AI

Philosophy

  1. Three key features of AI: (i) alien, (ii) encyclopedic, and (iii) generative.
  2. The joint hypothesis problem applies: output quality is a function of model ability and user input.
  3. Use AI most aggressively when the output is verifiable. Examples: formatting, extraction, citation checking, code with tests, and document summarization.
  4. The human is the router.
  5. Taste, breadth, and judgment become more valuable.

Tools

  1. Chatbots: (i) ChatGPT, (ii) Claude, (iii) Gemini, (iv) Grok
  2. Search and deep research: (i) Perplexity, (ii) ChatGPT Deep Research, (iii) Gemini Deep Research
  3. Plugins: (i) Claude for Excel, Word, and PowerPoint, (ii) ChatGPT for Excel, (iii) Microsoft Copilot, (iv) Google Workspace AI
  4. Agents: (i) Claude Cowork and Claude Code; (ii) Codex

Best Practices

  1. Role, problem, action, and constraints. The model needs to know who it is acting as, what problem you are solving, what output you want, and what rules it must follow.
  2. Durable artifacts. Use Markdown as the default working format, attach source documents when possible, and ask for citations, page numbers, and short quotes when the output depends on evidence.
  3. Separate generation from verification. Use AI to draft, extract, summarize, or code, but check important outputs with tests, source documents, a second model, or a separate critic pass.
  4. Use agents for divided labor. Drafting, critique, editing, search, and coding can often be split across agents, but coordination should run through clear Markdown handoff files.
  5. Manage input and attention. Use voice input for long prompts, keep context small and structured, and close loops before opening new ones.

Implications

  1. Macro Implications

    1. Acemoglu, "The Simple Macroeconomics of AI", 2024. Task exposure translates into much smaller medium-run macro gains than the largest AI forecasts imply.

      MoreLess

      The paper estimates an upper-bound TFP gain of about 0.66% over ten years, falling to about 0.53% after accounting for hard-to-learn tasks. GDP can rise somewhat more through investment, but the welfare-relevant productivity number is modest relative to aggressive forecasts. The distributional effect may still be large if AI widens the capital-labor income gap.

    2. Acemoglu, Autor, and Johnson, "Building Pro-Worker Artificial Intelligence", 2026. AI becomes pro-worker when it creates demand for new human expertise.

      MoreLess

      The paper distinguishes labor augmentation, capital augmentation, automation, expertise leveling, and new task creation. Only new task creation is unambiguously pro-worker, because it expands demand for human capabilities rather than commodifying them. The direction of AI is therefore not technologically fixed: developer incentives, buyer incentives, tax policy, antitrust, procurement, and rights over worker expertise all matter.

    3. Athey and Scott Morton, "Artificial Intelligence, Competition, and Welfare", 2025. AI market power can turn productivity gains into worker losses and upstream rents.

      MoreLess

      AI is often a priced upstream input, not a free productivity shock. If displaced workers absorb the substitution while concentrated AI suppliers capture fees, workers can suffer double harm from both displacement and AI pricing power. Broad gains require competitive AI pricing and productive absorption of displaced labor in other sectors.

    4. Baslandze, Edwards, Graham, McClure, Meyer, Sparks, Waddell, and Weitz, "Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives", 2026. For the typical firm, AI adoption is broad but shallow, and near-term employment and productivity effects are modest.

      MoreLess

      In a survey of nearly 750 executives, most AI spending takes the form of subscriptions, services, and training rather than hardware or capital deepening. Labor-productivity gains are positive but concentrated in high-skill services and finance, and the authors document a productivity paradox in which perceived gains exceed measured ones, likely reflecting delayed revenue realization. There is little evidence of near-term aggregate employment decline, though larger firms anticipate cutting routine clerical roles while smaller firms expect modest gains; the findings are expectations-based survey data, not realized outcomes.

    5. Bick, Blandin, and Deming, "The Rapid Adoption of Generative AI", 2024/2025. AI adoption is broad, but work-hour intensity is still modest.

      MoreLess

      By late 2024, nearly 40% of U.S. adults aged 18-64 had used generative AI, but only 1-5% of all work hours were AI-assisted. Self-reported time savings imply a potential aggregate productivity gain around 1.1% at current usage. The macro effect depends less on whether people have tried AI and more on repeated use, workflow redesign, and firm-level integration.

    6. Johnston, Holtz, Richmond, Ong, Tambe, and Chatterji, "The Shift to Agentic AI: Evidence from Codex", 2026. Agentic AI is diffusing rapidly beyond software developers and reshaping how much work a single user delegates.

      MoreLess

      Codex's active user base grew more than fivefold in the first half of 2026, with the fastest growth outside the original developer audience, and inside OpenAI adoption is nearly universal, largely displacing business ChatGPT use. Workflows are getting more sophisticated: over 10% of users run three or more concurrent agents in a given week and 26.6% use shareable "skills," alongside a near-tenfold rise in tasks that would take an experienced professional 8+ hours. The evidence is descriptive usage data from OpenAI, not a causal productivity estimate, and OpenAI-internal figures are an unusually early-adopting benchmark.

    7. Jones, "A.I. and Our Economic Future", 2026. The macro question is whether AI is another general purpose technology or a technology that automates intelligence itself.

      MoreLess

      Weak-links logic disciplines the extreme case: automating one task raises output in proportion to that task's initial economic share. Automating software alone is not the same as automating all cognition, and even broad cognitive automation may generate a large but gradual transition. The long-run issues become distribution, meaningful work, catastrophic risk, and what humans still own when intelligence is cheap.

  2. Broad Labor Market

    1. Autor and Thompson, "Expertise", 2025. Automation's wage and employment effects hinge on whether it strips away a job's expert or inexpert tasks, not simply on how many tasks it removes.

      MoreLess

      Using task data for 303 U.S. occupations over 1980–2018, the paper shows that automation eliminating inexpert tasks raised wages but reduced employment, while automation eliminating expert tasks lowered wages but increased employment. The mechanism is that removing tasks changes the expertise required of the remaining work, and the same task can be relatively expert in one occupation and inexpert in another. This expertise lens resolves the puzzle of why routine-task automation often cut employment yet raised wages in the affected occupations.

    2. Babina, Fedyk, He, and Hodson, "Firm Investments in Artificial Intelligence Technologies and Changes in Workforce Composition", 2023. As firms adopt AI, their workforces become more educated and technical, and their org charts flatten as middle management thins out.

      MoreLess

      Using resume and job-postings data, the authors show AI-investing firms shift toward workers with undergraduate and graduate degrees and toward STEM and IT specialization, while the share of middle-management roles falls and junior-level shares rise. Firms with higher initial shares of highly educated and STEM workers invest more in AI, suggesting AI complements skilled labor and substitutes for coordination layers. AI reshapes internal firm organization, not just headcount, flattening hierarchies as software absorbs monitoring and coordination tasks.

    3. Brynjolfsson, Li, and Raymond, "Generative AI at Work", 2023/2025. AI raises customer-support productivity mostly by lifting novices toward expert practices.

      MoreLess

      In a rollout to 5,179 agents, productivity rises 14% on average and 34% for novice and lower-skilled workers. The mechanism appears to be diffusion of tacit best practices from stronger agents, with evidence from recommendation adherence, outage periods, and convergence in communication patterns. The paper is a strong micro result, but it cannot observe wage, hiring, or aggregate labor-demand responses.

    4. Chen and Stratton, "Artificial Intelligence in the Firm", 2026. AI coding tools raise individual developer productivity, but the gains do not translate into more firm output or changed employment.

      MoreLess

      Using a proprietary dataset of about 200 million work events from 100,000 engineers at 500 firms and a staggered difference-in-differences design on Copilot/Cursor adoption, the authors estimate an 8.5% rise in coding activity per worker and an 8.7% reduction in time to task completion, with no measurable decline in code quality. Crucially, these productivity gains do not pass through to output, task composition, or employment. The disconnect suggests individual speedups get absorbed by organizational bottlenecks rather than expanding what firms produce (preliminary draft).

    5. Cruces, Fernandez Meijide, Galiani, Galvez, and Lombardi, "Does Generative AI Narrow Education-Based Productivity Gaps?", 2026. AI narrows education-based productivity gaps at the task-execution margin.

      MoreLess

      In a randomized experiment, the education gap falls from 0.548 to 0.139 standard deviations with AI access. A follow-up task without AI shows no evidence of over-reliance and some carryover for lower-education participants, but a residual education gap remains. The interpretation is task-level skill leveling, not the disappearance of human capital.

    6. Dell'Acqua et al., "Navigating the Jagged Technological Frontier", 2023. AI helps inside its frontier and can hurt outside it.

      MoreLess

      BCG consultants using GPT-4 do substantially better on tasks inside the frontier, completing more work faster and at higher quality. On a task deliberately placed outside the frontier, however, AI users are 19 percentage points less likely to reach the correct answer. The frontier is jagged, so the practical skill is knowing when to delegate, when to integrate, and when to interrogate.

    7. de Souza, "Artificial Intelligence in the Office and the Factory: Evidence from Administrative Software Registry Data", 2026. AI displaces office work but expands factory employment, and on net raises jobs by letting low-skilled workers do tasks that once required expertise.

      MoreLess

      Using Brazil's near-complete registry of commercially developed AI programs, a one-standard-deviation increase in AI exposure raises employment by about 2% immediately and up to 7% after three years. In the office, AI automates tasks, cuts employment, and hollows out the middle of the wage distribution; in production, it disproportionately boosts younger, less-educated, and lower-productivity workers operating machinery. Because production gains outweigh office losses, AI increases aggregate Brazilian employment, though the setting is a middle-income economy and may not generalize to advanced labor markets.

    8. Dillon, Jaffe, Immorlica, and Stanton, "Shifting Work Patterns with Generative AI", 2025. Individual AI access saves time but does not automatically redesign work.

      MoreLess

      In a field experiment across 66 firms, Copilot users spend about two fewer hours per week on email and less time outside regular hours. Meetings, document creation, email volume, and task composition do not move much. The distinction is important: tool access can create private time savings without changing the organization of production.

    9. Freund and Mann, "Job Transformation, Specialization, and the Labor Market Effects of AI", 2026. AI's wage effects are driven by job transformation, so they depend on a worker's full multidimensional skill portfolio rather than on occupational exposure alone.

      MoreLess

      In a general-equilibrium model with task bundling estimated on the skill distribution, projected wages are roughly unchanged at zero exposure, rise about 4% at moderate exposure, and fall by as much as 35% in the most exposed occupations, with large within-occupation dispersion. AI raises the return to social and non-routine manual skills while lowering the return to analytical skills, and the effects are mildly progressive, favoring low-wage workers. None of these results arises absent job transformation, the reallocation of time toward a job's non-automated tasks.

    10. Garicano, Li, and Wu, "Weak Bundle, Strong Bundle", 2026. The same AI capability can be substitution or augmentation depending on how easily a job can be split into tasks.

      MoreLess

      Labor markets price jobs, not isolated tasks. Weak bundles let AI peel away codifiable work, narrow the human role, and create a capacity shock in the remaining task; strong bundles keep AI inside the job and preserve more of the human revenue share. The empirical object is not just exposure, but coordination costs, accountability, shared context, and demand elasticity.

    11. Garicano and Rayo, "Training in the Age of AI", 2025. AI can break apprenticeship by automating the junior work that used to pay for training.

      MoreLess

      The key statistic is the expertise leverage ratio: AI-augmented expert value relative to AI-alone entry output. If the ratio is high enough, training remains viable; if it is too low, training compresses and can collapse once onboarding costs are included. The central question is whether AI raises expert value enough to replace the missing junior-production surplus.

    12. Gupta, Qian, Simintzi, and Sun, "Generative AI and Entrepreneurship", 2025. Generative AI reallocates the startup ecosystem, shrinking incumbent headcount while raising new-firm entry so aggregate startup employment is roughly unchanged.

      MoreLess

      Across nearly 95,000 U.S. startups founded 2018-2021, the most Gen-AI-exposed firms cut employment by about 8% after ChatGPT's release, concentrated in junior and implementation roles. Yet exposed startups raised productivity and scaled faster, decoupling growth from headcount, and venture capital shifted toward more but smaller checks (about a 12% decline in average initial funding alongside a 7% rise in deal count). Exposed sectors saw about 7% more active startups, so the net effect is reallocation rather than aggregate job loss.

    13. Hampole, Papanikolaou, Schmidt, and Seegmiller, "Artificial Intelligence and the Labor Market", 2025. AI substitutes for exposed tasks, but firm-level productivity gains largely offset the job losses, leaving net employment effects modest.

      MoreLess

      Using NLP to build task-level AI exposure measures for 2010-2023, the authors show tasks with higher mean AI exposure subsequently see lower labor demand, while exposure concentrated in a few tasks partly offsets this by letting workers reallocate effort. Identification leans on historical university hiring networks as an instrument for firm AI adoption. The headline is a composition story: displacement in exposed occupations is roughly cancelled by productivity-driven hiring at AI-adopting firms, so aggregate employment barely moves.

    14. Hitzig, Massenkoff, Lyubich, Heller, and McCrory, "Agentic Coding and Persistent Returns to Expertise", 2026. In agentic coding, domain expertise rather than coding training determines success, so AI rewards understanding the problem more than knowing how to write the code.

      MoreLess

      From ~400,000 Claude Code sessions across ~235,000 users (October 2025–April 2026), the authors find people make most planning decisions while the agent makes most execution decisions, and every major occupation completes coding tasks at nearly the same success rate as software engineers. More domain expertise raises the success rate, but the gap between expert and intermediate users is modest. Over the seven months, the debugging share of sessions fell by nearly half and the estimated value of the typical task rose about 25%, benchmarked against freelance job postings.

    15. Hosseini Maasoum and Lichtinger, "Generative AI and Occupational Entry Barriers: The Labor-Supply Channel of Technological Change", 2026. Generative AI reshapes wages not only by raising productivity but by lowering the expertise needed to enter occupations, expanding the pool of potential workers.

      MoreLess

      The authors build measures of AI-induced potential supply shifts and productivity gains from O*NET tasks and embed them in a general-equilibrium "productivity-scarcity race." Occupational productivity gains widen between-occupation wage inequality, while reductions in expertise requirements narrow it. Empirically, after adoption junior employment falls in adopting firms relative to non-adopters while senior employment is largely unchanged, an effect concentrated in the most exposed occupations and driven mainly by slower hiring rather than more separations.

    16. Humlum and Vestergaard, "Still Waters, Rapid Currents", 2025/2026. AI reorganizes tasks before it shows up in wages or hours.

      MoreLess

      Danish linked survey and administrative data show adoption, time savings, and new AI-related tasks in content generation, oversight, and integration. But earnings and recorded hours show precise null effects two years after ChatGPT. The early effect is beneath the surface of standard labor statistics: work changes before wages and hours do.

    17. Noy and Zhang, "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence", 2023. ChatGPT makes short professional writing faster and better, but mostly by substituting for effort.

      MoreLess

      Completion time falls by about 37% and quality rises by about 0.45 standard deviations. Lower-baseline writers benefit more, compressing the productivity distribution. Many treated participants submit lightly edited AI output, so the experiment shows immediate speed and quality gains more than deep human-machine complementarity.

  3. Financial Sector

    1. Babina, Fedyk, He, and Hodson, "Artificial Intelligence, Firm Growth, and Product Innovation", 2024. Firms that invested more in AI grew faster, and the growth came through new products rather than cost-cutting, concentrating gains among already-large firms.

      MoreLess

      Building a resume-based measure of firm-level AI investment covering roughly 64% of the U.S. workforce, the authors find a one-standard-deviation increase in AI investment is associated with about 19.5% higher sales, 18.1% higher employment, and 22.3% higher market valuation over 2010–2018. The mechanism is product innovation (more and better products), not improved operating efficiency, and the gains accrue disproportionately to large firms, raising industry concentration. Early-era AI acted as a growth technology for incumbents rather than a broad-based productivity leveler.

    2. Bartlett, Morse, Stanton, and Wallace, "Consumer-Lending Discrimination in the FinTech Era", 2022. Algorithmic FinTech lenders discriminate against minority borrowers less than face-to-face lenders do, but they do not eliminate it.

      MoreLess

      Studying GSE-securitized and FHA-insured mortgages, the authors find risk-equivalent Latinx and Black borrowers pay higher interest rates, costing minority borrowers over $450 million per year, concentrated in high-minority-share neighborhoods. FinTech algorithms also discriminate, but roughly 40% less than traditional face-to-face lenders, consistent with algorithms stripping out some taste-based pricing while retaining pricing tied to shopping behavior. Automating credit decisions reduces but does not resolve disparate impact, and can still price-discriminate through non-race variables correlated with group membership.

    3. Cao, Jiang, Wang, and Yang, "From Man vs. Machine to Man + Machine: The Art and AI of Stock Analyses", 2024. An AI analyst beats most human analysts at forecasting stock prices, but the best forecasts come from humans and machines together.

      MoreLess

      Trained on disclosures, industry trends, and macro data over 2001–2016, the AI analyst outperforms roughly 53.7% of individual analyst target-price forecasts, and trading on AI-versus-human disagreement earns monthly risk-adjusted returns of about 0.84%–0.92%. Combining the two, the "man + machine" model outperforms 57.3% of analyst forecasts and beats the AI-only model, with humans adding the most value for firms with intangible assets or financial distress where institutional knowledge matters. The result reframes AI in sell-side research as a complement that reduces extreme errors rather than a wholesale replacement.

    4. Eisfeldt, Schubert, and Zhang, "Generative AI and Firm Values", 2023/2024. Markets initially priced AI exposure as a labor-cost reduction for firms.

      MoreLess

      High-exposure firms rose about 5% relative to low-exposure firms after ChatGPT's release, especially when they had valuable data assets. The labor evidence points toward substitution: exposed occupations see lower job postings and lower relative wages, particularly when exposure is concentrated in core tasks rather than supplemental ones. The market reaction looks less like generic optimism and more like expected cost savings.

    5. Fedyk, Hodson, Khimich, and Fedyk, "Is Artificial Intelligence Improving the Audit Process?", 2022. Audit firms that invested in AI produced higher-quality audits at lower fees, but eventually displaced human auditors.

      MoreLess

      Using more than 310,000 individual resumes from the 36 largest audit firms, the authors find a one-standard-deviation increase in recent AI investment is associated with a 5.0% lower likelihood of a subsequent audit restatement, alongside reduced fees, with labor displacement materializing only after several years. Effects are stronger for audits of older, data-rich clients and in the retail industry, consistent with AI adding most value where large structured datasets exist. AI in professional services is quality-enhancing and cost-reducing in the short run but labor-substituting over a longer horizon.

    6. Fuster, Goldsmith-Pinkham, Ramadorai, and Walther, "Predictably Unequal? The Effects of Machine Learning on Credit Markets", 2022. Switching mortgage underwriting from traditional statistics to machine learning helps most borrowers but widens interest-rate disparities across racial groups.

      MoreLess

      Comparing default predictions from traditional and machine-learning models on U.S. mortgage data, the authors find Black and Hispanic borrowers are disproportionately less likely to benefit, even as the flexible model slightly expands overall credit access. The widening of rate disparities is driven by the model’s greater functional flexibility (its ability to detect within-group risk differences), not by algorithms proxying for race directly. Fairer inputs do not guarantee fairer outcomes: statistical flexibility itself can redistribute who pays more, complicating "debias-the-data" fixes.

    7. Huang, Hugon, Zhang, and Zheng, "Generative AI and Investment Research: Evidence from Analyst Reports", 2026. AI-assisted equity research is more accurate and more market-moving, with the largest gains for overloaded and lower-skilled analysts.

      MoreLess

      In a large sample of analyst reports, AI-assisted content is associated with greater earnings-forecast accuracy, and the benefit is concentrated among analysts facing heavier research demands and those with lower baseline skill. Using an exogenous workload shock from overlapping earnings calls for identification, the authors find AI usage causally drives the accuracy improvement. Forecasts from AI-assisted reports also trigger stronger capital-market reactions, suggesting investors treat this research as more informative, though AI adoption is inferred rather than directly observed.

    8. Sheng, Sun, Yang, and Zhang, "Generative AI and Asset Management", 2026. Hedge funds that adopted generative AI outearned non-adopters, concentrating the gains among the already-sophisticated.

      MoreLess

      Adoption jumped sharply after ChatGPT's 2022 launch, reaching roughly 60% of hedge funds by 2024, and a difference-in-differences design shows adopters earn 2-4% higher annualized abnormal returns while non-hedge funds see no benefit. The edge traces to funds' AI talent and ChatGPT's strength at parsing firm-specific information; GenAI usage improves price efficiency but initially widens information asymmetry. The pattern points to AI widening rather than narrowing competitive gaps in asset management.

Resources

  1. Workflow

    1. Aniket Panjwani, Getting Started with Claude Code (video). Practical walkthrough for using Claude Code.

    2. Getting Started with Codex (video). Practical walkthrough for using Codex.

    3. Adrien Matray, AI Guide for Economists. Practical guide to building reliable AI research workflows, from mindset and setup to verification, project structure, skills, and large datasets.

    4. Claes Bäckman, Claude Code Guide for Academic Economists. VS Code setup, Stata/R/Python/LaTeX extensions, CLAUDE.md conventions, and file-format handling.

    5. Xingjian Zhang, AI Coding Workshop. Workshop slides on AI-assisted coding practices.

    6. Gonzalo Ballestero, Practitioner Guide to Agentic AI. Lecture notes on agentic AI for academic research.

    7. Moran Koren, Theorist Toolbox. Claude Code skills for machine-assisted economic theory: gap-free proof writing, adversarial verification via Codex, and peer-review-gated proof projects.

  2. Reference & Learning

    1. Cathryn Lavery, The Non-Technical Technical Dictionary. 100 tech terms explained through plain-English analogies, for non-technical people building with AI.

  3. Landscape & Background

    1. Kevin Bryan, The Landscape of AI, 2026. Slide deck on AI capabilities, productivity, organizational frictions, and why technical progress does not automatically translate into growth.

    2. Suhas Pai / Hudson Labs, The AI Ecosystem Stack — Market Map, 2026. Market map of public and private companies across the AI stack, useful for tracking infrastructure layers, application layers, and current bottlenecks.

  4. Essays

    1. Leopold Aschenbrenner, Situational Awareness: The Decade Ahead. Essay on the strategic implications of frontier AI progress.

    2. Leaked internal Google memo, "Google 'We Have No Moat, And Neither Does OpenAI'", 2023. On open-source AI, fast iteration, low-cost fine-tuning, and why model quality alone may not create a durable moat.

  5. Prescient Reads

    1. Fred Brooks, The Mythical Man-Month, 1975. Coordination costs survive even when labor becomes abundant.

    2. Everett Rogers, Diffusion of Innovations, 1962. New technologies spread through social systems, not technical merit alone.

    3. Kurt Vonnegut, Player Piano, 1952. Automation can solve production while hollowing out work and status.

    4. Charles Petzold, Code, 1999. Interfaces change, but computers remain layered machines underneath.

    5. Joel Spolsky, The Law of Leaky Abstractions, 2002. Every abstraction eventually reveals the machinery below.