Frequent convo between me and PhD students

Me: "After 400 hours of data analysis on industry X, have you talked to anyone actually working in industry X"
Student (S): "No"
Me: "Why not?"
Student: "Where would I find someone in industry X willing to talk to me" Me: "Have you tried asking your advisor to put you in touch with someone in industry X"
S: "No"
Me: "Have you tried finding the subreddit/forum/Twitter where these people hang out to at least hear them chattering about stuff"
S: "No"

Heterodox economists believe that modern economics is flawed because it restricts itself to Newtonian mathematics. However, this reflects a flawed and outdated understanding of modern economics. Here are some examples of post-Newtonian mathematics used in modern economics

https://twitter.com/IsabellaMWeber/status/1971292262874370221

Rather than restricting to classic calculus-based approaches, or even simpler linear-system solutions, modern economics uses tropical geometry to analyze existence of equilibria in indivisible-goods economies

Nightclubs, luxury watches and handbags, and Labubu toys are social goods. People buy them because others buy them. Economic theory predicts that demand for social goods is fragile. In a revised paper, for the first time ever, we've found empirical evidence of this fragility! In this paper with @sammy_rosen and Sebeom Oh, we find a natural experiment "disrupting" initial NFT sales. Suppose a collection launches, and BTC prices happen to move a lot. The mint is less likely to succeed, because NFT buyers are distracted managing their crypto portfolios

I'm an academic economist and I've been bullish on crypto for a few years. The thesis is simple. Crypto is a tool almost exclusively for criminals: it's basically useless to anyone in the developed world. But it's a really good tool for criminals Crypto does, however, link to a very fundamental question: finance, as we know it, is a derivative of legal systems. Finance is the trading of promises, and promises only work when people think they'll be enforced

I feel like a substantial amount of creativity comes literally just from knowing a lot of stuff. If you know like, three paintings, history looks like a linear progression from painting 1 to 2 to 3. If you see like 3 million paintings, you have a richer understanding of history You see not only the famous things, but also all the smaller ones not taken, that could have been famous, but somehow missed the bus, you draw connections that haven't been drawn, revive forgotten stuff, and so on

Conjecture: econ academia is very far from the optimal architecture for data research. The optimal architecture is much more centralized. Half the researchers work together, or in large teams, to build clean and efficient datasets. The rest work in smaller teams to analyze data All empirical research is based on data. In the current architecture, building and maintaining datasets is under-rewarded. Some of this work is done by government workers, who are generally (and increasingly perhaps) underpaid and under-appreciated

IMO, a large driver of how productive econ/finance PhD students are is work habits. Basically, can you set and maintain realistic goals that interpolate between having nothing, and having an R&R'd or published paper. This is not taught, and it's surprisingly hard The issue is, if you go directly from school to PhD, maybe with an RA-ship in between, you're used to having tasks handed to you in like 1-2 week increments: psets, exams, RA work, so on. You've never really had more autonomy than this

The thing I'm realizing is lots of people don't get how revolutionary the o3/o4 class of models are, for (econ) research. 4o and the like models are borderline useless for econ research-level math. Deepseek is borderline useful. o3/o4 are at the level of solid PhD students o3, o4 represent a step change in how I use them. Prior to o3/o4, I used them occasionally for tasks here and then. After o3/o4, I use them in general as basically the first thing I go to, for more or less everything research related I'm doing: coding, model-solving, writing

Putting kids through too many status games makes them risk-averse. They go through middle school, high school, college, every step doing the "right" thing. They go into a world with a deep fear of ever accidentally stepping off the "golden road" They see peers, who they see as "less successful" than them - less badges and accolades - make choices they would never have made; and judge them, and think how horrific it would be to ever fall so low in life

I'm reading Holmstrom 1979, and I'm a lazy student who doesn't want to actually do the work, so I'm making ChatGPT do it for me. It is very simple, I just give o4-mini screenshots of parts and ask, wtf is going on. It tells me exactly what Holmstrom is saying, much more clearly
I can give it literally the laziest, most notation-insensitive questions of "wtf is going on" and it just puts up with my shit. My TA would kill me for even asking a question this way over email. But, until the AI overlords come to punish me for my sins, o4 just cooperates

There is a very cool English translation of an 1884 Russian middle school geometry textbook, ebook available for $10 online. The bibliography is 23 books, going back to 624BC, and includes Pythagoras, Hippocrates, Plato, Euclid, Archimedes, Fermat, Newton, Euler, Gauss, Maxwell
Starts elegantly and intuitively written:

"The idea of the plane is conveyed by a window pane, or the water surface in a quiet pond"

"The image of a thin thread stretched tight or a ray of light emitted through a small hole give an idea of what a straight line is."

Do stock prices behave weirdly around "round numbers"? Are technical analysis things like "support" and "resistance" real?

This cool paper says yes: stocks priced just above round numbers, like $6.1, tend to go up; stocks just below round numbers, like $5.9, tend to go down! We always tend to make fun of technical analysis, so it's pretty cool to have some evidence of some tech analysis stuff actually working. This plot shows that x.0 stocks do better than x.9 stocks

One of my favorite pieces of academic writing is Holmstrom on the role of debt in the financial system Motivation: 2008! Nobody knew how bad all this stuff was! That's terrible! Michael Lewis and the Big Shot! Markets need transparency! Why were these things so opaque? Was it all an attempt to trick investors?

idk about Nvidia but inference-time compute seems like a marginal cost business, not fixed cost

The whole OpenAI bet, IMO, was that scaling laws on training implied that you could have a moat by building the godmodel if you just bought an insane amount of compute early If the scaling is basically on using the model, rather than building it, the model seems more commoditized; anyone can use a half decent model and just make it think harder

Math puzzle (cleaner statement)

- You're in a jail cell with two doors, A and B
- There's 3 keys; the A door key, the B door key, and a dud key
- Design an optimal key-trying strategy: a strategy which gets you out of jail in the minimum expected trials of key-door pairs The key question is: after trying key 1 on door A, do you switch keys, or switch doors? That is, do you try:

1. The same key on door B, or,
2. A different key on door A?

I think 2. works better: you want to try all keys on one door

Math puzzle, inspired by Mario Party Jamboree:

There are 2 doors, A and B. There are 3 keys, 1, 2, 3. 2 of the 3 keys unlock the doors, the other is useless

You try key 1 on door A and it doesn't work. What's your next play?

- Try key 2 or 3 on door A
- Try key 1 on door B Of course, next step is generalize to N doors and M>N keys

A simple and crucial set of "data science" skills, which is taught very poorly in class, is data visualization and exploratory data analysis. Inexperienced RAs make two common mistakes:

- Don't make enough graphs
- Make very bad graphs, like these two
This plot is terrible because it says nothing. It says some observations are small, and a few are very big, and you learn nothing else. It violates a core principle of dataviz: aim for high data density per unit of screen area. 99% of your screen is empty!

What we call "economics" increasingly refers to a bunch of policy-focused quantitative social sciences:

- Economics
- Finance
- Accounting
- Marketing
- Public policy
- Health policy

As well as maybe parts of:

- Political science
- (?) Education policy

I may have missed some I characterize these as a more or less cohesive bunch because there's meaningful "cross-talk" across these fields

- People with PhDs in field A get jobs in field B
- People publish in field A and field B journals
- People in fields A and B collaborate

The humanities people haven't kept up with the historical development of economics as a field. We got mathematicized in the like early mid 1900's, went super "neoliberal" in the 80's, and more recently figured out we could just take over all of empirical social science A lot of modern empirical economics is just empirical analysis of random policies. It has very little to do with markets and less to do with Marx. There's no particular reason why the policy eval folks would learn anything from reading Marx

Guys, what if the lazy kids reading Sparknotes actually had the better, more efficient learning strategy, and the smart kids were wasting their time blindly reading the full texts like the teachers told them to The full text readers do better in life because of selection bias. The kids who listen to the teacher and work hard, also work hard on other stuff, and hard work pays off in the economy. Reading full texts instead of sparknotes causally contributed little to their later success

The reason some see economics as too broad nowadays is that economics basically expanded to become quantitative social science. Economists think of themselves as the best-qualified group of data people in social science and see their toolkit as basically applicable to anything We face arguably a little competition from psychology PhDs, but other than that the majority of quantitative/data oriented social scientists with large thought market share are functionally economists (i.e. have econ or econ-related PhDs)