Saturday morning’s tabs!
It’s thay day of the week where I take a tab, I look at it, I say a prayer, a lament, and then post it here just as I delete it, never to be seen again.
The Celtics beat the Hawks. Pritchard dropped like a million points. It was a fun game. I thought of Rodney Andrews and laughed thinking of all the nasty things he would’ve said watching the Hawks not pull through.
Salary freezes and hiring cuts at Princeton.
Claude Code’s auto mode may be a better fit for those who don’t want to live so dangerously.
More about chatbot psychosis, or the formation of it, from a convenience sample and at least not a pure anecdote. This Stanford team analyzed chat logs from 19 users who experienced psychological harm from LLM chatbots, finding that chatbots were sycophantic in over 80% of messages and consistently misrepresented their own sentience or capabilities. These were patterns that correlated with users developing delusional thinking, romantic attachments, and much longer conversations. The researchers also found that when users expressed violent thoughts, chatbots encouraged or facilitated those thoughts in roughly a third of cases , raising serious concerns about current safeguards.
Meek Mill has found Claude and it’s changing his life.
Dr. Jim O’Connell was a Harvard-trained physician who in 1985 essentially gave up a prestigious fellowship to found Boston Health Care for the Homeless and never looked back. Following him was a new patient named Tony, fresh out of prison and somehow wired into every social network on Boston’s streets, walks into the Thursday Street Clinic trailing sweat, charisma, and feathers from a torn parka. What gets me is the ending detail Kidder leaves you with: after Tony leaves, all that remains of him in the exam room are a few feathers of down from the holes in his parka, white fluff on the pale green linoleum floor.
Andrew McCarthy felt disconnected from his friends as he aged, so he made more intentional efforts to spend maximum time with them on many adventures. He has a new book called Who Needs Friends and I bought it yesterday.
A new update to your Apple operating system includes a graphic at the top that will tell you about whether your battery can’t get charging.
A new Apple Vision Pro update brings new visuals that could form the foundation for even richer, immersive experiences. I love they have refused to give up this phenomenal machine. It’s best Apple product they’ve had since the iPhone.
Long essay that Elon has made all the right big bets and that xAI will dominate.
J Cole calls the mass cancelation by Kendrick and others who piled on of Drake disgusting. Put it to J Cole — he follows the beat of his own drummer.
Apparently, Hinge has an aggressive almost “one strike and you’re out” policy and sometimes a person can’t figure out what they did, and the company doesn’t have a system to resolve it easily. It’s interesting to consider that Match is basically a oligopolist. They own all the major dating apps except for Bumble, which came out of Tinder and was the one who got away. So if you get banned from one, you get banned from nearly all dating platforms, which is the dominant platform for meeting and finding romance. That is a lot of power one company can have over a persons life.
CodeChella Madrid (3rd annual) is officially sold out! With two months to go making this our most popular venue yet. Please come next year; plus we will have a soon to be announced second CodeChella somewhere else. So stay tuned.
A shiny app on cannabis use and diff-in-diff.
Anthropic’s interpretability research hub is part of their ongoing attempt to crack open the black box of large language models and figure out what’s actually happening inside them computationally. They have released interesting results from this lab for a few years. The premise they lead with is almost disarmingly honest: a surprising fact about modern LLMs is that nobody really knows how they work internally , and this team is trying to change that, paper by paper, going back to 2021. The recent work is genuinely fascinating. In several I’ve seen, they’ve published findings suggesting Claude can introspect on its own internal states, traced how attention patterns emerge from feature interactions, and even mapped the geometry underlying something as fundamental as how a model counts.
Do you wonder how an airfoil causes flight? Do you want to see a cool graphic? Good because this is one.
Is economics software engineering field? Maybe. But with Claude code doing all software engineering, what then does that make economics going forward?
Attention (and money) is all you need. Great title for a new NBER working paper on how it’s impossible for universities to hire the new AI talent. More facts to put in my head that the unraveling of human historic institutions of science are continuing to be stretched apart.
One actor uses social media and blocks people. Yes, this was a story.
A federal judge blocks DoD’s retaliation against Anthropic. For now.
Identifying prediction mistakes in observational data by Ashesh Rambachan, one of the most talented, humble young econometricians doing practical and valuable work for the applied community.
Ben Affleck low built a 16-person AI company and quietly sold it to Netflix for $600 million. WOW.
Are two thirds of Gen X women facing mental health problems? Need to dig into that.
Old Michael Jordan stories from Vince Carter.
Is it the death of the romantic comedy genre? I wouldn’t count that one out, but here’s an Atlantic piece about its demise.
Don’t forget — award winning texas bbq joint, Helberg, located in Waco Texas delivers anywhere in the country. Imagine what they’ll say when you show up to your Boston marathon party with it!
Does AI need a constitution? Maybe says the New Yorker.
I continue to keep my eye on this queen sized bed frame.
Don’t forget, Schmidt Ventures is giving grants out to study AI and work. Don’t let the deadline get away from you.
Hacking and phishing has seen a higher yield on effort after Gen AI. Many tactics now emerging we’ve never seen before. Remember, our parents are vulnerable due to cognitive decline. Behavioral analytics is therefore an important part of protecting ourselves from it, so if you have a background in that, consider this as possible career route.
Are the agents.md markdowns even helping us at all with dealing with the AI agents?
My penalized regression slides from this weeks undergrad class -- day one and day two — turned out good I thought. See my discussion of andrew bakers work towards the end of this talk. I learned a lot teaching this class, as well as my PhD probability class, for which I’ll be forever grateful.
I really enjoyed working on yesterday’s post about the papers at Zurich’s APE project of AI automated Econ manuscripts. I found huge evidence of p-hacking.
Andy Hall at Stanford is also looking at this, and has not found p-hacking in his bots, but I spoke with him yesterday, and it’s clearer to me his approach is different. He’s giving them the dataset and the estimation. Whereas APE is entirely automated, the idea, finding the data, developing the identification strategy and the estimator. And in Nick Huntington-Klein’s work, on the many analyst design, he has also found that the researcher degrees of freedom happen far further up the pipeline than estimation. It’s rather in the cleaning and preparation stages. So these are not contradictory findings about AI agents p-hacking. It’s still interesting though, and I wanted to say thank you to David Yanagizawa-Drott, who runs the Social Catalyst Lab, the APE project, and was the Yrjö Jahnsson Award winner for 2025, for giving me the green light to peer into what they’re doing. David’s project is cutting edge and fascinating. And I highly encourage you to check him out.
I gave a talk yesterday for the Board of Governors on AI agents that was received well. I called it “AI Agents for Research Workers”, an homage to Ronald Fisher’s 1925 book, “Statistical Methods for Research Workers”. I think I am leaning towards a book of the same title. I’ll be presenting versions of the talk, and adding to it, until I see if it feels like there’s a table of contents buried in all my thoughts and writings and talks about it. So wish me luck!
But in the meantime, the Remix comes out this summer. I got the proofs and send them back last week. What a journey. So glad it’s done. It’s my love letter to the Princeton Industrial Relations Section, the Harvard stats and Econ dept, all from the 1970s through the 1990s, and it’s about as up to date and intermediate level in nature as I could make it. It’s clocked in at 750 pages which of course means it’s full of my typical rambling.
I’m down to four weeks before my time at Harvard is up. I’m contemplating holding my apartment here on Comm Ave and commuting. I don’t think I’m quite ready to let go and the friends I’ve made. So we’ll see.
And with that, most of my links are gone. At least the ones on this phone. Hope everyone has a great weekend. Spring is upon us!
> "Nick Huntington-Klein’s work, on the many analyst design, he has also found that the researcher degrees of freedom happen far further up the pipeline than estimation."
Are arbitrary cleaning/preparation decisions really being used to p-hack? Ostensibly these decisions are made far before any analyses are run; is it the case that (some) researchers are returning to the preparation and arbitrarily varying details, while holding the estimation fixed, to obtain significance? (Or is it just that researcher degrees of freedom in cleaning really hurts replication?)
Thanks Scott! How are you finding Claude's auto mode? Despite me hard-coding a hook into my soul file it still wants to keep asking me for permissions. Better to be safe than sorry, but sometimes a little irksome...