My favorite graph from our paper: the Cone of Adoption.

Lots of exciting work by @davidautor.bsky.social @avigoldfarb.bsky.social @lukasfreund.bsky.social @joshgans.bsky.social and many others, capturing the "directional" nature of AI.

We think job flexibility is key and best modelled as "collaboration" not "automation."

4. AI adoption depends on both worker and job characteristics. The same worker may use AI in one job but not another. Think academia vs consulting vs tech firm.
5. Adoption is faster when jobs and workers are more flexible, which is the case for knowledge work.

Relation to lit:

This simple insight leads to a lot of predictions.

1. Absolute advantage not enough for AI adoption. AI needs a comparative advantage relative to worker's current tasks.
2. No adoption up to some threshold, quick adoption afterwards.
3. Some workers only partially adopt.

Coworkers don't change the production function. They add to the set of outcomes that are feasible.

This formulation is unusual in labor, so we borrow Production Possibility Sets from international trade.

Worker + AI = convex hull of two PPSs.

GenAI is not a machine and shouldn't be modelled as such. Given the range of tasks it can do if properly supervised, it is more like a coworker.

We explore the economics of GenAI adoption among knowledge workers in a new paper:

koren.mk/publications...

As an economist who grew up under socialism, I agree.

Vinyl skills:

1. What is vinyl?
2. It has two sides.
3. Can place the needle right before the bridge of Karma Police.

Indeed, (almost) every regression is a GROUP BY query. Amazing job @gmcd.bsky.social

We can debate the magnitude but I don't see how you can say vibe coding does not weaken user engagement. If Claude (etc) selects and imports a library instead of a human user, how is that user engaged with the developer / user community of the library?

Vibe coding may be hazardous to open source Researchers argue AI coding tools disrupt community and hinder returns to maintainers Tailwind Labs CEO Adam Wathan recently blamed AI for forcing him to lay off three workers.…

Vibe coding may be hazardous to open source

Our paper with @gaborbekes.bsky.social @julianhinz.bsky.social and Aaron Lohmann is/was on top of HN. 15 minutes, I guess.

My favorite German veterinarian jargon:

tier spezifisches Eigengeruch

I keep repeating it to my dog.

Dependency

Also, I don’t see the harm the title could do even if picked up by a journalist who didn’t read the paper.

Open source has a brittle business model. It grew fast but it can also collapse fast.

People should be aware of that. Don’t take the “random person in Nebraska” for granted.

xkcd.com/2347/

2. I very much favor @saywhatyoufound.bsky.social titles (doesn’t show in my pubs i know). This is not a speculation. This is not a pun. Not clickbait. This is what we found.

Watering it down to “Quality Selection and Entry into Open Source in the Age of AI” would have been a disservice to readers

The Melitz model performs very well in predicting selection into exporting, and how competition induces exit.

We need more work on the empirical calibration, but the key assumption theta > sigma holds for sure.

If we want to treat the model seriously, though, it has to be put together seriously. Informed by empirical evidence, institutional details, understanding of existing similar models and their performance.

Our model is based on the Melitz model. We realized it captures the relevant mechanisms best

1. Economic theory, especially general equilibrium models, provide useful predictions for changes that are too large and systemic to be studied in a small experiment.

Theory is not a way to capture one particular mechanism with math. It is a coherent way of thinking about how the mechs interact.

Good discussion, thank you! You’re raising two points:

1. how much can we learn from an economic model
2. what is our responsibility in giving paper titles

I agree these are important and non trivial questions. We have grappled with them both.

My views, may not be fully shared by coauthors:

@gaborbekes.bsky.social would be very interesting to try at @ceu-economics.bsky.social

There will be almost no need to actually type syntactically correct source code. But a lot of need to think like a software engineer.

But: How do you teach programming without using programming languages?

Would be curious to see a vibe coding first curriculum.

Yes, and congestion, too. When a lot of contributions are vibed, it is hard to do a good job on all task.

Fair points. The model has dev cost going down due to AI, the result is that in the race between that and the demand collapse, the latter wins.

We also speculate on what alternative business models can help.

Vibe Coding Kills Open Source Unless We Make Drastical Changes would be accurate, though.

All audiophiles are rich, but not all rich are audiophiles.

Vibe coding kills open source.
Our most direct title yet. @koren.mk @julianhi.nz @aaron-lohmann.bsky.social
Theory paper with numbers and policy recs. First at arxiv.org/abs/2601.15494
Comments welcome.

@ceu-economics.bsky.social @kiel.institute

It’s not click bait if it’s true.

The only condition needed is that ai-assisted usage and direct usage of the *same* software package are closer substitutes than different packages. We are sure this holds (though plan to do more data work to confirm), that’s why we opted for the strong title.

GitHub - codedthinking/eventbaseline Contribute to codedthinking/eventbaseline development by creating an account on GitHub.

github.com/codedthinkin... and www.jonathandroth.com/assets/files...

Words you only hear in a vibe coding session:

"Now let me deploy and test"

A taylor’s measuring tape cut at 148cm

Hungarians will understand what this is for.

It ended being a rave party, but hey, had fun anyway.