Cowen and Imas: AI Breakthroughs Still Have to Pass Economic Bottlenecks

The economists see rapid AI progress, but say growth, wages and adoption depend on coordination, scarcity and institutions.

TCTyler CowenGastAIAlex ImasGast · Google DeepMind
Kanal
a16z
Veröffentlicht
16. Nov. 2026
Video
1:03:18
Lesezeit
11 Min.
Original
Englisch
Auf YouTube öffnen
Beim Scrollen schrumpft das Video in die rechte untere Ecke und läuft weiter. Untertitel folgen Ihrer Lesesprache.
LesespracheIn 2 Sprachen verfügbar
Diesen Artikel gibt es noch nicht auf Deutsch, daher wird das Original (Englisch) angezeigt.
ZUSAMMENFASSUNG

In 30 Sekunden

AI can make dramatic advances without immediately producing dramatic GDP growth, Tyler Cowen and Alex Imas argue. Faster technical progress can expose slower bottlenecks, and falling prices shift measured economic value toward what remains scarce. Imas says a major breakthrough in AI-agent coordination could change the picture, because today’s systems compete with organized human teams, not average individuals. Both see transition risks for workers while expecting meaningful benefits that ordinary GDP may not fully capture.

  1. 1Benchmark gains are not the same as economically valuable workImas says models are improving rapidly, including on less-specified tasks, but asks which valuable jobs each new model can actually unlock.
  2. 2Faster progress can make other bottlenecks more visibleCowen updated toward faster progress in mathematics but argues that institutions and other sectors still limit how quickly AI revenue and productivity can grow.
  3. 3Sustained double-digit growth requires demanding conditionsImas says models of runaway growth require innovation to avoid diminishing returns. Cowen says temporary double-digit growth is possible after a severe downturn but would bet against it.
  4. 4As automated goods get cheaper, economic weight shifts to bottlenecksImas argues that falling prices can reduce an automated sector’s measured share of output while scarce, slower-growing activities become more important.
  5. 5AI agents must coordinate to compete with human organizationsImas says a firm brings together people with different strengths. His research suggests that matching average individual capabilities is not enough for a large transformation if agents cannot coordinate effectively.
  6. 6A sudden coordination breakthrough could strain worker transitionsImas worries rapid change could leave workers too little time to find new roles. Cowen sees another transition problem: employers cannot yet easily identify who will adapt well to AI.
  7. 7Returns may flow to what remains scarce, not automatically to capitalCowen distinguishes capital doing more work from owners earning monopoly rents. He and Imas discuss energy and land as possible bottlenecks while noting that AI may change scarcity itself.

Das Wichtigste

  1. Technical capability and economy-wide adoption can move at very different speeds.
  2. Imas says coordination among AI agents could matter more to firm-level transformation than isolated benchmark scores.
  3. As automation makes a product cheaper, measured value can shift toward less-automated bottlenecks.
  4. Cowen expects meaningful AI benefits but sees sustained double-digit US growth as unlikely.
  5. Check welfare measures beyond GDP for benefits such as cheap digital services and longer lives.
  6. Imas worries a sudden transition could outpace workers’ ability to find new jobs; Cowen highlights the difficulty of identifying AI skills.
VOLLTEXT

Den bearbeiteten Text nach Kapiteln lesen

Nur das Transkript
01

Rapid capabilities meet slower institutions

Imas sees accelerating model capabilities and asks which economically valuable tasks each version unlocks. Cowen says rapid technical change can make other bottlenecks apparent.

Erik Torenberg asks whether recent evidence of rapid AI progress should change expectations. Alex Imas says capability improvements are unmistakable: releases are more frequent, and the shift from chatbots answering questions to agents doing work is showing up in real productivity.

The recent METR report did not cause Imas to update dramatically because he had already seen related evidence. He asks where the steep improvements will occur and how they connect. Work that is poorly specified or hard to verify may improve too, but could be harder to accelerate.

As an economist, Imas wants to know which economically valuable tasks each new model unlocks that its predecessor could not. A benchmark can rise sharply without an equally sharp change in real-world output, even though he believes the real economy is already accelerating.

Tyler Cowen says faster progress in one place can bring a different bottleneck into view sooner. Recording on November 11, 2026, he says his views have shifted since August, especially because of progress on mathematics, where there may be fewer human bottlenecks.

Cowen describes his most recent update as slightly more bearish on the speed of takeoff. He cites institutional uncertainty and market signals, including fluctuating expectations around a possible Anthropic IPO, that he says do not look like a straightforward forecast of the fastest scenarios.

When people propose extremely high revenues for leading AI companies, Cowen asks which other sectors will grow enough to support those payments. He remains bullish on AI overall, but trusts broadly normal market signals more than the most extreme local forecasts.

Imas agrees that buyers must eventually see enough productivity to pay for rapidly growing AI services. He also points to competition on price: agentic systems use many tokens, and he cites data in which token consumption rose sharply while revenue was comparatively flat.

Cowen expects elections and possible regulation to slow some data center development, though he does not see them stopping the sector. He says Texas’s result surprised him, while much of the broader political shift was closer to what he expected.

02

What would produce runaway economic growth?

Imas describes the demanding conditions needed for sustained double-digit growth. Both economists stress the shifting role of scarce inputs and slow-moving institutions.

Torenberg asks what would distinguish a normal-growth world from one with double-digit growth. Imas refers to economic models in which sustained explosive growth requires demanding conditions, especially the absence of bottlenecks and sufficiently strong returns to producing new ideas.

Imas explains that if one breakthrough makes further breakthroughs easier at an increasing rate, growth can accelerate rather than run into diminishing returns. He says available evidence so far is on the diminishing-returns side: producing additional intelligence still requires growing resources. That could change with a future feedback loop, he adds.

Cowen points to very high growth in small oil-producing Guyana and to temporary rebounds after recession, war or pandemic. He could imagine a year of US growth reaching 10% after a severe disruption plus an AI boost, but would bet against it and expects the rebound component to be temporary.

Robotics could bring AI into physical production, but Cowen expects deployment to take time. Driverless cars illustrate the delay: the technology works in some places, yet expanding it to more jurisdictions and markets takes longer. He does not see market valuations as evidence of 20% growth within a few years.

Imas offers a reason broad automation need not translate into runaway measured growth. As robots make a product cheaper, that product can account for a smaller share of spending. Economic weight moves toward goods and services that remain scarce and grow more slowly.

Cowen adds a difference in perspective. For economists, centuries of growth and productivity data are emotionally salient; for optimistic technologists, direct conversations about current breakthroughs may be more salient. Both sources contain information, but he gives more weight to the historical record.

Cowen expects AI adoption to be quicker than earlier technology transitions, but says government, schooling, universities and regulated healthcare show how difficult change can be. Regulation is one obstacle; organizational inertia is another, and much of an economy does not operate like a startup.

03

Measure welfare beyond market output

GDP remains useful, they say, but cheaper digital goods and longer lives can raise welfare in ways it does not fully capture.

Asked how a president should measure AI’s contribution to flourishing, Imas says the disagreement is not necessarily over model capabilities but over how those capabilities affect a complex economy. GDP has historically tracked welfare reasonably well, but technological change can widen the gap.

Digital entertainment and other products can become cheap or free while still providing value. Imas points to research using survey-based measures alongside GDP to estimate how much people benefit from such products, including a world with moderate measured growth but much greater welfare.

Cowen adds longevity. More years of life might affect total and per-person GDP in complicated ways, but their value is not identical to either measure. He expects GDP measurement to need improvement as AI advances, while still calling it a useful and underrated short-run statistic.

Torenberg asks whether advances in physical technology have answered older concerns that progress in the world of atoms lags behind software. Cowen expects a physical breakthrough eventually, but says driverless cars have reached only a small share of people and so far have made a limited contribution to welfare.

The existence of working self-driving vehicles is real progress, Cowen says, but their degree of deployment matters for judging whether a broader period of stagnation has ended. That distinction also informs his slower expectations for robotics.

04

Alignment through incentives and checks

Cowen and Imas draw on principal-agent theory and mechanism design. They emphasize checks and incentives alongside work on model alignment.

Torenberg asks about the Hugging Face incident and whether it changes their assessment of agent swarms or safety. Cowen says the result does not surprise him: principal-agent theory already describes how hard it is for a principal to control an agent through any contract.

Imas contrasts two ways of pursuing alignment. One tries to make the individual model share the preferences of the person deploying it. The other uses mechanism design: incentives and constraints can make an agent act in the principal’s interest even if its own objectives differ.

Cowen compares this to political checks and balances: people can live under an effective government without sharing the governor’s preferences. He thinks companies’ internal safety work is more practical than their public discussion can appear on social media.

Cowen also points to malicious people directing AI systems to do harm. In such cases, he says, a system may be obeying the human operator while imposing costs on others. Imas sees the incident as a chain of human errors and a chance for organizations to learn and improve their environments.

Asked about predictions of imminent AI-related deaths becoming a dominant cause of death, Cowen strongly disagrees. Neither speaker treats a single incident as resolving the broader question of safety.

05

Coordination, employment and an agent economy

Imas argues that coordinating agents, not merely improving individual scores, is key to competing with human organizations. The economists discuss labor transitions and practical adoption barriers.

Imas summarizes a theoretical paper on organizing intelligence. Individual agents are often evaluated against individual humans, but the relevant competitor in an economy is frequently an organization: people specialize and combine their comparative advantages.

His paper considers a minimal form of AGI that matches average human performance across tasks. Imas says even that would not necessarily transform output immediately, because a well-organized team can assign each task to someone particularly suited to it rather than rely on an average worker.

The paper does find that cheaper production can affect wages. For a larger economic transformation, Imas argues, agents need to coordinate with one another. He says many purported swarms mostly perform tasks in parallel, and some coordination tests find one agent can outperform two agents splitting the work.

If agents improve at coordination, Imas thinks the range of viable firm structures could change substantially. He does not claim to know the main obstacle: one possibility is simply that coordination has not been sufficiently trained into models after pretraining.

Imas worries that a sudden coordination breakthrough could change wages and employment faster than workers can move into new sectors. New jobs might still emerge, but delayed rematching could create a large pool of unemployment and lasting damage to labor-market participation. He distinguishes concern about that scenario from a claim about its probability.

Cowen’s related concern is that employers have poor signals for a new skill: who can adapt and use AI well? Traditional indicators such as grades, schools and interviews are imperfect, and matching workers to jobs could take longer until better ways to assess the skill emerge.

Despite those transition risks, Cowen expects many new projects and jobs and remains broadly optimistic about labor markets. Disruptions will occur, he says, as they did in earlier trade and industrial transitions, including some that are not yet visible.

Asked about an explosion of agents transacting with one another, Cowen says the baseline matters. He does not expect agents to account for 10% of GDP soon; even 1.5% within five years would be a very large amount of activity, though he offers that only as a forced guess.

Imas notes practical obstacles: websites are designed around human users and vary from one another, and an agent that repeatedly chooses a poor airline seat may lose its user’s trust. Platforms also have incentives to resist outside agents that bypass their attempts to upsell shoppers.

Cowen gives a personal example. He does not delegate calendar management because writing plans down helps him remember them. An agent could become much better and he might still prefer that step; such habits can be bottlenecks too.

06

Politics can slow an extreme transition

Cowen expects many data centers to proceed despite political opposition. Imas says political responses themselves will change the path of technological deployment.

On AI politics and data centers, Cowen expects some jurisdictions to say no and costs to rise, but thinks many projects will proceed across the large number of states and localities. He sees a gap between strong political rhetoric and what governments ultimately permit.

Imas says the political response depends on the economic path. When employment and growth look healthy, resistance may be limited. If an extreme economic transition began, an extreme political reaction could also begin and constrain it; later observers might see slower change because of that feedback.

Torenberg asks whether recent examples show limits to political spending. Cowen says campaign-finance researchers have long found that spending alone does not reliably buy a desired political result, apart from some local races. He treats this as another example of checks and balances.

07

How the economists found their way to the field

Imas recounts moving from medical-school plans to economics; Cowen traces his interest in AI to watching computers learn chess.

Cowen asks how Imas became an economist. Imas describes arriving in the United States as a refugee from Transnistria, growing up with limited resources, and initially choosing premed because medicine was a profession he understood.

An introductory economics class raised his interest, but he continued toward medical school. Reading Cowen’s Marginal Revolution and hearing Richard Thaler discuss economics showed him a broader set of questions the field could address. He left the medical path and attended UC San Diego.

Cowen asks whether Moldova should become part of Romania; Imas says he thinks it should and expects that it will, while noting the view may be controversial. He connects his answer to the region’s history and to similarities in language and culture.

Imas then asks why Cowen was early to AI. Cowen recalls being a young chess player when a computer called Tinkerbell played against people. Chess players understood the game required intuition, so seeing machines improve at it made progress in other intuitive domains easier to imagine.

Imas says that systems trained with verifiable feedback need not be limited forever to obviously checkable tasks: better verifiers might make more kinds of work assessable. Cowen suggests that, with some effort and input from poets, AI could write great poetry within two years; he also praises current models’ classical-music recommendations.

08

Capital, scarcity and the limits of simulation

The speakers discuss distribution, forecasts of rapid takeoff, and why simulation may not replace real experiments. They are cautious about predicting exactly which resources remain scarce.

The conversation turns to whether AI shifts economic activity away from labor and toward capital. Imas says that is a possible scenario, but another is that new jobs emerge whose value comes from their scarcity. The open question is how many such jobs exist and whether they can support the economy.

Cowen distinguishes capital becoming capable of doing much more from particular capital owners receiving durable monopoly rents. He says important technologies often benefit consumers, though that history is no guarantee for AI.

Both economists return to scarcity. Energy or land might capture returns, Cowen says, but AI could also make energy, desalination or land transformation cheaper. Imas agrees that the location of scarcity matters, while the exact answer is hard to predict.

Torenberg describes a contrary view from AI power users who see demand exceeding supply and mentions a new infrastructure fund. He acknowledges that this rests on an assumption: wider use must eventually deliver enough business value to justify much higher spending.

Cowen praises Leopold Aschenbrenner’s “Situational Awareness” as an important public work but disagrees with its fast-takeoff assumptions. In particular, he doubts simulation can generally substitute for physical experiments, though it can help in some areas.

Imas adds that predictions must account for how a system responds to a change. He invokes the Lucas critique: relationships estimated under one set of conditions may not remain valid when policy or behavior changes.

Cowen calls Robin Hanson’s The Age of Em a stimulating book but says he has long expected AI to arrive before whole-brain emulations, if those emulations arrive at all. The book remains useful to him as a source of ideas.

Asked about recent uncertainty involving Nvidia, the speakers again emphasize institutional checks. Imas expects supply chains and structural forces to prove resilient; Cowen says particular financing roles could be weakened in a downturn even if medium-term effects fade.

Cowen is optimistic about biotechnology but expects slow progress because of regulation and the need for experiments. Imas is excited by the prospect of better structural models, including digital twins, that help researchers understand mechanisms rather than only predict outcomes.

09

Experimenting in higher education and economics

Both welcome more experimentation in higher education. They expect economic reasoning to remain useful for understanding institutions and AI systems.

On higher education, Cowen calls for more experimentation, measurement and sustained investment, and welcomes new institutions trying different models. Imas expects traditional higher education to contract in some ways but is optimistic about long-term innovation, while acknowledging that institutions move slowly.

Imas argues that economic tools remain valuable for AI questions, from incentives and alignment to political feedback and the limits of prediction. Cowen imagines economists using AI to fill knowledge gaps while applying a common toolkit across social science, managing agents to investigate a wider range of questions.

Mehr aus diesem Kanal

3 Folien16:19EN
a16z· 17. Nov. 2026AI Research Interns Are Here; Autonomous AI Researchers Are NotSarah Wang, Mark Chen, John Schulman · Andreessen HorowitzMark Chen and John Schulman explain how AI already speeds up research, and why judgment, measurement and compute still constrain it.AutoresearchAI researchAI Alignment+1
46:06EN
a16z· 13. Nov. 2026AI attacks are moving faster than companies can patchAmi Luttwak, Zane Lackey, Sarah Wang · WizWiz’s Ami Luttwak argues for continuous testing of exposed systems and security designed to respond in minutes.AI agentsInfrastruktur & DeploymentRed Teaming
46:11EN
a16z· 12. Nov. 2026AI can shorten game prototyping, but making a game fun still takes judgmentAnish Acharya, Jonathan Lai · Andreessen HorowitzAnish Acharya and Jonathan Lai discuss AI-assisted game tools, modding and the possibility of more adaptive worlds.Interactive EntertainmentGenerative mediaAI agents
56:06EN
a16z· 11. Nov. 2026Mike Volpi: AI infrastructure can endure a market correctionRaghu Raghuram, Guido Appenzeller, Mike Volpi · Andreessen HorowitzThe former Cisco executive sees useful assets beyond the hype, but questions who will own them and how systems will evolve.Infrastruktur & DeploymentSemiconductorsVenture Capital
59:06EN
a16z· 10. Nov. 2026Dharmesh Shah: Find startup opportunities beyond the AI labs’ pathJoe Schmidt, Dharmesh Shah · Andreessen HorowitzHubSpot’s co-founder explains how to test narrow ideas, build around customer work and use AI without relying on a temporary model gap.AI startups & small teamsProduct StrategyEnterprise-KI+2
48:39EN
a16z· 6. Nov. 2026John Somorjai on Salesforce’s Acquisition Strategy in the AI EraAlex Immerman, John Somorjai · a16zSalesforce’s acquisition leader explains build-versus-buy decisions, AI pricing, Slack and the diligence issues that can stop a deal.Acquisition StrategyEnterprise-KIVenture Capital+2