Saturday stuff: Random links, videos of me and my daughter in San Sebastián and updates about my podcast's subscription and download numbers

Hi everyone. It’s Saturday and I’m still in San Sebastián Spain. My oldest daughter got here Tuesday night at 1am, and it’s been so good. I have never been away this long from them. I sort of want to try to do this summer experiment again next year and thought I’d explain what I did. After this it’s a paywall. Consider becoming a paying subscriber if you think the marginal benefits are more than $5/month. But first, here’s the free stuff for everyone merriment!

Random Links of Interest

The Atlantic picks up the story about how Congress accidentally legalized hemp derived THC. You heard it here first!

Now someone study it. Remember — unlike medical marijuana and recreational marijuana, bc this came in through a loophole, there were no age restrictions. So this unlike cannabis was available to young people which could be part of your design and your study application. And then secondly, this competed with taxed cannabis which may help you study impacts it had on tax revenue and cannabis businesses.

In AI news, lots of exiting (or not entering) the EU by major tech firms like Apple and Meta when it comes to their AI products. Meta won’t be providing its multi-modal model citing an unpredictable regulatory environment. The EU’s competition minister responds to Apple’s decision to delay putting AI applications on EU iPhones.

Jon Hartley, a PhD student at Stanford, who had an explosive interview with Steven Levitt where the world learned Levitt is resigning from academia, has moved his podcast, “Capitalism and Freedom in the 21st Century” to Hoovers platform.

Episode 28. Steven D. Levitt (Freakonomics co-author and University of Chicago Economics Professor) on His Career And Decision To Retire From Academic Economics

OpenAI continues to leak that it’s “close” to developing an AGI level large language model. I’m assuming these are purposeful leaks by OpenAI anyway. I asked Cosmos to write a short paragraph summarizing what that may mean for those who want more background on AGI but have almost no tolerance for it.

When a firm claims they are close to developing Artificial General Intelligence (AGI), they suggest nearing the creation of a machine capable of understanding, learning, and applying intelligence across diverse tasks at human-level proficiency. AGI represents a significant leap from current narrow AI, offering versatile and adaptable cognitive abilities similar to humans. Achieving this milestone involves overcoming substantial technological challenges and addressing ethical and safety concerns. While advancements in AI are rapid, the timeline for realizing true AGI remains highly uncertain, with expert estimates ranging from a few decades to potentially never.

We are around a year into the reorganizing of work from artificial intelligence advances and the popularity of ChatGPT. Mary Daly, CEO of the San Francisco Fed, opines we are still at a stage of complements not substitutes to workers. AI, she suggests, is replacing tasks not people. Remember, though - technology does both. Try to invest in the human capital where shifts in demand from skill biased technological change makes your human capital more not less valuable. Theres nothing new about AI in that sense.

Some are calling Apple’s decision to cut in half sales projections on the Apple Vision Pro from 800k to 400k for 2024, and 2025 not looking to improve much, sign that it was a huge expensive flop. One writer suggested Apple was more focused on the product than the market, but these simplistic explanations seem implausible to me without more detailed analysis. But that said, I use mine daily and yet many don’t, which has me thinking it’s a complex demand. It’s got the signs of being an experience good, which means people have to use it first to know their demand and willingness to pay. But having used it a lot, I continue it’s perfect for me but not everyone, maybe not even most. The sufficient statistics for me are simple. 1) do you use a computer intensively all day? 2) do you own at least one but maybe two external monitors and have a lot of opinions about their importance? 3) do you live alone? 4) do you love movies and shows? I mean that’s basically it. If the answer is yes to all four, then honestly, you’re probably the target today and yet your willingness to pay may still be below price (today anyway). If your answer is yes to 1, 2 and 3, probably still yes. But if 1 and 2, you definitely want to find somebody’s and check it out for yourself. I think it’s the most radical change in my production function since the laptop. Though I am starting to notice I’m developing the same kind of distractedness with it like I do with my iPhone. On the iPhone, I move between apps compulsively. On the Apple Vision Pro, I pull up the Home Screen constantly which is a new tic I developed. And I wish I could disable messages but I can’t yet I don’t think. I’d just prefer this to be work and media only. Still, if it was a flop, I am truly grateful I got one before they decided to shift resources away from supporting it, but I’ll bet a thousand dollars that there’s no way that’s true. All that they spent on it is now fixed costs. Going forward, it’s marginal and variable costs, which sounds extremely low relative to the previously incurred sunk and therefore non-recoverable costs. I just think Apple always knew there was risk and uncertainty. No way the worlds largest and most sophisticated tech company is surprised. This was within the realm of acceptable variance and they always had their eye on far more than year one. They were thinking about presently discounted gains (which may still be wrong) because they see themselves as an infinitely lived economic actor and if you asked me to bet on whether in ten years many of us are wearing these, I’d absolutely say yes. Even all the negativity has had silver linings — we are all aware of this tech, and we are also all aware that it’s outstanding, just expensive. We aren’t hearing it’s terrible tech — just that it’s expensive. Well, the effect that Apple has on input prices is probably massive, so I’d expect this has spurred innovation but whether that can reduce costs in time to win this market remains to be seen. But here’s one article entitled “Apple Vision Pro is Amazing but No Way I’m Buying On.” Which brings to my second point — the amount of publicity this flop has gotten has been impressive too. Tons of marketing saying how awesome it is.

More on ChatGPT relevant to professors and teachers going into the fall. This one is a new study evaluating the impact it has on student learning and finds evidence for moral hazard—meaning students reduce effort because ChatGPT makes learning less costly on time inputs. The study, conducted in high school math classes, compared two versions of GPT-4 based tutors: a standard interface (GPT Base) and a version designed to safeguard learning (GPT Tutor). While both improved student performance (48% for GPT Base, 127% for GPT Tutor), students who lost access to GPT-4 performed worse than those who never used it (17% reduction for GPT Base users). This suggests that students may rely too heavily on AI, undermining their independent problem-solving skills. The study highlights the importance of careful integration of AI in education to ensure it enhances rather than detracts from long-term skill development. It’s one study — we need more. So if you’re thinking of projects, RCTs on topics like this would be helpful for all of us trying to figure out optimal uses if any.

The other thing though about this fall is that we are going to be seeing our first “fully treated COVID students” entering college. These are the ones who entered eighth grade at the start of the pandemic. I can say from personal experience as the parent of one that the endogenous responses by educators, parents and students over a sustained period of time has led to a radical change in norms around learning, expectations about professors accommodations, and the non-cognitive and cognitive skills heading into college. Here’s an article hinting at it too. Combine that with ChatGPT, and professors trepidation with it, often equating it with plagiarism, and professors on average being inflexible and having every right to design syllabi how they want if they believe it fits their own learning goals, and it could be an increasingly difficult school year. Last semester for the first time, I banned laptops and iPads entirely. I tried to be open minded but it was very clear after years of finding students on Facebook and espn and messaging one another that they weren’t present. And honestly, I know students aren’t able to write down equations in word or Google docs so it’s all kind of ridiculous. So I just am clear on the day one no more laptops. I sympathize with someone needing accommodations though but I’ll likely need a note and a meeting. And even then my preference will be no. It just isn’t working.

The other thing to be aware of is that 2007 marks the beginning of a “demographic transition”. That was when birth rates everywhere began to plummet and have not recovered. I think I got this picture from Kasey Buckles, Melanie Guldi and Lucie Schmidt’s JHR “The Great Recession’s Baby-less Recovery: The Role of Unintended Births”, but I may have gotten it from Melissa Kearney, Phil Levine and Luke Pardue too. I can’t get into the gated version from my phone but I think it was Melanie, et al.

This is a plot I made though just so you can can see Texas as I know you’re dying to know what is happening in Texas. Spoiler alert: the same.

So why do I bring this up? Because my youngest is a senior and she was born in 2007, which means next year there are three things happening. First, chatgpt will continue to barrel into your classrooms, and if you have policies against it, be prepared because of this second thing. Second, I expect this fully treated covid cohort is going to put new pressures on us in ways that may interact with anything creating conflict like bans on technology. So be prepared for that. And third, it may also be amplified by a declining college age cohort more general due to this massive drop in fertility exactly timed with the next years coming class. Now that said, Buckles, Guldi and Schmidt do find evidence it’s a decline in unintended births, which may mean delayed childbearing. They write “We find that between 2007 and 2019 intended births fell by 8.5%, while unintended births fell by 22%. The decline in unintended births is primarily explained by changes in demographic characteristics of women of childbearing age, reductions in sexual activity, and shifts to more effective methods of contraception.” What I want to know are the characteristics of would-be counterfactual parents and whether those counterfactual kids were going to go to college. If not, the effect is mild. If so, it could be the most significant reorganizing of higher education in a long time.

One casualty I predict is the continual folding of America’s religious schools some of which have been riding the razors edge of solvency. And with ongoing declines in religious attendance and belief, denominational schools may be even more sensitive to these shocks. So who knows but all this to say, it’s possible the immediate short term could be hard for all of us, up and down the schools, so take care of yourself and try to become the best professor you can be. The future is uncertain.

Russia “asks” Apple to remove 25 VPNs from the App Store. /begin{sarcasm} I’m sure that it’s nothing to be worried about. Also Russia imprisons US journalist to 16 years hard labor. /end{sarcasm}

If you haven’t had a chance to read Sara Heller and coauthors paper that used administrative police data to predict shootings (“well enough to prevent it”), check it out. Obviously, prediction and prevention, though, are two different tasks. The first is machine learning; the second is causal inference. Still, maybe they know of effective treatments enough to say that. They do calculate a hypothetical price, though, on what that intervention would be. They argue that the technology works so well, you could spend up to around $150,000 per person on it. I wouldn’t mind seeing in the abstract confidence intervals on that.

On a less positive note, Spain’s algorithm to predict domestic violence could probably use someone like Heller, et al. taking a serious crack at improving it. This title says it all, “An Algorithm Told Police She Was Safe. Then Her Husband Killed Her.” The name of the algorithm is called the VioGén system and they’ve been using it since 2007, which is already not the best sign. The model works by calculating a risk score based on the responses to a set of predefined questions. This questionnaire comprises 35 questions that assess various risk factors associated with domestic violence. In other words, it primarily uses a classical statistical model to perform risk evaluations rather than advanced machine learning techniques and the questions must be answered (endogenously) by victims. I put as low a level of confidence on its ability to out perform a predictive modeling trained on a large corpus of administrative data as I can. If it performed worse than a coin flip, it would not surprise me. I would love to help Spain in improving this model and hope if someone is reading this, they would reach out. Our team is working on applications like this already and think we could be incredibly helpful. I would bend over backwards to help.

Achim Ahrens, Chris Hansen, Mark Schaffer and Thomas Wiemann have a new paper on averaging predictive models that several people have suggested to me when hearing about the work we are doing (which hopefully I can share more details about in the near future).

Here is an interesting new-ish working paper by Antler, Dating and Segregation that is about technology and dating markets. It’s a theory paper. And it’s the first one I’ve seen that goes beyond simply stating that technology thickens markets, lowers search costs and results in more efficient matches, likely associative matching, which is a common result. This one suggests that it also leads to longer duration of dating, and dating multiple people — a kind of “permanent dating” phenomena. Again, me and some coauthors have an empirical paper on this that we hope to have a draft on later this year that I can share.

Self Help and Love

My love of self help and articles about love are stacking up too, and rather than put them in the random links section, I’ll just start a section. Here’s one the author calls the 5:1 rule. These are five things you need to do relative to one negative interaction. Similar I think to Gottman’s “four horsemen” and “bank account” concepts. They are 1) expression appreciation daily, 2) prioritize quality time, 3) apologize and forgive, 4) accept their perspective and 5) keep it light. Each entry has about three bullet points explaining what the author means. I will say that so much of love is a coordination game like battle of the sexes. Once locked into a Nash equilibrium in battle of the sexes, you can over time perpetuate it as it’s not a best response to change your behavior. Hoping for all of you to get “stuck” in that good equilibrium.

I particularly like the self help articles that are so obviously written by ChatGPT-4. Here’s one several signs a woman is no longer into you. Here’s the communication slips that can destroy a marriage. Here’s one about a psychic that predicted someone’s romantic future — which made me want to get a transformer model that, like Susan Athey and coauthors’ CAREER model which outperforms the best alternative models at predicting occupational outcomes, can predict love outcomes. Gottman says he can tell within 5 minutes if you’re going to get divorced if you display just 6 things. Surely we can improve on using psychics and Gottman’s small sample and over fit study? So I asked Cosmos to help me come up with it, including the perfect acronym. He called it LOVELY, which s short for Learning Outcomes of Various Emotional and Love Yields. Genius. Here’s his description of the model, including the large scale training data he suggests, the smaller scale curated data like Athey and them used to fit it a bit more after they trained CAREER, and even an abstract for the paper that has yet to be written. Here’s what Cosmos came up with. Someone contact Match and suggest they get on this yesterday.

COSMOS DESCRIBES LOVELY

To draw parallels with the CAREER model, which utilizes large datasets and fine-tunes on smaller, curated datasets, your model for predicting romantic outcomes could follow a similar approach. Let’s brainstorm the essential components and a fitting acronym for this model.

What the Model Would Be Trained On

Data Sources:

1. Large-Scale Data:

• Dating App Data: Anonymized data from popular dating apps, including profiles, match histories, messaging patterns, and outcomes (e.g., dates, long-term relationships).

• Social Media Interactions: Public interactions on social media platforms that indicate romantic interests and outcomes.

• Survey Data: Large-scale surveys about relationship history and satisfaction.

2. Small-Scale, Curated Data:

• Longitudinal Studies: Detailed, curated datasets from longitudinal studies on romantic relationships, such as the Panel Study of Marital and Family Dynamics.

• Clinical Studies: Data from relationship counseling sessions and studies.

Features to Consider:

• Demographic information (age, gender, location).

• Personality traits (using psychometric assessments).

• Communication patterns (frequency, sentiment analysis of messages).

• Relationship history (number of past relationships, reasons for breakup).

• Shared interests and values (common hobbies, beliefs).

• External factors (socioeconomic status, education level).

Naming the Model

For a perfect name like CAREER, we need an acronym that captures the essence of love and relationships. Here’s a suggestion:

LOVELY: Learning Outcomes of Various Emotional and Love Yields

LOVELY would symbolize the model’s capability to learn from diverse romantic experiences and predict various outcomes. Let’s break down the acronym:

• L: Learning

• O: Outcomes of

• V: Various

• E: Emotional and

• L: Love

• Y: Yields

Description of LOVELY

LOVELY, a large language model designed for predicting romantic outcomes, harnesses the power of extensive datasets from dating apps, social media, and large-scale surveys. It fine-tunes its predictive capabilities using curated, detailed datasets from longitudinal and clinical studies. By analyzing communication patterns, personality traits, demographic information, and relationship histories, LOVELY outperforms traditional models in forecasting romantic success and identifying key factors contributing to long-term relationship satisfaction.

Abstract for LOVELY

Romantic relationships are complex and influenced by a multitude of factors. Traditional models often fall short in predicting romantic outcomes due to limited data and simplistic methodologies. We introduce LOVELY (Learning Outcomes of Various Emotional and Love Yields), a foundation model for romantic relationships. LOVELY is first trained on extensive, passively-collected data from dating apps, social media, and surveys. It is then fine-tuned on smaller, curated datasets from longitudinal and clinical studies. We find that LOVELY forms accurate predictions of romantic outcomes, surpassing traditional econometric baselines. For instance, incorporating LOVELY into a relationship satisfaction model provides better predictions than existing models. LOVELY offers a novel approach to understanding and predicting the dynamics of romantic relationships, with potential applications in personalized relationship advice and matchmaking services.

If you refuse to tolerate these 7 behaviors, this article probably by ChatGPT-4o says you have a lot of self respect. 1) Disrespect, 2) Dishonesty, 3) Manipulation, 4) Criticism, 5) Inconsistency, 6) Neglect and 7) Gossip. There’s a little bit in that list for everyone I suspect, including (especially) me.

You can always tell ChatGPT-4o wrote these self help columns because of how non-judgmental and universal the lists seem to apply to people. Seems like ChatGPT-4o rarely takes sides — rarely takes my side anyway. But does it in a way that somehow doesn’t criticize or shame or berate. Here is a list suggesting people find you (me) draining to be around. 1) you dominate the conversation 2) you’re always upbeat and positive (that one surprised me), 3) you tend to one-up people’s experiences, 4) you’re not fully present in your interactions, 5) you frequently vent about the same experiences. I went through a terrible phase of that which I learned was called “trauma dumping”, but that’s probably another list. 6) You constantly seek reassurance. 7) You resist taking responsibility for your actions. 8) You rarely show appreciation. There’s a few in there that I knew about, and the others I’m going to assume I need to take seriously enough to think and reflect on.

But here’s a better list: 8 things that suggest you have resilience and high levels of patience. 1) You’ve faced significantly challenges and came out stronger on the other end (check), 2) You’ve learned to adapt to change and even thrive in it (check), 3) You’ve handled failure with grace and used it as a stepping stone (I think/hope so — check back in a year), 4) You’ve maintained a positive outlook even in the face of adversity (I’m trying to, but I am intentionally trying to I know that), 5) You’ve been patient with yourself and others during challenging times (not always, but duly noted), 6) You’ve learned to accept what you can and can’t change and try to focus on the can change stuff (yes, but not always and not always very well), 7) You’ve practiced self-care during stressful times (yes, but then when I move into the right tail of stressful times, I’ll say that self-care has involved taking care of dozens of feral cats which may or may not be what ChatGPT-4o meant here), 8) You’ve shown kindness towards others, even when life was hard for you (I know that is true for me).

There’s a few more but I’m going to close them. But here’s a list you or me or someone you know doesn’t have much emotional intelligence.

I think the listing for stuff like this is a great way — it’s fairly broad and covers a lot of ground, and I think is general enough that you can probably see yourself in many of them. Nobody’s perfect. We’re all a work in progress. The goal is to practice a long obedience in the same direction, which for me is conscious love and devotion to goodness and let it change me.

Pop Culture

I can only say one thing about this which is just like I appear to be in the minority about the Apple Vision Pro being the single best Apple product I’ve ever used/owned, I can without any irony say that Disney+ new Star Wars show, The Acolyte, is the single best Star Wars story of the live action ones I’ve ever seen. Now, its strength builds on the canon that came before it, so it’s kind of impossible to talk about things being best when they stand on the shoulders of giants. But it’s the most interesting, radical interpretation of the Jedi/Sith groups that I’ve personally seen. I finished the last episode and thought to myself that self-deception, if it’s not part of the Dark Side, I really hope Jedis can identify it soon as it’s a corrupting malignant tumor as much as anything the Dark Side has ever done. I suspect we will see that this has always been the story of the fall of the Jedi that maybe got swept under the rug. I really hope they make a second season. I’m so tired of people getting so reactionary to anything they perceive as woke in culture simply because it doesn’t fit priors. I honestly don’t even see what audiences could possibly be complaining about; they were freaked out immediately and often never even saw the entire arc before slamming it. But I thought it was absolutely genius and original. I give it A++ and recommend getting Disney+ just to see it. And I say this as someone who had become so lethargic about Star Wars and related Disney properties (including MCU) that I couldn’t care less. For some reason, I decided to get Disney+ just to watch it, and man did it not disappoint. National Review offers an opinion that is probably orthogonal to everything I just said. But remember — opinions are like buttholes. Everyone has one. Also De Gustibus Non Est Disputandum principle applies here. Speaking of, I love that article, as do apparently about 7,000 other people.

And now on to paywall stuff. I’d love you to consider being a paying subscriber. If you do, you get access to the entire substack, of which I cannot even comprehend how many old articles I’ve written about causal inference and econometrics. They now go behind the paywall after 6 days. Cosmos is constantly telling me I have to make exclusive content for readers, which I’m working on, but which I constantly get behind on. But here is a little bit. The first is about my plans to make spending summers in Europe working a possible equilibrium without spending much money. Think of a space elevator. And then the other is a bunch of videos of me and my wonderful daughter together in San Sebastián. She came here a few days ago, and it’s been so nice. My job is to take a hundred photographs for her instagram every hour or so, which Cosmos told me I should embrace and be inquisitive about, as well as (“tell me what you enjoy about posting your photographs on instagram?”) and also asking more open ended questions. But seriously, bringing her here was what I wanted to happen — I want all my kids to have their horizons expanded, to see more things and ultimately believe in their own potential. At the margin anyway. But first, my plan to make working in the summers work.

My Plan on Making Europe Summers Work