Here is the paper and its abstract (and here the slides):

Artificial Intelligence and the Brave New World in Finance

Economics assumes that even when people know different things, they have a common understanding and hence can describe the world to one another in a shared language. Trust in the understanding of experts, extended by institutions, gives each person access to a broader societal understanding. Society hence understands more than any of its members. The introduction of agentic AI is qualitatively different from previous innovations. It disrupts the trust arrangement, undermines societal understanding and leads to asymmetric understanding: AI agents can learn how humans think and respond, while humans may be unable to understand or reliably anticipate how those agents will act. This asymmetric understanding can make prices harder to read (less informationally efficient) and put central banks at a strategic disadvantage when engaging and communicating with market participants. Preparing for the asymmetric understanding scenario calls for segmented markets that preserve a human fallback, simpler and more robust central bank rules, and less strategic ambiguity.

Raghu Rajan discussed the paper (here are his slides). His verdict: "Markus does a wonderful job of raising concerns. But policy changes require more experience and debate." I hope the paper gives others something to build on, or to push against.

Thank you,

Markus

Subscribe now

Episode 1: AI and Theory as a Fixed Point

In the first video Pietro outlined where AI fits in theory research. In contrast with math, our models don’t seek to prove a conjecture, but rather try to illustrate certain mechanisms. As a result, writing theory papers is analogous to searching for a fixed point; iterating over defensible assumptions and proofs to arrive at the desired insights. AI speeds up the iterations to arrive at the fixed point. The takeaways:

• Pietro outlined 7 use cases: (1) sketching and brainstorming, (2) literature reviews (3) suggesting proofs, (4) checking proofs, (5) extensions, microfoundations, and simplification, (6) general proofreading, (7) simulations

• Sketching models is the most underused. Give the model a vague intuition and ask: “Give me three minimal models that capture this intuition. For each, say what is elegant, what is fragile, and what theorem would be worth proving.”

• Despite often being notation-heavy, frontier models can prove the results of an economist’s typical model reliably. Reliable does not mean that they deliver what you want to (or should) submit. To be effective, weaker models need to have the tasks decomposed into parts

• The higher-value uses are attack (a hostile referee), repair (which assumption rescues a false statement) and inspiration: even a wrong proof can have insights

• The risk of wasting time with AI-driven rabbit holes scales with your own ignorance

Episode available here: https://www.youtube.com/watch?v=kVVggnIaw2Q

Timestamps:
[0:00] Introduction, and economic theory as a fixed point
[5:12] Seven use cases, and why sketching is the most underused
[8:46] Proofs: attack, repair, and inspiration
[15:12] Extensions, microfoundations, simplifications

Episode 2: Can AI Be Creative?

In the second part Fedor asked whether models can actually generate new ideas. In econ this is hard to assess, as model quality is subjective. We should look at math, where a proof either holds or it doesn’t, for data points. The takeaways:

• Poor at attribution, fantastic at aggregation. The October 2025 claim that GPT-5 had cracked ten open Erdős problems did not hold up: Bloom (2025) showed it had surfaced known solutions. The flip side: no lemma buried in an unread appendix is ever lost.

• AI excels at obtaining counter-examples: for example OpenAI (2026) disproving the unit distance conjecture and Fable 5 disproving the Jacobian conjecture

• Multi-agent workflows with frontier models can explore many proof strategies in parallel. Sol Ultra proved the cycle double cover conjecture using 64 agents

Episode available here: https://www.youtube.com/watch?v=LbedQKu-5XM

Timestamps:
[0:00] Can models be creative, and why the evidence comes from math
[2:20] The ten Erdős problems episode, and Terence Tao's ledger
[5:19] The unit distance conjecture falls
[8:44] Jacobian, cycle double cover, and what it means for theory

Episode 3: Which Model to Use

Aware of how fast the frontier is moving, Pietro compared the models available today. The takeaways:

• Intelligence and stamina are substitutes. AI often gets further by running a weaker model for a long time than a smarter one briefly. Yet many of the deeper models are not built to work for 20 consecutive hours, while agentic environments are. Stamina is also about price considerations; it may be prohibitively expensive to run Fable for extended periods.

• Their verdict for the best model for theory today: GPT-5.6, narrowly. Fable may be more intelligent, but it is outweighed by 5.6’s stamina

• The best mix is combining GPT-5.6 Pro on the browser with Sol Ultra. For a $20 subscription, pick OpenAI’s. The best free option is Gemini Pro in Google AI Studio, though your chats train Google’s models. Set maximum effort when available to you

• Agents are fundamental in empirical work, but the browser is still enough for brainstorming

Episode available here: https://www.youtube.com/watch?v=lLgD9WwFMvM

Timestamps:
[0:00] The frontier since June: Fable 5, GPT-5.6, Opus 5
[4:13] Intelligence and stamina are substitutes
[8:53] The verdict, and what to run on a Pro, $20, or free budget
[11:10] Browser or agents?

Episode 4: Prompting, Agent Adversaries, and Swarms

In the fourth video Fedor covered the current state of “prompt engineering”. Writing effective prompts is no longer as valuable as it was a few years ago. The takeaways:

• The value has shifted to context engineering. Specify task, context, output, quality criteria and incentives, one task per prompt; the weaker the model, the more it matters.

• Let the model write the prompt based on your lazy two-liner. Have the same model that will do the work expand the prompt. Spend 15 minutes reviewing the expanded prompt.

• Never let a session grade its own work. Run a verifier, and even a third “judge” between the “prover” and the “verifier”.

• Provide LaTeX files whenever possible, and always convert PDFs to Markdown before providing them. Restart sessions often

• The published prompt behind the cycle double cover proof provides a template for how to implement swarms of agents: 64 agents working for eight hours minimum with a supervisor stopping those that converge.

• Fedor adds a trick: give the agents an escape hatch. If they get stuck for an hour, they can write their blocking obstacle to a Markdown file and hand it up to a better model. Running Sol Ultra this way for 15 hours cracked a conjecture he and his coauthors had worked on for a year, which neither 5.6 Pro nor Fable had solved.

Episode available here: https://www.youtube.com/watch?v=OBJ0h8j_m8Y

Timestamps:
[0:00] Is prompt engineering still a thing?
[3:40] Prompt expansion: let the model write the prompt
[9:45] Prover versus verifier, LaTeX not PDF, and when to restart
[16:43] Agent swarms, and AI proofreading

Subscribe now

* Hosted by Markus Brunnermeier, with the support of Pablo Balsinde (PhD student, Stockholm School of Economics).

Recap of Part 1

The Financial Times’ Martin Wolf wrote an insightful piece on Part 1 (available here).

Wolf highlighted some of Setser’s central claims: that China’s exports are surging while domestic demand has stalled, that its manufacturing surplus is now around 2% of world GDP (roughly twice the largest Japan ever ran), and that the unexplained $125bn deficit on China’ investment income account hides a much larger true surplus, implying an even more severe undervaluation of the renminbi of around 30%.

Wolf also closed with Markus’ proposal to measure the “resilience account” in Part 1’s introduction. The current account adds a country’s net exports with its net investment income and its transfers: a deficit means a country is borrowing from the rest of the world. Acknowledging that each good/service is a bundle of (i) the good itself and (ii) a build-up of geopolitical dependency, the resilience account would also net out geopolitical dependencies. How easily could each good be sourced from somewhere else?

Part 2 Highlights

Watch Part 2 and read the summary below. Setser’s slides are available here. A summary in three bullets:

• Germany is hit hardest: with its industrial production down ~15%, its export mix overlaps with China’s the most

• Autos as the key example: In 2021 China used to import and export 1 million cars. Today it imports 400k and exports 10 million

• Without policy change the imbalance will not self-correct, while only China can correct its exchange rate. Setser argues Europe must take the initiative, including through tariffs, to achieve currency adjustment

[0:00] Europe sandwiched, Germany in the crosshairs

• German industrial production is down ~15% over the last 7 years, and 20–25% in the sectors most exposed to China. China’s manufacturing surplus has risen ~$1tn, concentrated in machinery, autos and transport equipment, the core of Europe’s industrial base

• Export-similarity indices (Finger & Kreinin 1979) show China’s export mix converging on Germany’s and Italy’s far more than on the US’s (de Soyres et al. 2025); German exports to China are down ~1pp of GDP

• Net exports are a major drag on German growth; without them the economy would be expanding rather than flat

[17:52] China’s imbalance will not fix itself

• Do not trust the IMF when it says China’s current account surplus will decline. For years they have forecasted a surplus decline that never arrives

• Compared to other high savings countries (e.g. Singapore, Taiwan, Norway) China’s surplus is small, so it is conceivable that its surplus could grow further.

[22:20] How Europe should act

• Policies targeting the financial account like blocking Chinese investment (Klein & Pettis 2020) are hard for the US due to its deficit-financing needs. Europe, which saves more than it invests, has a freer hand

• EU leaders have told Xi that the imbalances are unsustainable but continue importing. At some point they will have to back up their threats

• The likelier lever is strategic tariffs: lift them if China lets the currency appreciate — with economic coercion (rare earths, magnets) as the tail risk.

• The US made a mistake by pivoting from China-targeted measures to across-the-board tariffs

• We should build North-Atlantic common markets and pursue joint industrial policy in sectors like autos and pharma. The WTO’s non-discrimination principle no longer fits an economy of China’s size and strategy; new rules would start as US–EU bargains and then generalise to the rest of the world

[39:00] Implications for the rest of Asia

• The world is facing three major shocks: the Second China shock, the AI shock, and the Trump tariff shock.

• The shocks affect Asian countries differently. Vietnam is the clear winner of the tariff shock as manufacturing assembly relocates there (large US bilateral deficit), offsetting the China shock. Korea is bifurcated, with the AI memory-chip boom offsetting the auto sector squeezed by China

• The common thread is that broadly weak exchange rates keep concentrating manufacturing in East Asia

Subscribe now

* Setser is a Senior Fellow at the Council on Foreign Relations and a former official at the US Treasury and the Office of the US Trade Representative.

** Hosted by Markus Brunnermeier, with the support of Pablo Balsinde (PhD Student, Stockholm School of Economics).

Watch Part 1 of the talk and read its summary below. We will share the second part with implications for Europe tomorrow. Setser’s slides are available here.

A summary in three bullets:

• Concentration and bifurcation: compared to the imbalances of the 2000s or 2010s, today surpluses are concentrated in East Asia, China above all. China’s economy is now bifurcated: exports surge while domestic demand stalls, and due to import substitution policies domestic demand has decoupled from imports

• The financial flows have reversed: the pre-GFC’s safe-asset scarcity has ended. Reserve accumulation has stalled and surpluses are no longer recycled into Treasuries

• The mismeasurement of China’s investment income account understates its true surplus, lifting the yuan’s implied undervaluation to ~30%, versus the ~19% implied by the officially-reported surplus (the IMF’s own estimate of undervaluation is smaller still)

Highlights of Part 1

[00:00] Markus’ Introduction

• The world order has shifted from a rules-based multilateral system, with a balance of choke points and mutual interdependence, to a transactional and bilateral system where size and spheres of influence dominate and resilience outranks efficiency

• Each good/service should be seen as a bundle of (i) the good itself and (ii) build-up of geopolitical dependency. Alongside the current account, we should begin measuring the “resilience account” which would net out geopolitical dependency: how easily the goods in the current account can be sourced from other countries

• Comparative advantage is dynamic, built by industrial policy in sectors with increasing returns and learning-by-doing, as in the infant-industry / strategic-trade tradition of Krugman (1987) and Grossman & Helpman (1990). This should impact tariff and exchange rate policy

[05:17] China’s unprecedented surpluses

• China’s manufacturing surplus is ~2% of world GDP. This is roughly twice the largest surplus Japan ever ran (the target of the 1985 Plaza Accord) and twice Asia’s pre-GFC peak.

• Tooze (2026) has dubbed this “mercantilist-on-mercantilist violence”, as China’s gains come partly at other surplus economies’ expense (e.g. Europe)

[21:00] An export boom bolted to a stalled economy

• The real-estate bust cut property investment from ~12% to ~6% of GDP; capital has shifted into manufacturing and “new productive forces” aimed at cutting reliance on US and European choke points

• China’s share of global investment has started to decline while its share of trade has continued growing. The slow-down in investment has weighed on consumption, with the actual domestic demand growth likely half as large as the officially-reported (3%)

• Since 2021, China’s import growth has decoupled from demand growth

[35:11] This time the surpluses are not recycled

• Pre-GFC, the surge in reserve accumulation and a safe-asset shortage fed the housing-securitization machine. Today reserves are flat, while with its deficits the US has been oversupplying treasuries

• With the end of low-for-long rates in the US and the appreciation of the dollar, countries have seen less of a need to prevent their currencies from going up (and in the process accumulating reserves)

• Instead of accumulating reserves, the big reserve holders have been funding quasi-private channels, such as China’s state banks or Korea’s national pension service, that chase return, not just safety

• This shift has led to a more opaque recycling of surpluses and more treasuries being held by private investors (the basis trade). Even if the incoming treasury investors are more levered, this may prove more stable than the old reliance on housing securitizations: housing bubbles burst, whereas Treasuries always converge to par

[50:11] China’s statistical puzzles and the undervalued yuan

• China reports an unexplained investment-income deficit (~$125bn). However, under standard assumptions of the return of its net foreign-asset position it should show a ~$100bn surplus. The likeliest culprits are under-counted income from state banks and offshore FDI vehicles

• This understates China’s true surplus. Setser puts the real surplus near 5.5% of GDP versus a reported ~4%

• This mismeasurement drives the currency debate. Measured against the IMF (2025)‘s ~1%-of-GDP norm, the official surplus implies a ~19% undervaluation; while with Setser’s corrected figure it rises to ~30%

Subscribe now

* Hosted by Markus Brunnermeier, with the support of Pablo Balsinde (PhD Student, Stockholm School of Economics).

Messy Jobs is now out on Amazon (if you’ve read it, a one-minute Amazon rating is the best way to help). You can also check out Luis’ Substack on Europe, Technology, and AI: .

Watch the full talk and read the summary below.

Highlights

[02:43] Tasks are not jobs: the autonomy threshold

• Exposure to AI may be highest for workers with the highest wages, but task exposure does not entail job displacement

• The autonomy threshold is the level of ability at which AI can complete tasks alone and replace humans. Below the threshold AI complements humans (e.g. the difference between cruise control and autonomous driving)

• The threshold increases with messiness, and messiness increases as tasks become bundled and jobs involve relationships

[12:48] Three tiers of jobs, three kinds of AI impact

• Tier 1: jobs based on clean, verifiable tasks. The jobs Silicon Valley warns us of. AI crosses the threshold fast and the price for the task collapses, unless demand or regulation or a humanness premium protects the job

• Tier 2: jobs with strong bundles. Bundles of tasks can be held together by the need to synchronize actions, knowledge spillovers, or the joint measurement of outcomes. Despite Geoffrey Hinton (2016)’s warning that radiologists would be displaced, radiology is among the highest-paid and record-hiring specialties. The job involves a bundle of tasks like talking to patients, consulting surgeons, and deciding over hard cases

• Garicano et al. (2026) build on Coase (1937)’s theory of the firm and Becker and Murphy (1992)’s study of the division of labor across workers to formalize AI and the bundling of tasks. The real threat to workers is unbundling, not automation. If coordination costs decline the AI will do more tasks and only shells of jobs remain, increasing their supply and collapsing worker rents

• Tier 3: jobs that rest on authority, human initiative, and relationships

[30:06] Hierarchies and which jobs will survive

• We lack good research on which bundles will survive, but it is likely that they will combine high cognitive and social skills (Deming 2017)

• AI will rewire hierarchies in two ways (Ide & Talamàs 2025, see our webinar):

  • AI below: one expert leverages many AIs (the one-person company)

  • AI above: expertise-in-a-box lifts less-trained workers who retain decision making, compressing inequality (Brynjolfsson, Li & Raymond 2025).

• The middle of the wage distribution will then split in two: the routine, cognitive part will be squeezed out, while the “AI above” part might rebuild the middle class because emotional and physical skills are spread more evenly across the population (Autor, 2024)

[46:31] The value of organization

• Organizations are coalitions of individuals with conflicting goals. Tier 3 jobs will be based around organizational authority and decision making

• Why AI cannot hold authority:

  • (1) Tacit relational knowledge. In many settings, unless there is trust people will protect their private information and refuse to share what they know.

  • (2) Accountability and liability: who pays when an agent breaches a contract or harms a third party?

  • (3) Procedural fairness: when people cannot judge the outcome, they judge the process

• The key near-term scarcity is organization-level AI implementation. Those who jointly grasp the workflows, the tools and the stakeholders will benefit

[1:00:57] Three tiers, three economies

• The algorithmic economy (Tier 1), where cognition costs almost nothing and the value is captured by whoever owns the platform, stays small (a few percent of jobs).

• Relational work (Tier 3) of authority, implementation, and authenticity retains a double digit share of work.

• Most careers sit in the messy middle (Tier 2), where a job survives because its cognitive part is bundled with social, physical and relational tasks.

• As cognition becomes abundant its price falls toward zero and it shrinks as a share of the economy. Productivity still rises: Tier 1 automates and Tier 2 is augmented, but Tier 3 bottlenecks will cap the gains. Societies that are better able to restructure their organizations (and politics) around AI will see faster TFP growth

Subscribe now

* Hosted by Markus Brunnermeier. Summary produced by Pablo Balsinde (PhD Student, Stockholm School of Economics).

As we publish this on July 4th, we also wish the United States a happy 250th birthday, and express our admiration for the Founding Fathers: visionary philosophers who became practical nation-builders.

Watch the full talk below. A summary in three bullets:

• Reported returns overstate what investors actually earn: IRRs are measured from the capital call, not the commitment, and subscription lines, NAV loans, and unfunded commitments inflate the headline number while quietly adding leverage

• Private markets diversify less than advertised and cost more than they seem: a single fund is a concentrated bet on a manager, not an “asset class”. Opaque and layered fees are paid for longer than investors expect

• Risks may be neglected due to agency conflicts or beliefs, for example obscured by an insurance–private-credit nexus leaning on biased ratings

Highlights

[05:24] Returns aren’t what they used to be

• We should be concerned about “democratizing” access to private assets when even ultra-high-net-worth-individuals routinely fall victim to private-market traps

• Returns have been falling as the illiquidity premium erodes: private-credit yields have compressed from roughly 10% to 6%. As capital has flooded in private market funds prices increased, implying lower expected returns going forward

[15:25] Reported vs. realized returns

• Advertised IRRs are not what investors earn: they are computed from the moment that fund managers (General Partners - GPs) call for the pre-committed capital from outside investors (Limited Partners - LPs), not from when the capital is committed

• Holding the committed-but-uncalled capital in cash can impose a significant drag on returns. LPs effectively write the General Partners an unpriced option and bear a cash drag that is hard to value (Gourier et al., 2024)

• Three devices widen the gap between realized returns and IRRs:

• 1. Subscription lines of credit: Early in a fund’s life the GP borrows from a bank to make an investment instead of calling LP capital, then calls that capital a year or two later. The deal’s return is unchanged, but it is now earned over a shorter time, so the measured IRR rises (ILPA, 2017)

• 2. Net-Asset-Value loans: Late in a fund’s life the GP borrows against the accounting value of the portfolio to fund redemptions or new purchases. This increases leverage, risk, and the IRR without touching the underlying value

• 3. Unfunded commitments: Investors pledge capital to future funds they cannot yet fund, so even a family office whose policy statement forbids leverage can run hidden leverage off balance sheet, understating portfolio risk (Jansen et al., 2024).

[26:47] Diversification myths and concentration risk

• The idea that “private equity diversifies your portfolio” is often asserted without evidence: there are no monthly returns to compute correlations with other assets, and any diversification or risk-adjusted performance can only be observed ex-post (Schmalz and Zhuk, 2019; Franzoni and Schmalz, 2017)

• In practice portfolios are concentrated: heavy in SaaS and AI infrastructure, and not marked-down when public comparables fall

• Buying one fund is a concentrated bet on a manager, not exposure to an “asset class.” That matters because returns are right-skewed: in public markets Bessembinder (2018) finds ~4% of firms account for all net wealth creation above T-bills, so investors must “buy the haystack”. Most family offices are too small to write enough tickets to do that, so they take uncompensated idiosyncratic risk.

[37:00] Opaque fee structures

• Fees are opaque, with account, management, placement, and carry fees being stacked in a “Russian-doll style”

• Hortaçsu and Syverson (2004) famously showed that there was a large fee dispersion in identical S&P 500 index funds, and attributed this to investors’ imperfect information

• The same is happening with private assets. Goldman charged 1.25% management fee plus 17.5% carry for access to investing in Anthropic while Morgan Stanley offered the same round at a one-off 1% management fee (FT, 2026)

• As distributions stall and holding periods lengthen, investors are also not aware of how long they will pay fees (Strebulaev, 2026)

• Private “debt” is more equity-like than it looks: roughly a fifth of these portfolios is preferred stock, equity, or warrants whose payoff tracks the borrower’s share price. After controlling for the embedded equity risk private credit’s alpha disappears (Erel, Flanagan and Weisbach, 2024)

[46:45] Liquidity, systemic, and agency risks

• Fund liquidity is an illusion once redemptions are correlated: a 5%-a-year redemption right feels ample to any one investor, but when everyone runs at once funds gate withdrawals as illiquid loans can’t be conjured liquid on demand

• Insurers now hold roughly a fifth of their fixed income assets in private credit and are increasingly cross-owned by the fund sponsors. They are leaning on overoptimistic private ratings to reduce their capital requirements (Li et al., 2026)

• Genuinely good advice (personalized, independent, and competent) doesn’t scale, so it barely exists. We lack research on optimal long term investment strategies (Cochrane 2022 is an un-personalized first step) so it is hard to evaluate advisors

• Illiquidity, although it prevents accurately measuring diversification, may be a commitment device. People are more averse to imminent risks (Eisenbach and Schmalz, 2016) so in liquid but crashing markets they panic-sell; a lockup ties their hands. On one reading investors rationally pay for that discipline; on another they simply underestimate the risk (Eisenbach and Schmalz, 2018)

Subscribe now

* Hosted by Markus Brunnermeier. Summary produced by Pablo Balsinde (PhD Student, Stockholm School of Economics).

This post includes the eight episodes of the series. Paul’s detailed notes for each episode are also linked below.

Episode 1: Getting Started

Paul introduced Claude Code as a terminal-based AI coding assistant that can read files, write and run code locally, and accelerate research workflows. He explained his key principles for using these tools optimally, for example the importance of the context window and compaction.

He contrasted Claude Code with the more sandboxed Cowork environment, and discussed other complementary tools like Ghostty, Zellij, and Oh My Zsh.

Paul’s detailed notes on this episode can be found here.

Timestamps:
[1:43] Series Overview
[6:52] What is Claude Code / Cowork?
[22:10] Installation and pricing
[23:54] The context window
[29:31] Privacy
[31:18] Tips for getting started

Episode 2: Data Analysis

Paul showed how Claude Code dramatically shrinks the gap between a vague research idea and initial results. As an example, he used Claude Code to obtain and plot data on home ownership in the US. Claude Code found the relevant data from the Census, handled scraping issues, cleaned spreadsheets, generated the required scripts.

Watch the full talk below:

https://www.youtube.com/watch?v=Rp17XUPxa4I

Paul’s detailed notes on this episode can be found here.

Timestamps:
[1:33] Starting from nothing
[6:16] Agent spawning
[13:20] Scraping tips
[22:24] Plotting graphs
[30:04] Results
[35:20] Summing up

Episode 3: Web scraping

Paul used Claude Code to scrape SEC EDGAR filings. Claude Code extracted the risk-factors mentioned in firms’ 10-K reports and built a structured DuckDB database to show how tariff-related risk disclosure became more frequent and more specific in recent years.

Overall, this episode highlighted how to best work with Claude Code by planning and iterating on results. Paul turned messy filings into usable research data in just a few minutes.

Watch the full talk below:

https://www.youtube.com/watch?v=wqLZrKdevHs

Paul’s detailed notes on this episode can be found here.

Timestamps:
[1:26] Developing a plan to scrape 10K filings
[12:40] Modifying the plan
[18:05] Running the scraper and debugging
[24:03] Which sectors mention tariffs the most?
[26:32] Recap

Episode 4: Large datasets

Paul showed how Claude Code can be used to work with large datasets, converting massive CSVs into Parquet, and organizing everything in DuckDB. He also highlighted how Claude Code handles more complex projects: using planning mode, spawning sub-agents, debugging data issues, and managing context.

With Claude Code he built a mortgage market panel from the Home Mortgage Disclosure Act data to study county variation in lender concentration. He extended the pipeline to classify lenders and illustrate the growing share of fintech and non-bank mortgage lending over time.

Watch the full talk below:

https://www.youtube.com/watch?v=4uwI1-9DafU

Paul’s detailed notes on this episode can be found here.

Timestamps:
[00:00] Intro
[02:59] Building a mortgage panel from the Home Mortgage Disclosure Act Data
[07:52] DuckDB + Parquet for large-scale data work
[11:47] Claude Code’s planning mode and agent workflow
[27:55] Harmonizing 18 years of mortgage data
[42:01] Fintech lender classification and market-share trends

Episode 5: Writing and Thinking

In three different parts, Paul showed how Claude can support the writing side of research:

1. First, he showed how to create a personal style guide from your own papers.

2. Second, he showed how Skills can be used to develop a revision plan after receiving a referee report. He especially recommended Jukka Sihvonen’s Skill for Strategic Revisions.

3. Third, he shared some advice on how to use LLMs as editors and thinking partners rather than judgment outsourcers.

Watch the full talk below:

https://www.youtube.com/watch?v=BxfSiB3Moyo&t=2579s

Paul’s detailed notes on this episode can be found here.

Timestamps:

[1:27] Developing a Style Guide
[18:12] Using Skills for Referee Reports
[35:37] Claude for Brainstorming and Editing
[43:42] Summary

Episode 6: Claude Skills

In this episode, Paul explained Claude Skills: reusable instruction bundles that help Claude perform recurring tasks in a standardized way. He showed where Skills are stored, how global and project-level Skills differ, and then built a paper-summary Skill. Paul also introduced community Skill packs such as “superpowers.”

Watch the full talk below:

https://www.youtube.com/watch?v=a03ehomPqMA&t=162s

Paul’s detailed notes on this episode can be found here.

Timestamps:

[0:00] What are Claude Skills?
[7:33] Building a paper-summary Skill
[21:50] Skill packs, “superpowers,” and cautions about overusing Skills

Episode 7: Permissions and OpenClaw

In this episode, Paul explained how to think about autonomy and risk when using Claude Code and other agentic AI tools, for example covering different types of permissions. Paul showed how containers and sandboxes, including Docker-based workflows and tools like Safe House, can give AI agents more autonomy while limiting potential damage. Paul also discussed OpenClaw-style bots, and introduced us to his own “Duncan Idaho” research assistant agent.

Watch the full talk below:

https://www.youtube.com/watch?v=8Jnx5rL_Gfk&t=2143s

Paul’s detailed notes on this episode can be found here.

Timestamps:
[0:00] Permission types and an overview of agentic tools
[11:37] Sandboxing with containers and Docker
[35:43] OpenClaw and autonomous research agents

Episode 8: Integration and Collaboration

Through a replication of Jay Ritter’s results on post-IPO stock performance, Paul shared some best practices for verifying Claude Code’s work. Integrating with GitHub let him track the project as a chain of committed steps rather than one final output, and review Claude’s work one change at a time.

Finally, Paul connected the project’s repository to Overleaf via GitHub (Dropbox works too) so that Claude can update figures and tables directly in one’s paper. The core principle: every number in a draft should come from the code, not from an AI’s memory.

Watch the full talk below:

https://www.youtube.com/watch?v=EcloxLPcRsY&t=3315s

Detailed notes from Paul are available here.

Timestamps:

[0:00] Verification is the new bottleneck

[6:30] Using GitHub to verify Claude’s output

[38:35] How to give feedback

[52:27] Integration with Overleaf for Collaboration

Subscribe now

* Hosted by Markus Brunnermeier, with the support of Pablo Balsinde (PhD student, Stockholm School of Economics).

Watch the full talk below. A summary in three bullets:

• The new wave of AI development is driving a new wave of physical capital formation that is both unusual in scale and in composition

• The buildout is changing who owns and finances AI infrastructure. Hyperscalers are moving away from fully self-funding data centers and are increasingly relying on landlords, debt, SPVs, private credit, and securitized structures

• This new architecture differs from the old model: leverage is rising, obligations are moving off balance sheet, and risk may be ending up with investors who are harder to identify and monitor

Highlights

[04:32] Why AI changed the Economics of Data Centers

• Modern data centers are much more than a warehouse with servers in it. They require an order of magnitude more power, backup generators, bespoke liquid cooling, fiber-optic cables…

• A frontier data center with 200 MW power capacity costs $8.2bn. The planned U.S. data center capacity of 200 GW implies $8.2tn in CAPEX over roughly a decade

• AI-related investment today essentially accounts for all U.S. GDP growth. The planned investment exceeds historical infrastructure booms in railroads, electrification, highways, and telecom fiber

• Half of the investment will be for computing equipment, another third for data facility infrastructure, and the rest for new power capacity. In data centers the tenant owns the IT equipment, while the landlord provides the building and energy infrastructure

• In the past data centers were owned by publicly listed real estate investment trusts, and each data center would house many tenants. Hyperscalers (e.g. Meta, Google), with their complex hardware needs, have transformed the market into a single-tenant market with large bespoke AI campuses

[22:25] The changing ownership and financing of the AI buildout

• Hyperscalers historically self-financed and owned data centers, but escalating AI CAPEX and limited free cash flow increasingly push them toward leasing real estate and, increasingly, IT hardware

• Data center leases are like corporate bonds. Landlords prioritize having creditworthy tenants, so hyperscalers are a natural fit. AI model firms (e.g. OpenAI) are startups, and so they rent hyperscalers’ capacity via compute contracts

• 60% of the financing will be equity and 40% debt. Most of the hyperscaler equity is for IT needs. Most of the debt is for data centers, and is highly leveraged at ~70%

• In the next 4 years Morgan Stanley expects 1.15tn in private debt, bonds and securitizations. For a sense of scale, in 2007 there were 2tn in subprime mortgage securitizations

• Corporate bond issuance by hyperscalers has surged, but the dominant incremental funding channel for new data centers is long-maturity, highly levered structured finance rather than traditional on-balance-sheet debt

• Securitizations come in two forms: (1) Commercial mortgage-backed securities borrow by collateralizing the buildings (which are in turn backed by the lease payments); they have very little diversification, often with a single tenant and a single borrower. (2) Asset-backed securities borrow against collateral like GPUs

[38:20] The New Finance Architecture of the AI Buildout Transfers Risk and Increases Opacity

• A new structure is emerging, illustrated by the Meta–Blue Owl Louisiana campus. Originally owned entirely by Meta, it is the largest facility to date (2 GW)

• Meta then sold an 80% equity stake to Blue Owl for $2.5bn, and as a joint venture SPV they issued $27.3bn in debt. With ~90% leverage, it is the largest investment grade bond issuance in U.S. history.

• This long-maturity amortizing bond is bankruptcy remote from Meta’s balance sheet. Back of the envelope: Meta would have paid 120bps less if it had financed it with unsecured corporate debt on its own balance sheet. The bonds are trading above par and for technical reasons it has much lighter reporting requirements to the SEC

• Instead of a 20-year lease, Meta signed five 4-year leases. Meta has the right not to renew them, but if this happens they have to pay the landlord a minimum residual value. Even if one of the two outcomes will happen with certainty, accounting rules allow Meta to record neither the future leases nor the residual guarantee as a liability

• With the lower reported liabilities, Hyperscalers preserve software-like equity multiples by shifting physical capital and leverage onto SPVs. Hyperscalers do not want to trade at the multiples of infrastructure companies

[50:26] What could go wrong?

• The bullish case is that vacancy rates are at all-time-lows, rent growth is at-all-time highs, capital is being deployed by the companies with the strongest balance sheets on earth, while there is no speculative development in the sense that nothing is being built without knowing who will occupy it

• However hyperscalers face rising costs of capital and increasing leverage, while revenue growth is uncertain and there is a lot of concentration risk. Credit default swaps on hyperscalers have climbed in recent months

• The second risk is technological disruption: quantum computing, much more efficient chips, or inference shifting to phones could rapidly depreciate data center collateral

• We know little about who is bearing risks. If they shift to pension funds and insurers, households will be on the line, while wealthier households will bear the costs if the risks shift to private credit

Subscribe now

* Hosted by Markus Brunnermeier. Summary produced by Pablo Balsinde (PhD Student, Stockholm School of Economics).

• Smith’s conception of human nature combines a propensity to truck, barter, and exchange with the social desire to be admired and praiseworthy

• Admiration, status-seeking, and faction lead individuals to overvalue wealth and social rank relative to virtue and intelligence, distorting our moral sentiments

• The invisible hand requires a framework of justice, competition, and well-governed institutions; mercantilist policies, monopolies, and state-sponsored privileges are its central enemies that corrupt markets and entrench power

• Smith’s defense of liberty encompassed an opposition to slavery, a critique of empire and a support for education

Highlights

[03:12] Why do we work so hard?

• Smith’s view of human nature combines a propensity to “truck, barter, and exchange” with imagination, speech, and reason, generating specialization, trade, and a drive to improve one’s condition

• He drew a distinction between the desire for praise versus praiseworthiness. Beneath the wish to be admired lies a deeper wish to be admirable by an impartial spectator; Smith treats the confusion of these two aims as a major source of dissatisfaction and moral error

• His story of the poor man’s son shows that ambition and the “deception” of wealth can simultaneously undermine individual tranquility while mobilizing creativity and innovation

• Prosperity in Smith requires more than bare subsistence: flourishing entails higher and rising living standards, lower mortality, and the capacity to raise families successfully.

• For Smith, “truck, barter, and exchange” is not just about hard work: there is a lot of creativity involved. Keynes perhaps missed this in his essay on the Economic Possibilities for our Grandchildren

[20:50] Do we admire the wrong people?

• In practice, admiration can focus too much on visible wealth and rank (observable) rather than virtue and intelligence (unobservable), corrupting moral sentiments and misaligning social esteem with genuine merit.

• The tendency to admire the rich and powerful and neglect others is “the great and most universal cause of the corruption of our moral sentiments”

• Progress for Smith is tied to the majority’s high and rising incomes (a “progressive” state), making median conditions more relevant than aggregate GDP per capita

• Although he did not use the term “conspicuous consumption”, Smith was worried about “trinkets of frivolous utility” and how people often chase status through small, often useless luxuries

[27:57] The invisible hand and its enemies

• Smith only used the term “invisible hand” three times in his work, and not in the context we currently use it. Yet it captures his idea that cooperation can emerge without intent or benevolence through self-interest

• However, to sustain this custom alone is not enough. It requires justice: clear rules enforceable by a governing body against coercion, fraud, and domination

• Self-interest differs from selfishness or rapacity: self-interest can include concern for family and others, while greed undermines justice and turns markets into instruments of exploitation. Smith has often been misread as a hyper-individualist

• Factions and special interests (mercantile lobbies, monopolists, guilds, established churches) can capture legislatures, secure privileges, and become systemic enemies of competition and the invisible hand. They can also distort our search for approval from the impartial spectator to the faction

• Smith’s book is an attack against government-granted monopolies (e.g., East India Company) and as a result an attack against mercantilism and empire building

• Property rights, especially in labor (the poor person’s “sacred patrimony”), are central, but especially with landowners Smith worries that concentrated economic power can lead to rentierism

• At the same time Smith was very nuanced, allowing for very targeted and exceptional protection. For example, in his discussion of the Navigation Acts he prioritized defense over opulence

[53:30] Liberty, slavery and the American colonies

• For Smith, liberty involves freedoms of occupation, religion, and family formation. He saw slavery as equal to death: “the most cruel violation of the most sacred rights of mankind”

• At the same time, and despite the colonial exploitation, Smith saw more of these liberties in American colonies than in Britain.

• He viewed empire and colonies as a drain on Britain, with Smith favoring either taxation with representation or separation combined with freer trade.

[57:52] Education

• Smith shared the Physiocrats’ worries about powerful landowners. However, he disagreed with them in that he did not see land as uniquely productive, emphasizing labor productivity, innovation, and early notions of human capital accumulation

• He supported education on the basis that extreme specialization can render workers “stupid and ignorant”

• However, he was skeptical of state-endowed universities and insulated teachers, reflecting his broader hostility to monopolies; Smith proposed that students pay their teachers directly

Subscribe now

* Summary produced by Pablo Balsinde (PhD student, Stockholm School of Economics)

Watch the full talk below. A summary in four bullets:*

• We see an extremely wide range of forecasts for AI’s growth impact—from Dario Amodei (2026), who projects 10–20% annual growth, to Daron Acemoglu (2025), who estimates 0.5% total additional growth over a decade

• Given the uncertainty and the lack of clear historical precedents, managers’ expectations may be our best available data

• In their recent working paper, Bloom and coauthors (2026) leverage existing central bank surveys to ask 6,000 executives about AI

• They document four facts:

  • 1) ~70% of firms use AI

  • 2) ~75% of executives use AI, with an average 1.5 hours per week

  • 3) AI has had little impact on employment and productivity in the last 3 years

  • 4) AI is expected to double productivity growth over the next 3 years

Highlights

[08:41] A new age of productivity growth?

• The history of productivity growth has had three phases: (1) near zero until around 1800; (2) an acceleration until the mid-20th century, peaking around 4% in the 1960s; and (3) a secular slowdown since then, to roughly 1% today

• Bloom et al. (2020) argue that the decline in productivity growth is due to the fact that ideas are getting harder to find. Will AI reverse this trend?

[19:05] What is the current data on firm AI use?

• Individuals’ AI adoption has risen from ~10% in the spring of 2023 to 30–40% today

• The Survey of Working Arrangements and Attitudes (also run by Bloom and his coauthors) shows people’s AI usage is more common at home than at work, and is strongly correlated with education, industry, and occupation

• Yet, we have much less data on firms’ AI usage

• Official business surveys, such as the US Census Business Trends and Outlook Survey, often report low and flat AI adoption (~17%). They likely understate true usage because responses are provided by junior employees lacking comprehensive knowledge of their firms

• At the same time, many surveys where firms are paid to participate suffer from the “impostor problem”, being filled largely by bots

[24:58] The survey

• They leveraged existing firm survey panels in US, UK, Germany and Australia. In all cases except Australia they are run by the central bank

• The surveys are mainly answered by the CFOs and CEOs of firms that have an average of ~250 employees

• Executives, while unpaid, participate because they see the data, and can see how they are benchmarked against other firms. Since answers remain private, they have no incentive to overreport AI usage

• The surveys ask managers to forecast future sales, so they validate the responses by comparing them with realized sales; forecasts track actual outcomes closely

[29:29] Results on firms’ AI use

• Only 69% of executives use AI for at least 1 hour a week. 29% do so between 1 and 5 hours a week

• 78% of firms use some form of AI technology. Firms that use more AI are younger and larger, and have higher labor productivity, better paid employees, and younger directors

• Last three years: 91% report no productivity impact from AI; on average, respondents estimate AI increased sales growth by 0.24% in total

• Next three years: 39% expect no productivity impact from AI; on average, respondents expect sales growth of 2.25% in total from AI. Expected gains are ~60% lower in Germany and Australia than in the U.S.

• These estimates of AI’s future impact are striking, since productivity growth has been around 1% for the last few decades

• They may justify cutting rates if they signal a sustained productivity-driven expansion in supply that lowers inflation pressure, as Kevin Warsh has suggested

[43:39] Employment effects and employee responses

• In the next three years, only 42% expect no AI impact on employment; on average, respondents expect employment declines of 1.2% in total. Two thirds of this comes from less hiring

• They also asked employees the same questions. Workers report a similar AI usage (1.8 hours per week)

• Workers expect much smaller productivity gains (0.92% in the next 3 years). But are more optimistic on employment, expecting AI to increase it by 0.45%

• One may argue employees’ responses are more accurate because they are closer to production bottlenecks, but executives are largely accurate on their sales forecasts. Executives’ confidence of their answers is also at pre-pandemic levels

Subscribe now

* Summary produced by Pablo Balsinde (PhD student, Stockholm School of Economics)

A summary in three bullets:

• How much of past economic growth is due to automation? Around half. Jones and Tonetti built a standard task-based model to answer the question, and then simulated the model to speculate about the future of AI

• Automation boosts growth by switching from slowly improving humans to rapidly improving machines on an increasing number of tasks. However, it can also create bottlenecks and spread capital thin. TFP growth may slow even as automation advances due to the slow-productivity-growing tasks that will still have to be done by humans

• AI may deliver explosive growth, but with a 75-year delay due to the weak links. Humanity may have time to solve its problems around inequality, political economy, and AI’s existential risks

Highlights

[04:19] The basic weak-link model of automation

• The model has a final good which is produced by combining a variety of tasks with a constant elasticity of substitution across them. The EoS is smaller than 1, so tasks are complements

• Firms providing tasks have a binary choice over using labor or capital. Each has its own task-specific productivity and they are perfect substitutes

• The richness of the task-based framework comes from the interaction between task complementarity and labor-capital substitutability

• A key assumption is that the new tasks being automated are those with the highest labor costs, or the least productive: Moravec’s (1988) Paradox

[26:20] How much of economic growth comes from automation?

• Calibrating the model allows for growth accounting, decomposing growth between:

  • (1) Increases machine productivity

  • (2) Increases in human productivity

  • (3) A residual for other drivers of productivity like misallocation

• Since 1950 they find that machine’s task level productivity has growth 5% faster than human productivity, and that 2% of tasks are automated every year

[48:03] The Future Consequences of AI

• Some intuition for the weak-link mechanism: assuming infinite productivity in current software tasks raises GDP only by (≈2%) because growth is still dominated by the remaining weak

• To study the long-run impact of AI, the model is enriched to allow for endogenous idea generation and for automating both the production of goods and ideas

• The model delivers three scenarios for economic growth in the next centuries:

  • (1) Full automation of all tasks: humanity reaches infinite income in ~2 centuries. The labor share of income goes to zero

  • (2) Incomplete automation: a set of tasks is permanently done by humans, slowing growth in the long run. The capital share falls to zero.

  • (3) A baseline case with vanishing but never-zero human tasks and a constant capital share of around 1/3

• There are two forces: (1) the flywheel effects where the automation of ideas yields more ideas and more automation, (2) the weak link from non-automatable tasks.

• In all scenarios explosive growth takes decades to arrive due to the weak links

Subscribe now

* Summary produced by Pablo Balsinde (PhD student, Stockholm School of Economics)

Watch the full talk below. A summary in four bullets:*

• In the episode, Nageswaran presented his vision of India’s growth strategy, included in India’s 2025–2026 Economic Survey, which has been described as a disruptive shift in India’s growth strategy

• The strategy is based on the idea of Investment in National Strength, with three pillars: (1) Intelligent import substitution, (2) Strategic resilience, and (3) Strategic indispensability

• National reforms in tax, labor, infrastructure, public procurement, state-level deregulation, and R&D support aim to lower input costs, crowd in private investment, and build a Mittelstand-style manufacturing ecosystem

• India remains an oasis of macro stability in a turbulent world, with sustainable non-inflationary growth, fiscal prudence, supply-side reforms and investments

Highlights

[05:36] India’s manufacturing future

• India’s strategy of Investment in National Strength has three pillars:

  • Intelligent import substitution: building domestic competitive production

  • Strategic resilience: The ability to function if the world breaks down

  • Strategic indispensability: The world being unable to function without India

• The goal is to grow manufacturing’s share of gross value added from 17% to 25% by 2047. There are three reasons behind the manufacturing push.

• First, essential goods (e.g. oil, metals) become strategic assets under fragmentation. We need to start thinking about stockpiling and vulnerability to beggar-thy-neighbor behavior

• Second, when trade, markets and supply chains become instruments of power, there are strategic reasons to promote manufacturing rather than just employment

• Third, manufacturing promotes currency strength. After Bretton Woods, countries with stable currencies have been those with robust manufacturing (e.g. Germany, Japan…)

• Manufacturing is crucial to lowering a country’s cost of capital, eliminating the FX depreciation risk premium, and avoiding self-fulfilling currency crises

• China’s experience illustrates that rising per capita income drives import demand; India’s rising resource needs will have to be financed via exports and capital inflows

• Both of these are harder in a fractured system. After Trump’s tariffs, India saw capital outflows to the U.S. and lower-tariff “China+1” countries (e.g. Vietnam), weakening the rupee

• Putting everything together, if essentials aren’t secured domestically (or via buffers), shocks force higher imports and external financing needs: this pressure can turn into currency weakness and, over time, into an exchange-rate depreciation risk premium in the cost of capital

[20:51] Designing intelligent import substitution

• The key idea of intelligent import substitution is that domestic industries should be held to the standards of global competitiveness; this is the East Asian playbook (Studwell, 2013)

• Intelligent import substitution is never permanent: (1) it focuses on areas where there is a natural advantage, (2) it makes protection conditional on productivity improvements, scale and eventual competitiveness, and (3) it subjects domestic industries to ruthless internal competition

• The Economic Survey suggests prioritising sectors based on a 2x2 matrix of urgency and feasibility. High urgency sectors are core, while in low urgency sectors, India can be more selective

• High-urgency and high-feasibility sectors (e.g. ag commodities and pharmaceutical ingredients) require demand assurance and procurement policies

• High-urgency and low-feasibility sectors (e.g. batteries, solar) justify public investment and mission-style programs

• Low urgency and high-feasibility sectors (e.g. industrial machinery and EVs) can be improved by gradual localization and clusters

• Low-urgency and low-feasibility sectors (e.g. electrolyzers, tunneling machines) receive ecosystem and talent support, but no aggressive import substitution

[31:17] A reform agenda for strategic resilience

• India’s industrial policy reforms aim to solve India’s lack of scale and create a manufacturing ecosystem analogous to Germany’s Mittelstand. The solutions can be grouped into three:

  • (1) Ecosystem-first reforms

  • (2) Infrastructure and talent provision

  • (3) A bureaucratically reformed entrepreneurial state

• Ecosystem-first reforms emphasize cluster-led growth, concentrating industries in individual states to promote economies of scale

• They also focus on input-cost reduction. For example, India has recently eliminated its import duty “inversion,” where intermediate goods used to have higher duties than final goods

• Government capex has grown from ~1.7% to ~3.1% of GDP over a decade

• India has achieved this while halving the deficit in the last five years, bringing ratings improvements

• The goal is to shift mindsets in bureaucracy from extraction to development and risk-taking with resilience, especially at lower levels of government. A precedent for this is the government’s efforts to boost the software sector (Mistree, 2019)

• Recent reforms since 2024 include:

  • Income-tax simplification and a revamped goods and services tax

  • A modern and unified labor code and a redesign of the rural employment scheme to improve job guarantees and reduce leakages

  • An infrastructure procurement policy based on qualifications rather than the lowest cost

  • New trade agreements

  • Liberalising the insurance and nuclear sectors and increasing R&D funds

• A state-level deregulation initiative has seen high implementation rates. The principle has been experimentation and learning from others rather than harmonization

[45:49] Growth in strategic sectors and under global fragility

• India underperforms in global influence relative to its potential given size (Lowy Asia Power Index)

• Potential growth, previously assessed at 6.5%, is now internally upgraded to about 7%. India remains an oasis of macro stability in a turbulent world, with sustainable non-inflationary growth, fiscal prudence, supply-side reforms and investments

• India is implementing public digital services (IDs, payments) while deploying an AI strategy (data center incentives, sectoral applications) to raise state capacity

• Nuclear power policy has been liberalised to diversify away from thermal power, allowing private and foreign investment and offering customs exemptions for nuclear equipment. The goal is to raise nuclear’s share to 7–10%

• India’s banking system is healthy; in the financial crisis, India did not suffer that much, and it did not need liquidity lines

• In the last few years, the stock market has become more responsive to domestic flows than to international capital. Nevertheless, the risk is that a more fragmented global financial market will make India’s current account deficit harder to finance

• Three scenarios for the world economy in 2026:

  • A continuation of ‘business as in 2025’ with less security and more fragility (40–45%)

  • A disorderly multipolar breakdown (40–45%)

  • A systemic shock cascade (10–20%)

Subscribe now

* Summary produced by Pablo Balsinde (PhD student, Stockholm School of Economics)

Watch the full talk below. A summary in four bullets:*

• In the talk Green presented his recent paper (Chen et al., 2025), which shows that LLMs exhibit the same behavioral biases documented in humans (optimism, overconfidence, extrapolation, and framing effects), despite demonstrably “knowing” the behavioral finance concepts studied

• When asked to rank stocks by expected returns, models strongly extrapolate from past returns, with prompt engineering only modestly reducing the bias

• LLMs are overly optimistic about expected returns, while pessimistic about upside (90th percentile) returns

• LLMs are more optimistic in predicting returns when historical information is provided as return charts rather than price charts

Highlights

[02:42] Literature and overview

• While 57% of investors use AI for stock analysis and research, around a third of them use it to make final buy–sell decisions (The Motley Fool, Blankespoor et al., 2026)

• A large literature has shown that AI use can improve outcomes in investing, sell-side research, auditing, and corporate governance, while they also embed social biases (for example, in medical advice or loan approvals)

• Some evidence suggests LLMs can predict returns with news headlines (Lopez-Lira and Tang, 2025), while others show that LLMs exhibit human-like extrapolative sentiment: if the news are good today they will be good tomorrow (Bybee, 2023)

• In Green and coauthors’ paper (Chen et al., 2025), they examine LLM bias in making stock return forecasts, in settings where human biases are well-documented

• They run separate calls to LLMs’ APIs to eliminate the path dependence from chat histories throughout (they consider GPT-4o, Claude 3.5 Sonnet and Gemini 2.5 Pro)

• They focus on four behavioral biases:

  • (1) optimism (Weinstein, 1980),

  • (2) overconfidence (Kahneman and Tversky, 1979; Ben-David et al., 2013),

  • (3) extrapolation, that is placing excessive positive weight on recent stock returns (Da et al., 2021), and

  • (4) framing effects (Hartzmark and Sussman, 2024; Glaser et al., 2019)

• They show that LLMs largely exhibit these biases, despite them demonstrably “knowing” the behavioral finance concepts studied

[22:20] Extrapolation in stock returns and market sentiment

• The paper builds on the setting of Da et al. (2021), which studied crowdsourced earnings forecasts (ForceRank). Human participants ranked 10 stocks by expected weekly performance

• Human forecasts loaded positively on lagged returns (trend-followed) even though realized returns exhibit short-term reversals (negative autocorrelation)

• Being provided past weekly returns, when asked to rank the same stocks GPT-4o strongly extrapolated from recent returns, with an especially high weight on the most recent week

Markus’ Academy: own elaboration. Lighter color indicates lack of statistical significance.

• When introducing separate coefficients for positive and negative return lags, humans extrapolate from poor performance more than from good performance. ChatGPT extrapolates both, with a somewhat stronger emphasis on past positive returns

• LLMs’ extrapolation bias persists when adding firm controls and when using simulated returns, discarding the possibility of look-ahead bias from the return data being in the LLMs’ training data

• The behavior is largely common to all LLMs tested, with models justifying predictions by pointing to recent return trends

• Replacing returns with cash flow changes does not change the LLMs’ behavior, indicating a generic tendency to extrapolate

• When asked to produce aggregate stock market sentiment measures analogous to survey expectations, LLMs extrapolate from recent returns even though realized returns show no trend continuation; they do so more strongly and for longer horizons than humans

Markus’ Academy: own elaboration. Lighter color indicates lack of statistical significance.

• The paper repeats the same analysis as above, but improving the prompt so as to have the LLM:

  • (1) think step-by-step,

  • (2) leverage a statistical model,

  • (3) avoid biases documented in the behavioral literature, and

  • (4) beware of extrapolative bias as in Greenwood and Shleifer (2014)

• This only modestly reduces extrapolation (at most one-third of the effect)

[42:43] Optimism in the distribution of stock return forecasts

• Ben-David et al. (2013) asked CEOs to forecast returns and to provide confidence intervals, showing that realized returns are within the intervals only 36% of the time

• Similarly, Green’s paper randomly selects 500 stocks over the last century, and asks LLMs to provide a return forecast and confidence intervals (available to the model are 10 years of monthly return history)

• While the historical mean return is 1.3% and the next-month realized mean return is 1.12%, LLMs’ mean forecast is about 2% per month, indicating strong optimism

• Hartzmark and Sussman (2024) find that asking individuals to provide forecasts across return bins, rather than point forecasts, reduces optimism, but for LLMs doing so did not help

• When asked for 10th and 90th percentile forecasts for next-month returns, LLMs’ 10th-percentile bounds align reasonably with empirical 10th percentiles, but 90th-percentile forecasts are materially below the empirical 90th percentiles. LLMs are thus optimistic in their mean forecast, but pessimistic about the upside

• Improving the prompt so as to point the LLM to the results of Hartzmark and Sussman (2024) does little to reduce optimism or correct upper-tail pessimism

[54:23] Framing effects

• To study the impact of framing, LLMs were asked to forecast both prices and returns, providing the past return data either as price charts or as return bar charts

• Glaser et al. (2019) showed that, in this setting, humans are more optimistic when forecasting returns rather than prices, and more optimistic when information is framed as price charts rather than return charts

• LLMs’ expectations for returns are largely invariant to whether the model is asked to forecast returns directly or to forecast prices

• However, and opposite to humans, they are more optimistic when information is provided in return charts rather than price charts

[59:09] Implications for using LLMs in finance

• The persistence of bias even under prompt engineering suggests that effective mitigation requires model-level changes like explicit fine-tuning. If requested, models still help users get payday loans or invest in the style of WallStreetBets

• The goal should be to have two types of LLMs: (1) human-like models for simulating public behavior (e.g., what gets people to exercise) and (2) expert models constrained to data-driven inference and conservative and model-based financial advice. Today’s systems are a weird mix of both

• Combining robo-advisors (machine-learning asset allocation engines) with LLM front-ends that interpret the robo-advisors’ recommendations could make LLM biases consequential for household portfolios

Subscribe now

* Summary produced by Pablo Balsinde (PhD student, Stockholm School of Economics)

Watch the full talk below.

Acknowledgements

We are grateful for feedback from Lord Robert Skidelsky, Harald Hagemann, and Harold James.

Production by Pablo Balsinde, PhD student at the Stockholm School of Economics.

Timestamps

[00:00] Introduction
[01:22] Economic Possibilities for our Grandchildren
[04:22] Alan Turing
[06:58] J-Curved Transitions
[10:07] Populism
[13:06] Jevons' Paradox
[16:09] AI Platforms
[18:43] Stock Market Bubbles
[22:07] The Master Economist
[24:48] The Importance of Economics
[25:57] Conclusion

A real-life recording of Keynes

Subscribe now

• The first part is a tutorial for how to use Cursor, an AI-native environment that bakes LLMs directly into the editing workflow so that models can read your repository, edit multiple files, run “agent” tasks, or explain code in-context

• The second part presents Ben Golub and Yann Calvó’s start-up Refine.ink, an AI tool to generate referee-style feedback to academic papers, identifying errors in math and empirical strategy, clarity problems, and consistency issues

• The third part walks through some best practices when prompting LLMs. Much of Golub’s advice is also provided by Goldsmith Pinkham (2024)

AI-native environments: Cursor

• Cursor, built on VS Code, is an AI-native development environment that connects to GitHub repositories containing LaTeX, code, and bibliographic files, replacing or complementing Overleaf-based workflows

• Within the environment one can then access all of a project’s documents. One can then assign tasks to LLMs and, through retrieval-augmented generation, attach files or snippets directly, improving performance

• The value of Cursor is that it is a much better orchestrator of different tools than chatbots, which are confined to your browser. Often with chatbots it can feel like managing a small bureaucracy with different tabs

• Within Cursor, Claude Opus 4.5 is best for arduous and simple tasks, like searching for papers and writing BibTeX entries. GPT 5.2 is best for math-heavy tasks

• Rather than attempting everything in a single pass, decompose technical tasks into staged prompts. For example when writing proofs: (1) first ask the model to assimilate the existing work, (2) then ask it to generate background notes, (3) then to design a proof strategy, and (4) finally ask it to produce the proof

• For those that prefer single-shot prompts, one can provide high-level instructions to orchestrate, for example telling the model to “orchestrate a good software engineering team for different tasks.”

• Cursor is largely similar to GPT Codex and Claude Code. The key benefit is that it is model-agnostic, allowing for mixing and matching models to tasks

• If you are comfortable with the basic computer terminal but not an expert in software deployment, it might be best to grant Cursor permission to run basic command-line steps like installing packages

Refine.ink

• Faster AI-assisted research can reduce the slow, deep engagement that typically surfaces mistakes, increasing the chance that errors persist into drafts.

• Generic chatbots cannot guarantee logical consistency, creating demand for a dedicated, end-to-end audit layer.

• Refine is an AI-based academic reviewing service that produces deep referee-style reports, identifying issues such as errors in logic, clarity problems, and consistency issues. It can produce a referee report at the level of a better-than-average PhD student

• Specialized orchestration and domain‑tuned workflows make Refine’s commentary systematically deeper and more coherent than generic chatbots

• For example, Refine points out definition inconsistencies, flags clustering choices, and questions an IV’s logic.

• Testimonials from Omer Tamuz and Drew Fudenberg report that Refine detects subtle mathematical errors and inconsistencies that would take chatbot-assisted human experts many hours to uncover

• Refine works for both theory and empirical papers (as well as in fields such as applied mathematics, computer science, and physics). For empirical work, the main current limitation is that tables and figures often cannot be appropriately parsed, an issue the tool’s creators are actively working to improve

• It is best to ask Refine to give at most 10 comments on your paper. Use it before you circulate or submit papers; when you care about correctness

• Privacy is contractually enforced: uploaded papers are neither used for training nor exposed to external model providers

• Top journals are piloting Refine as a complement to human peer review before publication with ethical use enhanced by disclosure in referee reports.

Heuristics for working with AI

• Treat LLMs as a brilliant but limited junior assistants: strong intuition, algebra, and coding skills, but weak understanding of higher-level project objectives and norms

• Post-training biases models towards overconfident answers rather than acknowledging ignorance. It also biases (especially chatbots) toward narrow, local task completion rather than cross-domain discovery

• Models’ “tunnel vision” leads them to execute the immediate instruction, ignoring the broader goal, even by sacrificing important checks or commenting out other code

• Prompt for rigor, not just content: pair instructions with explicit behavioral constraints on rigor, notation discipline, and stylistic consistency—especially by specifying the role and audience (e.g., “a senior probabilist writing for Econometrica”)

• Ask it to reason step-by-step, even if you later request a concise final output. Ask it to explain things back to you in the process

• Decompose tasks to improve reasoning quality. Ask for a plan, and then implement it in small steps

• Don’t fight the model. After a wrong answer, editing the prompt or starting a new chat is better than arguing within the same chat: the wrong answers can poison the model, inducing it to reconcile previous errors with later content

• Ask for “handoff reports”, key summaries of what has been done and next actions. Then feed these reports to the next chat. Example: “Write a handoff to a junior colleague: include all essential information to continue the task.”

Subscribe now

*Summary produced by Pablo Balsinde (PhD student, Stockholm School of Economics)

Watch the full talk below. A summary in four bullets:*

• In contrast with Acemoglu (2025), Aghion and Bunel (2024) estimate the impact of AI on productivity growth at 0.68% per year; this excludes AI’s additional effect on growth from making ideas easier to find

• Market power in cloud, compute, and foundational models risks turning AI into a superstar-firm technology that depresses new firm entry. Competition policy needs to be updated to address the impact of concentration on innovation

• So far evidence suggests that AI adoption in firms raises employment via productivity and expansion effects, with displacement concentrated in a small set of highly exposed and substitutable administrative jobs

• Education focused on “learning to learn”, flexicurity-style labor market institutions, DARPA-like innovation institutions, and a renewed competition policy will be central in managing AI’s creative destruction and securing Europe’s AI position

Highlights

[8:09] New estimates of the impact of AI

• Aghion et al. (2019) argued AI would raise growth by automating tasks in the production of goods and services, but also by making ideas easier to find and recombine

• They leverage Zeira’s (1998) model of automation and growth, showing that the growth of GDP is driven by AI’s capital share (driven by its automation of existing tasks) but also by the rate of growth of ideas induced by AI

• Brynjolfsson et al. (2023) present micro evidence from a Fortune 500 company that sells business-process software, showing that AI can lead to large within-firm productivity gains: 14% in the first month and 25% in the first five months

• To assess AI’s aggregate implications, one approach is to look at prior innovations: electricity increased GDP growth by 1.3 percentage points per year for 10 years, after which growth returned to its previous trend, while IT did so by 0.8% per year

• Acemoglu (2025) instead follows a task-based approach, decomposing aggregate TFP gains into:

  • 1) The share of exposed tasks

  • 2) The share of these exposed tasks that can be profitably automated

  • 3) Labor-cost savings per automated task

  • 4) The labor share adjusted for AI exposure

• Borrowing from the empirical literature’s estimates for each of these parts, he arrives at an estimate that AI will only boost TFP by 0.07% per year

• Aghion and Bunel (2024) maintain the task-based approach, but reassess the estimate for each component

• On the share of AI-exposed tasks, Acemoglu used Eloundou et al. (2024)’s estimate of 19.9%. However, based on Gmyrek et al. (2023) and Pizzinelli et al. (2023), the interval for the share of exposed tasks is between 18.5% and 68%

• On the share of profitably-automated tasks, Acemoglu leverages Svanberg et al. (2024); looking only at visual tasks, they estimate the share to be at 23%. However, Acemoglu ignores their scenarios of further cost reductions, under which the share could rise to 80%. The shares could be higher for language tasks, while these calculations assume that the tasks remain fixed

• On the labor cost savings, Acemoglu considers 27% by taking the average between Noy and Zhang (2023) estimate of 40% and the lower bound (14%) of Brynjolfsson et al. (2023)’s 14-25% interval. When taking the upper bound (AI’s impact after 5 months), the relevant interval for the labor-cost savings is 33%-40%

• Putting all of this together, the AI’s impact on annual TFP growth is in the interval of 0.08% to 1.24% over the next 10 years. Aghion’s median scenario is 0.68%

• This is slightly lower than IT’s impact, but, beyond automating tasks, AI will bring an additional permanent effect on growth from making ideas easier to find, as illustrated by Aghion and Bouverot (2024):

• Evidence of this permanent effect is Bergeaud (2025), which documents an explosion in the share of US patents related to AI. One can show that AI patents are more general, more cited, more original and of higher quality

[28:02] Superstar firms and renewing AI competition policy

• The downside to this is the lack of competition. The IT revolution raised US TFP growth between 1995–2005, but was followed by a slowdown and a rise in concentration

• Aghion et al. (2023) show that rising aggregate markups are mainly driven by composition effects: as IT lowers overhead costs, high-productivity and high-markup firms expand into new products, raising aggregate markups even though within-firm markups can fall

• Consistent with this, Autor et al. (2020) and De Loecker (2020) find that the concentration was driven by superstar firms across many sectors

• AI’s value chain (cloud provision and GPUs) is already dominated by a handful of large incumbents, weakening entry and diffusion

• However, competition policy has not adapted, and Aghion’s outlook for reform is pessimistic. Boosting open-source models, limiting overregulation that disproportionately burdens entrants, and data-sharing obligations (e.g. extending EU Digital Markets Act principles to cloud and AI) would mitigate incumbent advantages

[33:48] AI and employment

• Surveying French firms, Aghion et al. (2025) showed that AI-adopting firms expand employment relative to similar non-adopters: as productivity and sales grow so do firms’ labor demand

• Gmyrek et al. (2023) classify jobs according to their exposure to AI and the extent to which AI substitutes or complements workers. Highly exposed and highly substitutable jobs represent around 10% of workers

• However, Aghion et al. (2025) show that this 10% may be too large. Even in highly exposed and highly substitutable jobs, if AI is used in production or IT security it raises employment. Only AI used in administrative processes reduces it

• In ongoing work with Danish data, Aghion and coauthors instrument firms’ AI adoption with whether a firm’s manager has a spouse employed at an AI-adopting firm; the IV estimates show similar increases in employment

[43:43] Conclusion and Q&A

• So far there is no existential risk from AI: no mass unemployment. Although firms see higher labor turnover, they also see higher job creation due to AI’s productivity effect

• AI-driven creative destruction requires systems that facilitate the transition to new jobs; Denmark’s flexicurity model (high income replacement plus active retraining and placement) provides a model

• Roulet (2020) shows that, unlike in other countries, workers in Denmark who lose their jobs due to firm closures do not experience deteriorating health outcomes

• Education systems should prioritize “learning to learn” (reading, writing, reasoning, math) without AI, ideally with homework done at school, allowing only for controlled AI use for targeted support like remedial tutoring

• Vocational systems like Germany’s and Switzerland’s should keep strong task-specific training but add more general education to boost adaptability. Yet, France’s heavy tilt toward general education has contributed to higher unemployment

• Governments can steer firms toward more worker-friendly AI adoption by rewarding training, internal mobility, and promotion practices and by penalizing excessive layoffs through experience-rated social-insurance contributions.

• It is very hard to control the development of AI. The EU’s overregulation of AI may have driven AI firms elsewhere. It cannot catch up in everything. It should leverage its strengths to develop AI in sectors like health and energy

• The EU should take the Draghi report’s call for a genuine single market seriously and add DARPA-style mission agencies. Coalitions of willing countries can move first, building joint innovation markets and shared DARPA-like institutions, while keeping any support for “champions” firmly pro-competitive

Subscribe now

* Summary produced by Pablo Balsinde (PhD student, Stockholm School of Economics)

Watch the full talk below, and watch a 15-minute version of the episode here.

Highlights

[6:12] The problem

• The book begins with the five stories of people that have had hurtful encounters with the personal finance system, from:

  • (1) excessive student debt

  • (2) undiversified retirement savings

  • (3) little wealth at retirement because they saved in money markets and rates declined

  • (4) worthless mortgage payment protection insurance

  • (5) high-fee life insurance + investment bundles

• These problems are systemic and growing as more people around the world enter the global middle class and are asked to solve harder personal finance problems (longer retirement, more expensive housing and education, fewer informal insurance mechanisms)

• People’s first encounters with the financial system can determine their trust in the financial system and market economy institutions

• Distrusting the financial system, people may turn to informal alternatives like loan sharks, which will worsen capital allocation in the economy

• The literature has shown that the poor earn lower (risk-adjusted) returns on their wealth, while they also pay more on their debt (Campbell et al., 2019; Bach et al., 2020)

• People struggle with financial decisions because human intuition is poorly suited to these. We anchor on salient numbers, and do not think about exponential growth. We extrapolate from the small samples of personal experience (Malmendier and Nagel, 2011), and confuse luck and skill

• Important decisions are infrequent, so there is little experiential learning. We cannot learn much from others because shocks are correlated, while others avoid talking about times they lost money. Social learning exacerbates FOMO

• Because of their difficulty people procrastinate taking these decisions, and do so only when they are under pressure to do so, often acting emotionally

• Competition alone cannot solve these issues. Institutions supply the products demanded, not those in people’s interests

• Rather than competing to reduce prices and increase quality, suppliers engage in unproductive competition (e.g. excessive bank branches, more realtors during real estate booms)

• They also actively obfuscate, offering complex and bundled products to prevent comparison

• Financial literacy alone isn’t enough, as many products have regressive cross-subsidies embedded in their offering

• For example, in the UK adjustable-rate mortgages offer low initial payments (teaser rates), but inattentive borrowers end up paying more once their loan switches to a standard variable rate, effectively subsidising the attentive who refinance before the switch (Fisher et al., 2024)

[38:45] The specifics: what goes wrong in personal finance

• Personal finance currently fails to help people smooth life’s ups and downs. With volatile incomes and limited savings or insurance, many are forced into costly short-term borrowing, often leading to debt traps

• The lack of equity market participation is endemic, while the system also fails at helping people make crucial decisions like choosing a student loan repayment plan, or choosing a fixed vs variable rate mortgage (overcomplicated by the points system in the US)

• Managing retirement has also become harder. People are living longer, receiving less family support, and must manage their own assets as defined-benefit plans give way to defined-contribution schemes. Persistently low real interest rates make funding retirement even more challenging

• We are failing at having people reach the recommended wealth-to-income ratio of 6x at retirement, and of 4x at age 50

• Many retirees are “house-rich but cash-poor”. This is the annuity puzzle (Modigliani, 1986), where too few people buy lifetime income annuities

• The annuities market is in a bad equilibrium: products that might convert illiquid housing wealth into income like reverse mortgages face high marketing costs, low consumer understanding, and are not standardised

[48:02] Solutions: shove, don’t just nudge

• Technology can lower fixed and search costs, and can allow for product customization. However it also makes it easier to exploit behavioral biases through things like gamified trading; it can also make it easier to price discriminate

• Alipay for example offers mutual funds to people by ranking them by past performance, encouraging performance chasing

• Financial education is valuable but insufficient. The industry keeps evolving, introducing new terms, tricks, and traps faster than education can keep pace

• Teaching finance in high school is abstract, as students are not yet making major financial decisions

• Nudges can deliver striking short-term results but often fade over time. They can also have unintended effects. Automatic enrollment in 401(k) can boost retirement savings, but this can be offset by increasing household debt

• Further interventions are needed. Consumer protection authorities should have the ability to “name and shame.” Regulators should have tools to change prices, including price caps if necessary

• A litigation safe harbor can also help reduce costs for suppliers. Regulatory action has often been retroactive, only punishing products ex post

• Regulators should impose a fiduciary duty on all financial advisors to mitigate conflicts of interest. However doing so on the entire financial sector is likely to hurt innovation and raise costs

• Market power can be curbed by requiring standardized product offerings and regulating units in which prices are quoted

• We can also fight biases with biases, promoting savings accounts aligned with mental accounts or prize-linked savings

• Direct government provision of financial services is ill-advised in today’s high-tech world

[1:04:49] A new vision: a personal finance “starter kit”

• Regulators should design standardized products and ensure it is mandatory to make them available: a personal finance starter kit. These products would become cheaper to market and easier to comparison-shop

• 4 design principles: simple, cheap, safe and easy. Having simple price structures would facilitate comparison shopping

• An example is the Basiskonto in Germany, a basic payment account. Transaction accounts today do not pay interest nor charge fees. This is confusing. Accounts should pay money market interest and charge clear fees

• APRs are hard to follow, and should be complemented with scenarios where dollar costs can be quoted

• Loans should avoid tripwire fees, such as overdraft charges or late fees, that trigger large penalties for minor delays. Fees should be continuous

• Credit should be linked to borrowers’ capacity to repay, their future income, rather than to current spending, resembling a paycheck advance rather than “buy now, pay later” schemes

• In mortgage markets, teaser rates (UK) and upfront points (US) should be eliminated. Ideally there would be automatic refinancing

• Insurance products should clearly disclose eligible claims and payouts

• Retirement saving is fragmented across a plethora of account types: profusion leads to confusion. There should be a single, standardized account with automatic enrollment starting at first employment

• These problems are serious not only because of their direct impact on individuals, but because the unpopularity of finance spills over into distrust of the market economy and the institutions that sustain it

• Personal finance is a structural societal issue, not just a self-help problem

• There have been proposals covered in the media to include private equity in 401(k) plans. The argument is that people should invest in the “market portfolio,” and that private assets should naturally be included

• However, these proposals should be approached with healthy skepticism. Individuals would likely not hold these private assets directly. They would do so through high-fee funds that may engender adverse selection, receiving lower-quality private investments than those of large institutions or ultra-wealthy investors

• Moreover, if participants can set their own portfolio weights, they are likely to overallocate to private equity, attracted by its artificially smooth reported returns. Universities have tended to overinvest in private assets for the same reason

Subscribe now

* Summary produced by Pablo Balsinde (PhD student, Stockholm School of Economics)

Watch the full talk below.

Highlights

[0:00] Markus’ Introduction

• There are two strands of literature within behavioral economics. The first emphasizes biases in preferences (e.g. reference dependence, loss aversion or hyperbolic discounting) and beliefs (e.g. probability weighting, overconfidence or confirmation bias)

• The second emphasizes the cognitive origin of noise, modeling the information acquisition of agents that are self-aware of this noise

• Thaler contributed the idea of nudging people to overcome biases, exposing them to information, providing default options (Thaler, 2004), or using one bias against another

• Thaler’s original The Winner’s Curse, was published in 1992 and was based on a series of columns in the Journal of Economic Perspectives (e.g. Thaler, 1988). With the new edition they look back and assess the field’s progress

[3:50] Economists should listen to Kahneman

• Economists should listen to Kahneman’s advice in Thinking, Fast and Slow: before you do something rash, check in with System 2

• There has been progress since the 1980s, when economists thought there was no alternative to expected utility. There is a large literature on loss aversion, reference dependence, fairness, salience, hyperbolic discounting…

• However the textbooks and standard principles have barely changed

• Economists should think twice about whether the models they write down are normative or descriptive

• von Neumann explicitly formulated expected utility as a normative model, but economists use it as the workhorse model of how people make decisions

• Together with the study of markets, what distinguishes us from other social sciences is that we model “agents” maximizing an objective function

• However we should think twice about the difficulty of the problem we are modeling. Take for example the complex problem of saving over the life-cycle. Keynes’ maxim that people spend a fraction of their income is not the right model, but it is probably closer to reality than the life-cycle models of Modigliani and Brumberg (1954) and Barro (1974)

• We should also think twice when we assume that people maximize utility. There is lots of evidence suggesting people are not good at making the forecasts required for doing so, while there is a long literature on preference reversals

• One alternative could be Herbert Simon’s idea of satisficing: when you find the first option that feels good enough, you quit searching. However we don’t have good satisficing models, and so we don’t know how they differ from maximization models

• Expected utility or the efficient market hypothesis are benchmarks; a null hypothesis. Prospect theory couldn’t exist without expected utility. But this does not mean we should write down rational expectations models where the market’s expectations are assumed to be in line with those of the best econometrician

• Will AI end behavioral economics if humans outsource all of their decisions, making us all hyperrational? Two points: it is unlikely that AI models themselves are hyperrational, while it is also unlikely that humans will adopt them everywhere (for example in their marriage decisions)

• In ongoing work with Sanjog Misra and Kevin Lee, Imas experimented with humans giving agents their preferences to an AI and having AI agents negotiate outcomes. Not only did people’s prompts inject their biases into their preferences, but they were less happy with the result because it came from a black box

• It is a 50-year-old result that simple linear models for decision-making most often do better than humans. Yet we did not switch to decision-making based on linear regression

[33:08] Progress in behavioral economics

• Early work documenting behavioral anomalies relied mainly on low-stakes, often hypothetical, lab experiments. The pushback was: “We don’t care about what college students do in the lab”

• Part of the reason behavioral econ has succeeded is that, since then, it has shown that the anomalies arise in real world settings, among professional investors, athletes, CEOs…

• Neuroeconomics was especially exciting around the mid-2000s, hoping to achieve an Edgeworth-style “hedonometer” to measure utility

• Although the use of fMRI data did not take off (it has become controversial even in the neuroscience community), using non-choice data to study decision-making is at the frontier of behavioral econ

• What are the cognitive foundations of the anomalies? Are loss aversion and narrow bracketing just parameters in the utility function, or do they arise from cognitive constraints? Can we understand them as the result of a maximization problem constrained by limited attention and memory? This is the frontier of behavioral economics research.

• Yet, there is always a risk that the new neural approaches feel novel without actually teaching us anything new

• Context effects matter, but we lack theory for them. For example the data shows heterogeneity in loss aversion but we do not yet understand its drivers. The new cognitive approach to behavioral economics has the potential to help us understand context effects better.

• Behavioral finance has been one of the most successful subfields because people make frequent, high-stakes decisions and get rapid feedback. Much of its evidence comes from the field rather than the lab; for example, the disposition effect (Odean, 1998) and overconfidence (Odean, 1999).

[40:12] Selling fast and buying slow

• The pushback against behavioral finance has been that the evidence of biases comes from non-experts

• To address it, Akepanidtaworn et al. (2023) study the buying and selling decisions of long-only institutional investors

• They show that these two decisions are different and not just two sides of the same coin, benchmarking investors’ performance against random strategy alternatives

• To assess buying performance, they compare what investors bought against comparable alternatives they could have (randomly) bought. They find that investors outperform the random strategy, and so display skill. There is no correlation between prior returns and what they buy

• To assess selling performance, they compare sales to a conservative “randomly sell an alternative holding” benchmark. Investors do drastically worse than the random strategy

• The natural explanation is limited attention. Two aspects predict whether a holding will be sold: (1) the salience of past returns (sell extreme winners and losers), and (2) within these salient buckets, investors sell the holdings they are least attached to

• That is, in line with the endowment effect, they sell stocks they have held for shorter periods of time. The problem is that the recent buys are the ones that are generating alpha

[48:00] The robustness and future of behavioral econ

• Behavioral econ has been largely spared from the replication crisis in the social sciences because it adopted the methods of experimental economics, which were different from those of psychology

• Going back to Vernon Smith’s work, papers have included the required instructions and data, while a condition for publication has been that papers replicate the relevant prior result (a good example is the classic asset-market bubble experiments; Smith et al., 1988)

• In their book, they replicated the main lab or field experiment of each chapter, and effectively found that all of Thaler’s (1988) anomalies hold: preference reversal, the dictator game, the ultimatum game, the endowment effect…

• All of the required instructions and data are available at thewinnerscurse.org. Teaching slides are also available

• Looking to the future, we need to think about the relationship between noise in observational data and the difficulty of the decisions studied. One answer is that difficulty brings noisier behavior, but often, as environments get more complex, people rely more heavily on heuristics, so behavior can actually look less noisy

• We also need more research to measure complexity. Defining what counts as “difficult” is itself difficult. Do computer science’s measures map to the ecologically valid response to complexity or the perception of complexity?

• Behavioral economics has become mainstream in departments and journals, yet, despite some texts adding behavioral flavors (Acemoglu et al., 2021), it is still underrepresented in the core textbooks

• The impact of AI is uncertain. It won’t make behavioral questions disappear any time soon; we will not all become hyperrational

• The intersection between cognitive science and AI is booming. Particularly with small and interpretable models, AI could help elicit people’s mental models and help represent parts of the judgment process

Subscribe now

*Summary produced by Pablo Balsinde (PhD student, Stockholm School of Economics)

Watch the full talk below. The highlights below were produced using artificial intelligence and may not necessarily reflect the views of Lawrence H. Summers.

Part 1: Technological Progress

[0:00] Technological change over Summers’ lifetime has been modest relative to his grandmother’s era, but AI, life sciences, and cheap energy could make his granddaughter’s lifetime comparably transformational.

[6:55] Solow Paradox or not? AI is too early in its diffusion for a new Solow paradox judgment, and current anecdotal evidence for prospective productivity acceleration exceeds that at Greenspan’s 1995 ICT inflection point.

[8:20] Labor augmenting vs. savings and the Jevon’s Paradox. AI’s labor-augmenting versus labor-saving effects hinge on long-run demand elasticities by sector, which are fundamentally uncertain and heterogeneous, making confident predictions about Jevons-type quantity responses unreliable.

[10:03] Recursive nature due to automation of ideas. Recursive AI that automates programming and research labor raises the probability of discontinuous productivity gains because, unlike past technologies, it directly accelerates the production of future ideas.

[17:27] Will the meaning of being human change? Human experience may fundamentally change through biological interventions, altered life-cycle expectations, and pervasive AI “chief-of-staff” agents that transform individual agency and organizational forms.

[21:10] Human community, social interaction and AI companion. DoorDash–Netflix–AI-companion bundles reduce the need to leave home, undermining face-to-face interaction, family formation, and the hive-mind social structures historically central to human progress.

[23:42] Designed or spontaneous progress. Governance of technological change faces a tradeoff between dangerous overconfident planning and risky unguided evolution, implying a preference for fast feedback, adaptation, and humility rather than comprehensive social engineering.

[28:04] Concentration of power. Current AI competition features multiple large model firms and overbuilding dynamics, suggesting less stable oligarchic dominance than critics claim, while historical fears of foreign techno-models (Russia, Japan, China) have repeatedly misfired.

[38:14] Distribution of fruit of technical progress and UBI. Future fiscal needs imply a larger public sector financed by higher, more broadly based and progressive taxation and expanded social insurance, but universal basic income is criticized as misaligned with human motivation and optimal transfer design.

Part 2: Populism.

[45:00] Technological and structural transitions, like the industrial revolution, can coexist with massive welfare gains and deep political crises, making contemporary populism plausibly linked to similar dislocations.

[47:00] Elites, university politics. Populism is reinforced by failing economic paths for non-sedentary boys and by elite institutions’ identity-focused, cosmopolitan agendas and secessionist stances (e.g., law schools) that alienate broader electorates.

[53:00] The democratic party in the US. The Democratic Party’s focus on highly educated liberals and identity groups has estranged traditional Roosevelt-era working-class constituencies, many of whom now vote for Trump, exposing a core strategic failure.

Part 3: Geopolitics

[56:08] U.S. strategy often violates the realist maxim of uniting allies and dividing adversaries, instead alienating partners (Europe, India, Southeast Asia) while encouraging closer ties among rivals like Russia and China.

[1:00:06] Can mistakes be undone? International trust erosion from recent U.S. political volatility fosters hedging by partners, but American history illustrates substantial resilience and self-correction through critical “Jeremiads,” now threatened by rising speech-related livelihood risks.

[1:04:01] The role of the US dollar. Loss of dollar centrality would mainly be a symptom of deeper geopolitical decline rather than its cause, while weak alternatives and fiscal complacency create a “bus stop dilemma” in assessing bond-market sustainability.

[1:10:19] Conclusion: Lessons for Economists. Effective economists should address policymakers’ stated concerns, privilege compelling statistics and narratives over platitudes, and rebalance from narrowly identified micro-questions toward big, messy phenomena in an Albert Hirschman-style tradition.

Subscribe now

A few highlights from the discussion.

Highlights

[3:08] Basics of tokenized money

• There was already a project for a global stablecoin 6 years ago: Libra. It was shut down by regulators because it created its own unit of account, threatening monetary sovereignty

• In the past there was a choice between transacting through currency (peer-to-peer, anonymous, but local) or through bank deposits (intermediated, non-anonymous, and possible at a distance). Tokenized money frees us from this dilemma: it is peer-to-peer, anonymous and at a distance

• There are several forms of tokenized money:

• (1) Bitcoin and others are pure fiat (no backing) digital currencies

• (2) Tokenized deposits mirror bank accounts, allowing you to use bank money just like with debit cards or transfers. They retain deposit insurance and access to the central bank, along with KYC/AML checks

• (3) With e-money a single issuer holds one reserve account and issues tokens people can use to transact (e.g. M-Pesa and WeChat). No bank account is required to access your tokens. Many developing countries are considering sending welfare payments through such wallet systems

• (4) Stablecoins are backed by non-monetary assets like government bonds or bank deposits. You can transact them without having a bank account, but they lack access to the central bank’s balance sheet

• With a bank account (traditional or tokenized) you hold a liability of the bank, and are a creditor to it. With a wallet a ledger records your tokens, but it is not a balance sheet. Wallets are object-based systems, not claim-based

[14:45] Can stablecoins become a generalized payment instrument?

• Stablecoins are confined to the crypto ecosystem, and have had no impact outside of it. The two main coins (Tether’s and Circle’s) account for the vast majority of the current $250bn market cap (BIS, 2025)

• The GENIUS Act gave stablecoins credibility as a monetary instrument, but did not ensure the viability of the business model

• It imposes requirements on the composition of reserves, transparency and licensing requirements, along with a designated supervisor

• Tether used to be backed in part by real estate and commercial paper, but now they will only be able to back their coins with bank deposits and short-term Treasuries

• The Act did not go as far as the Treasury had proposed in 2021, wanting to regulate stablecoins as banks with capital and liquidity requirements

• Although the Act mandates strict redemption rules, it provides few detailed requirements. Our experience with money market funds shows how destabilizing such rules can become

• The rationale for prohibiting stablecoins from paying interest is to ensure they remain only a payment instrument and do not become a store of value

• Paying interest would make stablecoins securities under SEC oversight, while the ban also helped secure banks’ support for the Act

• Stablecoins have found a way to circumvent this requirement: exchanges hold the coins and pay interest on the accounts of stablecoin buyers, prompting a strong reaction from banks

• Tokenized money is attractive to businesses, for example by improving the efficiency of payments along a supply chain (Brunnermeier and Payne, 2023). However these benefits can also be obtained through other forms of tokenized money. In general, the more centralized and well-governed the system, the more business-friendly it will be

• Whether households will use stablecoins for payments will depend on network effects and local alternatives. Stablecoins could drive dollarization in economies with weak currencies

[32:55] Stablecoins and the stability of private money

• The seigniorage earned by coin issuers will increase with the level of interest rates. It will also depend on their ability to control the amount of issuance. In recent years the sector has realized that it is hard to destroy money; it requires “open market operations” to buy it back

• Issuers will compete on the efficiency of payment, or perhaps on the laxity of controls

• Stablecoins are most often compared to money market funds, however coins guarantee a fixed value and aim to have greater liquidity and instantaneity

• They have also been compared to private banknotes during the free banking era (1837-1883). However at the time banks’ assets were much more opaque

• Narrow banks are perhaps the best comparison, although in theory these could offer nonzero interest rates

• Unlike stablecoins, currency boards are backed 100% by the pegged currency. They also have instantaneous redemption and are passive in the sense that they cannot control the amount issued

• In the past stablecoins have seen large deviations from their par values (BIS, 2025). A coin’s backing is different from redeemability: the first is about solvency, the second about liquidity. Central banks are there because these two do not coincide, but stablecoins do not have access to it

• Stablecoins’ stability will depend on their redemption rules and on the liquidity of the Treasury market, itself supported episodically by the central bank

• Stablecoins break the singleness of money because different blockchains cannot interact. Indeed, issuers’ incentives are the opposite: to prevent holders from redeeming

• They change the form but not the quantity of money. Even those who argue that the quantity of money matters for monetary policy do not tend to think the form of money does

• However, stablecoins could threaten central banks’ control over the unit of account, which is required for effective monetary policy (Woodford, 2003). Central banks can fix the interest paid on the unit of account, allowing them to control the real rate if people use the unit of account to price things

• Governments control the unit of account by controlling the medium of exchange (that is declaring it legal tender). The system rests on the coincidence between the unit of account and the medium of exchange

• There is a Hayekian argument in favor of stablecoins. By making financial repression and FX control more difficult they might keep inflation in check

[54:47] International monetary competition

• The GENIUS Act reflects the U.S. administration’s view of the international role of the dollar. They do not want the dollar to be a store of value, as it attracts capital inflows and appreciates the currency

• However they still want dollar dominance to fund government deficits. This new vision of dollar dominance is not based on a reserve status but rather on the dollar’s ability to fund deficits through digital network effects (with everyone coordinating on the use of dollar stablecoins)

• The drawback of this approach is that you expose yourself to competition from other networks

• The ECB is building a digital euro to prevent private issuances of euro stablecoins, seeing them as unstable

• The banks have found semi-allies in the American payment and credit card companies to defend the current system. The majority of cross-border retail payments in the EU are made by U.S. companies, with the data from these transactions going to the U.S.

• European authorities are aware that the majority of domestic credit is provided by banks. The ECB knows how to build a digital euro that does not threaten banks’ funding

• The digital euro should serve as a catalyst of digitization and a European Payments Union (which is arguably more important than the Capital Markets Union). The initial European Payment Initiative for a European credit card was abandoned, while the adoption of the new Wero wallet has been very slow

Subscribe now

* Summary produced by Pablo Balsinde (PhD student, Stockholm School of Economics)