It’s happening!

But how many would they collect in Nashville?

Very disappointed. Cedric needs to be shut down.

100% no. No no no

Also… ouch…

Interesting and depressing finding from @danielmangrum.com and colleague, but not surprising at all.

What a headline… oof

We call her Peggy Hill

I really expected the river to still be a bit green though...

Wheels up soon to #AEFP2026. I’ll be chairing a session in the first block tomorrow and presenting in the second. If you’re interested in higher ed finance, I’ll see you at both!

*”fine to use” TBD by data vendor of course

You can install a local version of established models that run siloed without the ability to “phone home.” The inputs don’t get fed anywhere for training or leave the local install. They’re just evaluated. So they are fine to use with sensitive and confidential data.

Underrated massive improvement from LLMs on research productivity? Documentation! Everything of mine is vastly more organized now. Within and across projects and workflows. Maybe that’s just a tell of how bad it was for me before but a huge improvement to have organized digital headspace.

Came to say the same thing. IMO its a narrow view on the current dominant mode people use LLMs but these good models are going to be local installs with no training very soon and that alleviates a lot of concerns.

I’ve found performance to be miserable with pdfs. I just point it to the tex file and have it read the input tex/images that the tex links to.

Ignoring spatial spillovers doesn’t just attenuate estimates, it also undercounts scale. “Control” states are already betting and experiencing credit distress. Methodologically, this matters for any staggered DD where state policies spill across borders (marijuana, minimum wage, firearms, etc.).

Spillovers create a fiscal asymmetry: states that haven’t legalized bear costs from spillover betting without capturing tax revenue. We simulate legalization for not yet legal states. Missouri (with KC and STL sitting on legal-state borders) gets the highest revenue-per-delinquency ratio.

Using the NY Fed Consumer Credit Panel, we find legalization raises overall delinquency by 0.31 pp. For under-40s, credit card delinquency rises ~1 pp and auto loan delinquency rises 0.55 pp. Scaling by the 3.1% take-up rate implies ~10 pp overall delinquency increases among induced bettors.

Residents of non-legal states near legal borders cross state lines to bet on their phones. Counties within 15 miles of a legal state see spillover spending ~14% of the direct effect. The effect fades nearly linearly with distance and hits roughly zero by 60 miles.

Since the 2018 Murphy v. NCAA decision, 38 states have legalized mobile sports betting. Using transaction data covering 80M+ accounts, we find legalization increases sportsbook spending ~10x and raises betting participation by 3.1 pp.

But the effects don’t stop at state lines.

Staff Reports Technical working papers

🧵 Jake Goss and I have a new paper where we study how legalized mobile sports betting affects consumer credit and how betting spills across state lines into places where it’s still illegal. newyorkfed.org/research/staff_reports/sr1184

“Create a baby boom with this one simple trick!”

This also holds with code/workflows too. The quality of the initial development is okay but it improves dramatically when you have independent agents critique and improve the initial workflow.

Yeah in my experience, the best way to get quality output is to have an independent agent review the output of initial agents. You don’t really good quality output until you have reviewer tasks.

For reference, I’ve used variants of this style prompt to produce simulated reports/editor decisions on my own papers and the quality is generally pretty good. The key is to not have any agent review its own work. Shocker: agents are quite happy with the work they produced.

Mind sharing the prompts (even over DM)? I’ve found that performance with tasks with opus are substantially better if you assign independent tasks to follow up and critique each other. An agent rarely likes to critique itself but parallel agents are way better at finding flaws in other agents’ work

I previously thought that everyone was turning hyper productive using Claude Code but then I read the quotes and replies to this and apparently that’s not the case.

New research alert!

Are smartphones changing knowledge measures on surveys?

tl;dr YES!

This is work with my awesome coauthors Olivia Valdes and Jeremy Burke--probably one of the most fun papers I've been a part of.

Ken Walker got them in range so he could kick any FGs and kept him from kicking any long ones. Without those 3rd down conversions and long runs, Meyers is maybe 2 for 3 in a loss.

Maybe if he kicked just one from 50 plus there would be a case. Every kick was a kick that would be automatic from the median nfl kicker