Saturday links
I’m getting off to a bit of a late start today because this morning I started my workshop on causal inference at Mixtape Sessions. I wanted to wake up early, but did not wake up early, and therefore did not get to this. But I wanted to try to now.
First, I am using otter.ai for the first time in the workshop and that’s been a little buggy. But I’m going to keep at it. Otter.ai recall sits on your zoom meetings and records the entire thing. Then later you have a transcription as well as a small little LLM to talk to about the call, who will also give you a summary of things to do. I’ll get the hang of it though.
This article discusses a new type of neural network architecture known as Kolmogorov-Arnold Networks (KANs). It caught my eye because of Kolmogorov who I knew from statistics. KANs replace the typical weighted edges in neural networks with learnable functions, allowing them to represent more complex relationships. That’s as far as I got.
Lots of people are saying that the new Strawberry model is very good at deception and other forms of risk maybe too. It is true that OpenAI rated it as “medium risk” on its Defcon security scale, and they've sworn to never make anything that went above that. So that’s interesting.
I thought this substack by Timothy Less on Strawberry was helpful. He notes, about halfway down, that he is finding that it is worse at spatial reasoning that GPT-4o. His examples come from simple chess play with the model, and it cannot get even basic things correct, even while GPT-4o can. He thinks the reason is that GPT-4o has so many chess games in its training data that the pattern recognition is doing something, but the reasoning can’t reason in spatial dimensions. Something about the tokens being in one dimension.
Lance Eliot also wrote about the Chain of Thought reasoning that Strawberry is doing. He writes a ton about the new consumer LLM models, actually, and I hope I can remember to keep checking him out as his writing is accessible for a lay person like myself .
I updated my students’ AI homework assignment, due every other week for this history of thought class I’m doing, when Strawberry came out. I wanted them to use Strawberry alongside ChatGPT-4o to summarize and critique the classical economists, before then writing their own essay. I thought it might be good to have them see both LLMs taking a stab at critical writing as well as providing references, and that has been a really positive experience for them I can tell. They have been surprised by the different quality of the final product, but they’re also getting something out of the chain of thought it produces (Strawberry I mean). They seem to typically see improvements, and then by the time they write their own essay, they’ve read two LLMs interacting with the text, which they have to compare with one another and with the original text, and then write their own. It’s doing what I want which is that the hope is through that circle of reading and evaluating they’ll start forming their own opinion about the text itself, what it means, and what they do or do not agree with the initial summary.
I encourage you to upload an article to Google’s new notebookLM and just flip the notebook guide at the bottom to make you an audio summary. It makes a podcast of two people talking intelligently about the essay. It’s pretty mind blowing. I was absolutely stunned speechless to be honest. It was the first time yet as an educator I thought “well, I should probably start looking for a backup plan as a teacher”.
New article looks at whether LLMs can produce “novel research ideas”. TL;DR - yes.
New paper on marriage market sorting over the last 50 to 60 some odd years. Includes discussions of matching models, demographic trends, and descriptive analysis. Also discusses the role that online dating has played.
Here’s why super rich people still get mortgages.
Here’s someone predicting the nail in the coffin for coders with the new o1-preview, but I personally think this is pretty premature by orders of magnitude to be calling for things like that. The thing to keep in mind is isoquants typically do not have corner solutions except for perfect substitutes, and I just have a hard time seeing that. We can have adjustments across the isoquant to a place with a different marginal rate of technical substitution and not go to literally zero workers in a field. Also, not to mention that I got locked out of o1-preview for two full days solely because I kept asking it to help me design a recruiting event for Baylor’s MSECO program. Which by the way, you should absolutely apply to come to get a masters in economics, but I’ll explain the value proposition of our program another time once I can get back into Strawberry.
Speaking of the effect of AI on jobs, a new paper by a team from the Census just published about it in Econ Letters. Headline findings are here:
We find that about 27% of firms that use AI are using it to perform tasks previously done by workers, and this fraction is expected to grow substantially (to nearly 35%) within the near future. However, the impact on employment is modest with only about 5% of firms changing employment levels. Moreover, a slightly higher fraction report an employment increase rather than a decrease. The fraction of businesses changing employment due to AI is expected to rise to nearly 12% in the near future, with an expected increase in employment being slightly more common than a decrease
Pretty short-run effects to keep in mind. This paper is based on a survey conducted in 2023 and 2024. You be the judge as to whether 27% of firms using AI are substituting away from workers after only this short period of time or not. That kind of seems large to me, though, and suggests rapid adjustments are being made already.
This study really caught my eye though. Researchers found that exposure to an LLM chatbot caused believe in conspiracy theories to fall. The article was just published in Science. And if you want to try the LLM yourself, you can do so here.
Be careful messing around too much with Strawberry. If you try to figure out what is under the hood, you might get banned.
Alex Bartik and coauthors use LLMs in economic research studying the costs of housing regulations. Remember — you are limited by your own imagination in how you use LLMs for research going forward. If you just use it to write paragraphs or code, then that’s all you’ll get out of it.
Danielle Li, Lindsey Raymond and Peter Berman have a working paper (I think it’s still working paper anyway) framing hiring as balancing the need to exploit known successful candidate groups and explore under-represented ones to improve hiring quality and diversity. They call this a contextual bandit problem. They build an algorithm that evaluates candidates’ statistical upside, and examine how well this approach can enhance both hiring rates and demographic diversity, unlike traditional supervised learning algorithms, which improve hiring rates but result in fewer Black and Hispanic applicants being selected. I definitely want to read this closely.
Thanks for tuning in! Now back to the second half of day one of my workshop.