growth is likely too low.
  • Low double-digit growth could be plausible in optimistic scenarios.

However, that assumes:

  • Not just cognitive AI,
  • But physical automation (robotics),
  • Because most economic activity is not purely digital.

The Cognitive Dissonance Problem

Unemployment remains below 5%.
Productivity hasn’t surged.

Korinek explains:

  • There is a gap between frontier capabilities and diffusion.
  • That gap will shrink over time.
  • But capabilities themselves may continue accelerating.
  • Both extreme optimism and skepticism are speculative positions.
  • He personally expects continued capability growth with significant economic impact.

Substitution vs. Complementarity

In earlier research, Korinek argued that sufficiently advanced AI is more likely to substitute for labor than complement it.

Clarification:

  • This prediction applies to AGI-level systems, not early AI systems.
  • Deep neural networks scale far beyond biological constraints.
  • There is no clear natural ceiling below human intellectual capability.

If scaling continues, systems may eventually perform most cognitive tasks.

The Lump of Labor Fallacy

Economists reject the idea that a fixed number of jobs exist.

Korinek clarifies:

The key issue is not fixed jobs — it’s total demand for human labor.

If AI substitutes broadly:

  • The labor demand curve shifts downward.
  • Wages may fall.
  • Employment may fall.
  • Or both.

A less severe outcome:

  • Real wages rise.
  • But labor grows more slowly than total output.
  • Labor’s share of GDP shrinks.

Which outcome occurs depends partly on the speed of automation.

Key Indicators Korinek Watches

  1. Overall capability benchmarks.
  2. Whether models can learn dynamically (instead of frozen weights).
  3. The length of tasks AI can autonomously complete — reportedly doubling roughly every 7 months.

This “task horizon” expansion is particularly important for economic substitution.

AGI as Economic Inflection Point

Korinek believes:

  • Taking AGI seriously remains a minority view in economics.
  • If AGI is reached, it would not be the end of change, but the beginning of major transformation.
  • Many colleagues remain skeptical, but more now consider it possible.

Recursive Self-Improvement & Hyperbolic Growth

Korinek has modeled scenarios where AI:

  • Improves its own software,
  • Accelerates hardware development,
  • Unlocks new energy sources,
  • Improves robotics,
  • And feeds these improvements back into itself.

These reinforcing loops can create:

  • Super-exponential growth,
  • Hyperbolic trajectories,
  • Eventually limited by physical constraints.

This would represent historically unprecedented economic acceleration.

Advice to Graduate Students

Korinek tells students:

  • It’s uncertain whether traditional economic research jobs will exist in the same form by the time they graduate.
  • There is fundamental uncertainty about the labor market trajectory.

CEO Guidance

Public company CEOs should:

  • Stay close to frontier AI capabilities.
  • Not rely only on filtered summaries.
  • Personally experiment with advanced systems.
  • Follow capability improvement over time.
  • Begin experimentation and deployment within their organizations.

Diffusion is slow, but informed decision-making requires direct exposure.

Final Question: Is AI the Bubble or Everything Else?

Korinek’s answer:

If forced to choose, probably everything else.

However:

  • Diffusion is always slower than frontier observers expect.
  • His median view: transformative change,
  • Moderated by real-world friction and slower adoption.