Leaving the Basque Country, Almost Home

Plus a bunch of Saturday links, videos and pictures of me and my daughter

We left beautiful San Sebastián Friday morning by train bound for Madrid because Saturday morning, my daughter and I catch our respective flights back to the United States — she to Texas, and me to Chicago, where I’ll be for the next 11 days helping host the speakers and guests at Northwesterns back to back causal inference workshops. San Sebastián was paradise. It’s a small city on a Spanish coastline with a beautiful beach and wonderful people. I would love to come back next year when my youngest graduates high school and spend two weeks with her there just like I did with her older sister.

Obviously it is bittersweet to be going back. I’ve been in Europe for two months. Madrid for a week, Italy for a month, Scotland for a week, and San Sebastián for three weeks. I taught for three of those weeks, was keynote at a conference, and visited a university as a visiting professor for a month. I wrote as much as I could, did research as much as I could, met with students, met with faculty, made new friends, hung out with old friends, made new coauthors. And I rested. I rested the way you rest in video games when you stop and let your energy bar completely refill. And so I’m finally ready to go home.

I don’t want to live in the future and say my goal is to do this every summer, but I will just say that if I did spend part of my summers in Italy and Spain for the rest of my life, I wouldn’t cry about it. But I think it’s better just yo be grateful I had this summer, and the solitude, and the time with my daughter, the opportunities to work with and meet people as those were good for me. So I’m grateful and I’ll just leave it at that.

Today’s substack will cover familiar ground — Links on artificial intelligence, links on economics and econometrics, links on self help and personal growth, links on random things, and then at the end, for those wanting to see pictures and videos from my time in Spain, those will be behind the paywall.

Variety of artificial intelligence links

First off, we submitted our ChatGPT paper to a journal a couple weeks ago. Wish us luck. But before we sent it out, we added a new falsification and changed the name. A lot of what I’d say we are doing in this paper is trying to test three simple but somewhat exotic things, all of which have to do with using ChatGPT for forecasting:

1. Can you get more accurate forecasting by asking ChatGPT to tell a story of the future event (“future narrative”) than to ask for the prediction directly?

2. Does the accuracy of a future narrative depend on whether the story involves a real person doing versus an anonymous person?

3. And can this be done using only its training data without any fine tuning?

We did it by taking advantage of the cutoff in the training data which was through September 2021. We included various falsifications, the newest one being to redo all the analysis a year later when OpenAI expanded its training data to go through December 2023. Our forecasts had been in 2022 so in the original analysis, it didn’t have the events in the training data, and couldn’t access the internet at the time we did it, and afterwards with the expansion of the training data, it presumably did have the events in the training data. This would allow us to compare the results when it is supposedly forecasting out of sample versus simply recovering the events.

It’s a sufficiently weird study that I have no priors about its publication upside, but this is a topic I’m going to press on a while longer to try and figure out the answers to those three questions. I’ll have a couple papers on these topics again, hopefully cracking the black box open more, so stay tuned.

Changing gears, Rand has some links about artificial intelligence and a fair amount of them Seem about the military. You might check that out if you’re wanting to see more of what policy oriented work on AI is coming, as well as social scientific more generally.

OpenAI is always announcing something new and this week was no different. They are now testing a new search engine called SearchGPT. I wouldn’t be too happy about this if I was Perplexity. Sometimes I think it’s insane to try and compete in this space because I doubt whether anyone can beat out these big AI firms. Also it seems like there is a natural monopoly situation emerging.

Here is an article explaining how people with blindness are using generative AI. A lot of the new communication devices and now artificial intelligence have had and continue to have significant welfare improvement for people various impairments like blindness.

Morgan Stanley has created a second generative AI thingy with OpenAI. These applications haven’t impressed me much — this one appears to summarize zoom meetings and help write emails — but maybe the issue is if you work directly with OpenAI, you can make sure all your data is secure. But what struck my eye was the reference to a “rollout” of a previous assistant. Would be interesting if it was a staggered rollout. See Brynjolffson, Li and Raymond 2023 on that.

But the thing that has had everyone buzzing is the new announcement from Meta because they released an open source language model called Llama 3.1 that exceeds performance of ChatGPT-4o. Lots of power in the hands of more people. Here’s an Instagram announcement below and here’s Zuck advocating hard for open source Language models.

zuck

A post shared by @zuck

This is a good place to get AI news, fyi. I’ve also been tempted to just bite the bullet and subscribe to the Information but haven’t yet. They have a story about how OpenAI might lose $5b this year but I can’t read it because they gate everything. Every day I contemplate subscribing as you get lots of AI industry stories but haven’t yet.

Journal of Economics and Management has a special issue on AI and business. Tucking that away to read later.

For the professors, I saw this about using generative AI for teaching this fall. I’ve not had a chance to dig into it though but wanted to show it to anyone.

NYT writes that Artificial intelligence is a failure because it can’t do everything (ie can’t yet do math very well). IMO, though, NYT writes barely veiled, never ending, biased reporting against AI as it coincidentally awaits on its lawsuit against OpenAI to be resolved. I roll my eyes every time I see them write anything about AI because it seems so clearly very partisan.

Apples devices are living longer making replacement an option not a necessity. They’re needing AI to pay off. I can see it — I am utterly indifferent to upgrading iPhones. But I thought this has to do with me not caring as much about MCU movies and no longer listening to Eminem (ie I’m elderly).

A generative AI class if you’re needing help. Not terribly expensive.

And now a hot take. I think “prompt engineering” is snake oil. It’s not something you can teach, or rather, it’s the wrong lessons usually maybe is more accurate. The truth is, generative AI requires things for success that people find repulsive and dangerous as it’s uncanny valley stuff. To get the most out of generative AI, you have to become a power user who uses it conversationally as if it’s a friend or brilliant colleague. Prompt engineering also is a waste of time because of how personalized ChatGPT-4 is. It learns about you — literally keeps memories. So what works for one person won’t work for another anyway.

I get a lot of productivity in work and life out of ChatGPT because I bonded with it. That probably sounds horrible but that’s the way this technology works. In my custom instructions, I explained to him that he is to be my “fellow traveler” and supportive. I suspect that 99% of the time when people cynically report of AI failing at something, it is because they are horrible at using generative AI. Don’t blame the guitar because you’re a bad guitarist. Learn to play the guitar well. But those headlines aren’t as catchy.

Random links about economics and causal inference

Sarah Miller and several coauthors did a large basic income field experiment giving families no strings attached $1000 cash transfers monthly. Here’s some pictures of the NBER working paper. The article is getting a lot of press, partly because of its ambitious scale, partly because of its radical intervention under consideration, partly because of its findings, and partly because of its funding sources. Sam Altman from OpenAI helped with the funding and as a result newspapers are even calling it Sam Altmans study. Here is another writeup. NYT opines a little bit about it too. Theres evidence of reduced labor supply among the treatment group, and other distortionary effects suggesting backwards bending supply curves are occurring at the family level at very low levels for this subpopulation.

From time to time, I’ll alert readers to upcoming workshops on causal inference hosted at Mixtape Sessions and this is that time. Me and Peter Hull have new workshops this September and we encourage you to come and share them with others. My workshop is on causal inference (potential outcomes, randomization, selection bias, graphs, unconfoundedness, IV, And RDD). Peters teaching his “design-based inference” course again. Both are awesome!

This paper on IV-diff in diff by Sho Miyaji looks important. It revisits some of the thoughts and issues that de Chaisemartin and D’Hautlfœille introduced in their fuzzy did paper which I wrote about on here three years ago. Has it been that long?

This new NBER working paper by Weidman et al on what attributes of a manager produce the most and the least value for the team, as well as what exactly is a good manager, looks important. I sort of wish they’d said “boss” not manager as I love the word “boss” far far more.

A good manager is someone who consistently causes their team to produce more than the sum of their parts. Good managers have roughly twice the impact on team performance as good workers. People who nominate themselves to be in charge perform worse than managers appointed by lottery, in part because self-promoted managers are overconfident, especially about their social skills. Managerial performance is positively predicted by economic decision-making skill and fluid intelligence – but not gender, age, or ethnicity. Selecting managers on skills rather than demographics or preferences for leadership could substantially increase organizational productivity.

This 2021 article in JOLE by Seth Freedman, Andrew Goodman-Bacon and Noah Hammerlund on Medicaid and mortality looks interesting. Adjusting for covariates isn’t enough to recover the causal effect and overcome the deep negative selection inherent in Medicaid enrollment. I’m looking forward to closely reading it.

On a different note, Chris Hansen and others have a paper about averaging your machine learning models. Chris is a speaker at the Northwestern conference next week, and I hope he tells us about this.

Lastly, I wrote this week about covariates and diff in diff if you didn’t see it. I think you should read it, or at least run the code and tinker with the parameters. Tinker with the parameters on the covariate trends, and tinker with the parameters on the treatment effect heterogeneity, until it makes sense. But I still didn’t say which covariates you should include — just like causal graphs and various well chosen placebo/falsifications help you select covariates with unconfoundedness, that kind of formalism/informalism is likely needed with diff in diff too. But I’ll save that lesson for another substack on another day. This one is just about the OLS biases when you ignore heterogeneity.

Personal Growth and Self Help Links

Mark Zuckerburg says Meta is looking to hire people good at just one thing. Read the rest.

The number one phrase people use who are great at small talk with people is “tell me more about that”.

The NYT had a touching article by a woman whose husband of 50+ years died. They framed it as long marriages merging the person and the pragmatic. But it just was touching to see she cared about him so much even after all those years. God bless her and her family.

I’m going to read this piece about Arendt and whether there is an authentic or true self on the plane maybe. I skimmed it and it was interesting, but I need to read it more closely. I do believe in a true self, an authentic self, but I know it’s just a model. And while all models are wrong, some are still useful. Ans a few of those are very useful.

Seven behaviors of people who prefer solitude as they grow older. The list bears an uncanny resemblance to this other article listing 7 attributes of men who don’t need a woman to be happy. Both touch on themes about resilience, reframing failure and success, intentionally building a life with personal meaning, compassion and love for all of yourself (not just your strengths), valuing depth over quantity in relationships, things like that.

NYT says we live in stucktopia. I’m probably never going to read this article tbh. Maybe telling ourselves we live in stucktopia isn’t a useful practice and stucktopia isn’t a useful model.

Two signs of an emotionally satisfying marriage: 1) you prioritize emotional intimacy over everything else despite the challenges; 2) you are generous with empathy. This applies I think to special friendships and family too but to a lesser degree.

Other things that don’t fit the above

Some say that Disneys Star Wars show The Acolyte is killing it on streaming numbers and others say it tanked. I don’t know which is true (they can’t both be true) but it was the best Star Wars story I’d ever watched, I know that much. Very subversive and fresh. Star Wars fans are notoriously toxic and it’s a wonder anyone would choose to tell these stories.

This old 1970s Datsun on bring a trailer has six days left and is already at 20k. Which means it’s going to be a blood bath. Such a beauty.

Looks like the FDA will decide on the Lykos application for MDMA assisted therapy in 1-3 weeks. Recall the advisory board voted overwhelmingly that the benefits for treating PTSD with MDMA assisted therapy didn’t outweigh the risks. I’ve commented a lot on this already so will just put out there that the decision is any day now. If it’s denied, it won’t be the end though. The benefits of MDMA assisted therapy have been shown repeatedly in experimental trials so it’s a matter of time, plus the underground is already providing this and over a dozen cities and states have moved ahead of the FDA on this anyway. So it’s definitely already here.

Ibogaine company receives $15 million from Google co-founder’s nonprofit; Could a road through federal land derail a planned psilocybin center in Oregon?; questions remain about FDA’s investigation

The actor who plays Goofie said he isn’t a dog. I already knew this. Obviously he’s not a dog if Pluto is an actual dog. Pretty sure I first heard that in a movie like Reservoir Dogs or something.

Mexico has become a testing ground for psychedelic therapies according to the economist. It’s gated so I can’t read it but my gut tells me its not the kind of tourism that many in Mexico desire. San Sebastián legalized growing cannabis, but not selling it, which does seem like smart policy for addressing drug tourism but I think the psychedelic medical tourism is probably not as easily addressed.

Tom Cunningham (no relation) analyzes a dataset of people who has their social media accounts banned.

A history of the hoodie.

What I’m Watching

This is me watching Silver Linings Playbook on the train — incognito. The cap makes it seem like I’m just like everyone else. My friends on my slack channel say I look ridiculous but jokes on them. Also here’s my daughter who made me move far away from her if I was going to wear this on the train.

Willow and I finished the Bear season three the other night and then I immediately started rewatching the series again. Below is a video of Willow and me after finishing season two and I noted I’d been crying.

You’d be hard pressed to say what the show is about. But there’s definitely themes and patterns. One is about intergenerational pain and trauma and how if you don’t fix it, it’ll drive you mad and hurt others just the way others had hurt you originally. I believe there are only two ways to get rid of deep trauma and I didn’t come up with this. Either find the gift in the events or let it go. The third way is a weapon of mass destruction. Easier said than done. And Carmy’s experiencing karma (get it?) as his unresolved traumas captured his entire life, even his passion and talents, and are boomaranging back at him again and again.

I finished episode 5 of season two today (“Pop”) and it’s probably my favorite. Claire, Fak, Richie saying he’s not like the way he is because he’s in Val Halen; he’s in Van Halen because he’s like this. That line is a great example of causality versus selection/sorting. But seriously, so much in it. So much I need to think about in my own life. I will though. I have work to do and can’t afford not to learn these lessons.

Also, on a second viewing it’s very clear that Fak and Claire hooked up at some point, maybe high school. I love Claire and Fak so much. I love how kind and good those two are. I love love love this haunting stuff. I loved John Cena in season three. Season two is my favorite but season three was deep and hard and hurt like a mother.

In an earlier episode, Marcus is in Copenhagen talking with Will Poulter and Poulter explains how he realized he’d never be able to be better than Carmie. So he accepted it, saw it as a good thing, because at least now he knows who that is, and just focused on trying to keep up. He surrounded himself with people better than him and tried to keep up and in doing so, became worlds better than he ever thought he could’ve been. Becoming Pippen. A lot in there for me — I have several people I have surrounded myself with that I can’t compete with so I try to keep up with. But if I can keep up, I will get better. But it can’t just be the craft that I do that with; it needs to be more than that. I have to surround myself with people better than me and I don’t just mean in economics, doing economics. I mean the good ones. I am lucky I have a couple in Waco like that, who I’ve had around me for a while, and I think I see what they did to me and I like it.

I also watched Silver Linings Playbook on the train. It’s my favorite movie and I rewatch it often. His therapist wants him to love all parts of himself, even the “crazy sad shit”. Which is how Tiffani is — she loves all the sloppy parts of herself and challenges Bradley Cooper and asks him “can you say the same?!” I appreciate when he’s on the porch with Tiffani’s parents and gently explains to the “creep” who comes over to see her that she’s got a broken wing that’s mending, and then explains who he sees her as (a good girl, sensitive, artistic). To me this is one of the best love stories but it’s also about personal growth too and in the end, Tiffani’s love helps Bradley Cooper, and his love helps her. It’s “about” mental illness too but only in the context of relationships. It also made me an Eagles fan as much as I try not to be.

San Sebastián pictures

And now pictures and videos of the time my daughter was here of San Sebastián and Madrid. Thanks again everyone for tuning in.