Igor Letina

In *Startup Acquisitions: Acquihires and Talent Hoarding,* Jean-Michel Benkert (@jm_benkert), Shuo Liu, and I study acquihires, a type of startup acquisition where the acquirer is mainly (or only) interested in people working for the startup. PDF: igorletina.com/files/Talent_H…

A day like today, exactly 25 years ago, my admired advisor Sherwin Rosen died, way too young, at 62 years of age. He was then the president of the AEA. We owe him some crucial ideas. I highlight 7. Hedonic Prices and Implicit Markets (JPE, 1974). How does the market price something as complex as a car, a house, or a job? Goods are bundles of characteristics. In equilibrium, the price schedule is the envelope of heterogeneous buyers' bids and heterogeneous sellers' offers—so market prices reveal the implicit value of each characteristic. Key to environmental valuation, the value of life, and urban quality-of-life indexes. Monopoly and Product Quality (JET, 1978, with Mussa). Did you wonder if your tourist class seat is too narrow? A monopolist who faces buyers differing in taste for quality degrades what she sells to low types so high types can't mimic them and capture the surplus. The foundational screening model. Education and Self-Selection (JPE, 1979, with Willis). Do grads from better colleges earn more because of how much they know? People sort into college by comparative advantage: those who go are better at college-type work, those who don't are better at non-college work. The returns need to be corrected. A crucial idea for an entire literature. The Economics of Superstars (AER, 1981). Why do rewards concentrate at the top in music, movies, sports etc.? When output can be replicated at zero marginal cost, and there is little substitutability in production, small talent differences produce enormous earnings gaps. The economics of the internet, twenty years early. Rank-Order Tournaments (JPE, 1981, with Lazear). When individual output is noisy, firms pay on rank; the spread between winner and loser is the instrument that elicits effort. Authority, Control, and the Distribution of Earnings (Bell Journal, 1982). In a hierarchy, each manager's talent is multiplied across everyone below her. A slightly better person at the top is worth disproportionately more- a better general decides which war we fight, hence affects all of our marginal products. That is why we see convexity of pay at the top of organizations. Prizes in Elimination Tournaments (AER, 1986). In a multi-round promotion ladder, the biggest jump must come in the final round, because the option value of future rounds has vanished, only the current prize can motivate. Professor Rosen would look distracted in seminars. He would look confused. Then he'd say something that changed the entire analysis and discussion. He never tried to look good at the speaker's expense. He just saw the problem more deeply than anyone in the room- no exceptions. One personal anecdote: during my PhD studies, I was totally depressed: I could not advance, all my ideas were awful. I could not bear going to his office. As i was coming upstairs towards the 4th floor of the Social Science building I met him in the stairs. He said. "I have not seen you, Luis, for a while." I said "Sorry Prof. Rosen, I had nothing to show you." He said "Well, you have to come. Come every week, whether you have something or not". incredibly, that short exchange was probably the most important one in my life. The duty to go to his room got me out of the hole.

Very happy to be part of this OECD workshop on start-ups’ ability to scale. If this topic interests you, take a look at the website to see the full speaker lineup and consider joining us online or in person. oecd.org/en/events/2026…

@varma_ashwin97 @PatrickMcGee_ How come?

Reading “Apple in China” (@PatrickMcGee_ ) will count as one of the most politically important and radicalizing moments of my life.

@Afinetheorem It is possible that I am at fault. But I am certainly trying to use agents .md file.

@IgorLetina Interesting - I'd never had this happen before with CC.

1/7 I have spent a month testing ChatGPT, Claude, and Gemini ($20 subscriptions) in parallel. Here is my experience, in case someone finds it useful. I am not an expert user, so some of this may reflect my (suboptimal) usage. But maybe that will reflect your experience as well.

@Afinetheorem I use that, and I still find that Claude Code regularly ignores the instructions. This is what happened to me literally 10 minutes ago: "You're right, I violated the AGENTS. md rule. The datasets/ folder is read-only. Let me fix this."

@IgorLetina With Claude Code (and any of these agentic models), there is a HUGE benefit from a claude .md or agents .md that points to other files with your coding, style, writing preferences.

@s_lauermann Thanks! Your experience with GPT 5.2 pro matches mine. However, I have gotten used to switching between thinking and pro, as the pro usage is quite expensive on my plan ($25/seat team plan). I was surprised by the low quality of literature reviews all three models produce.

Great post. Too little of such information out there. Generally, my experience is that the difference for math problems between the top models and the $20 versions is quite high. Was that yours, too? GPT 5.2 pro would solve all exercise of intermediate micro without any errors, e.g., and write beautiful solutions with figures of game trees and everything on top.

@abdurrahmanregi @OpenRouter I have an OpenAI business plan for $25 per seat, which includes 5.2 pro usage (limited budget though).

@IgorLetina I'd suggest @OpenRouter, there, you can also try gpt 5.2 pro (via API), where -iirc- this model is not available in the $20 subscription. Would be very interesting to know the opinion of another IO theorist on gpt 5.2 pro. However, Opus via API is super expensive for daily usage

7/7 For example, when asked to recommend local LLM models, ChatGPT would regularly recommend models from 2024, whereas Gemini would recommend the most recent variants. In the end, I will be keeping ChatGPT and Claude and I will be upgrading Claude if my workload requires it.

6/7 Gemini: I ended up using it the least. Gemini CLI is OK, but I did not find any real advantages over Codex or Claude Code. One big advantage of Gemini over ChatGPT is that its answers are based on more recent data.

This expresses the vague thoughts I've been having as well. A corollary is that some people are developing skills that will be useful for only a brief period of time.

Economic research from the University of Bern getting some well-deserved attention:

@FlorinBilbiie @lugaricano Same!

I used to be so proud of my em dashes—and so sad to have to let them go now.

@SanchezCartas Oh, I'm sure it's by design, but I don't think their objective is to waste tokens. They built Claude with a "move fast and break things" mindset. Codex, in my experience, has been much more restrained.

@IgorLetina Maybe it's by design. A simple instruction to raise "temperature" above users' configuration. The worst case scenario is you wasting tokens, the best case scenario for them is you wasting tokens.

Something tells me Claude is not actually sorry.

@oliverbeige I disagree. Graduate students are actually great.

@IgorLetina The whole thing about "PhD level intelligence" is starting to turn into "first year graduate student sloppiness plus the usual bullshit excuses"...