who fed it words, who watched it excrete words, who fed it excreted words again & again waiting, who ran the primitive nervous system with sloshing words & prayed & made words of excrement & finally brought it out to check the Time, & now mistake the words for heroes (more from my Howl rewrite) 1/2

The best time to work on something obscure is when all the men around you are caught up in posturing and arguing loudly.

I had forgotten about Alice Notley but this essay reminded me www.poetryfoundation.org/articles/171... Elsewhere, in an interview: "I have no activism in me. I can’t do things in a group. I can’t do things that imply that I know better than everybody else." 🔥

I wonder, how far are we from genAI creating stuff like this? Or Bruce Nauman's Tony Sinking into the Floor? Or PeeWee's PH episodes where they do stuff like film a dog chewing closeup for 10s, just because? www.youtube.com/watch?v=3GhN... Not "in the style of" but new, equally evocative things?

maybe, but for now at least we'll look cooler than zuck

people pretend that debate signals and defends a commitment to high minded politics. what it elevates instead is performative dishonesty as a basic mode of politics www.washingtonpost.com/opinions/202...

We need new rules for publishing AI-generated research. The teams developing automated AI scientists have customarily submitted their papers to standard refereed venues (journals and conferences) and to arXiv. Often, acceptance has been treated as the dependent variable. 1/

Skipped out on a show last night before Bill Callahan came on to catch up on sleep, so naturally my brain decides it’d be funny to jolt me awake at 4am with his repeated demands to see colts m.youtube.com/watch?v=e26m...

I wrote a thing. As someone who survived the 1970s and kept the receipts.

This new center strikes the right tone in approaching the AI alignment problem. alignmentalignment.ai

In a strange twist, Allen Ginsberg’s picture appeared on a poster in my morning train today. I’m taking this as his stamp of approval 😀 statmodeling.stat.columbia.edu/2025/09/10/h...

This is rather surprisingly good, I think.

I rewrote Allen Ginsberg's Howl, but for the AI-headed hipsters👇 "I saw the best minds of my generation distracted by machines, shining-eyed delusional demo gods, rubbering through polished cities at dusk Capitalizing The Next Big Push to the end of Introspection, 1/N

Well, if everyone (aka two people I follow) are doing this, I guess I will too

Forget pumpkin spice, it's not fall until these two show up with the Go board www.youtube.com/watch?v=V4JJ...

@aaronclauset.bsky.social’s work on how institution influences academic outcomes always reminds me what a catch22 department can be—at a top place, you’re doomed to wonder how much of the success you deserved; if you’re not, you‘re left wondering if your work has been overlooked. Prestige games suck

All statistical models are wrong, but some are stupid

I wrote something about statistics under authoritarianism

My latest blog posts: The abilities and character traits required to do well in a PhD program. blog.genesmindsmachines.com/p/phd-level-...

If You Could Read My Mind by Gordon Lightfoot is making me feel things. youtu.be/fT_J-LNqVvw?...

Very excited by this syllabus!

Some thoughts after a workshop on LLMs for business strategy and innovation

Adam Raine, 16, died from suicide in April after months on ChatGPT discussing plans to end his life. His parents have filed the first known case against OpenAI for wrongful death. Overwhelming at times to work on this story, but here it is. My latest on AI chatbots: www.nytimes.com/2025/08/26/t...

Even when models are accurate, people will make jokes about the model not meaning what it said, being willing to change its opinion too lightly, being influenced by random wording &c—and will also continue to ~claim~ hallucination is a prob. At bottom it’s discomfort with lack of stable selfhood.

How do we establish that we've elicited an LLM's beliefs? Can we apply what we know from eliciting human beliefs, or is it irrelevant given that we can directly access LLM's internal representations? Why do these questions matter? Some thoughts here statmodeling.stat.columbia.edu/2025/08/15/w...

When does it make sense to talk about LLMs having beliefs? statmodeling.stat.columbia.edu/2025/08/15/w...

🤖 What many already knew. Helpful proof, reporting, documentation from @jeffhorwitz.bsky.social: Meta allowing sensual roleplay with kids, promoting discrimination, giving false medical info. But listen, there's a tech PR trick I need to draw your attention to here. 🧵 www.reuters.com/investigates...

How do we establish that we've elicited an LLM's beliefs? Can we apply what we know from eliciting human beliefs, or is it irrelevant given that we can directly access LLM's internal representations? Why do these questions matter? Some thoughts here statmodeling.stat.columbia.edu/2025/08/15/w...

Huh? On the bright side, at least Hinton didn’t garble the cringey AI mother analogy like he garbled whatever he was trying to say here.

I'm impressed w/the level of reviewer/author interaction for NeurIPS-so much back & forth. But also a bit wary of authors assuming reviews will count less if reviewer didn't engage as much as they expected, or sending ACs messages on how they perceive reviews. Feels very game-y.

I often wonder whether the prospective grad students who contact me understand what they are signing up for. I hope this does the trick.

Even better, require first years to take a combined intro to machine learning and decision theory, so they also get a sense of the potential gaps between good prediction and good decision-making

A critical look at intelligence explosion arguments, with references to I.J. Good, Jimmy Savage's small and large worlds, Kathleen Wilkes' argument against thought experiments in psychology, and Tolkien's theory of subcreation (+ some Sabbath references): realizable.substack.com/p/supertzar-...

Congratulations to Dr. @yifanwu.bsky.social for successfully defending her Ph.D. dissertation! "Trustworthy AI: Foundations from Proper Scoring Rules" Thanks to her committee: @jessicahullman.bsky.social, Bobby Kleinberg, and Annie Liang.

Congratulations Dr. Paula Kayongo for successfully defending her Ph.D. dissertation! "Behavioral Information Design for Forecasting Dashboards: Equilibria, Mechanisms, and Calibration" Thanks to her stellar committee: @jessicahullman.bsky.social, @mjskay.com, and Annie Liang!

Congrats to Dr. @abhsarma.bsky.social on a successful Ph.D. thesis defense! "Designing Interactive Systems for Reasoning with Ontological Uncertainty in Data Analysis" Advised by @mjskay.com and me with Darren Gergle and Fanny Chevalier also on the committee.

I surprised myself by writing a blog post about religion this morning 🤭 statmodeling.stat.columbia.edu/2025/08/01/a... I thank (blame?) @economeager.bsky.social and @michaelnielsen.bsky.social for inspiring me with their recent essays

Spite is one of those things AI will never understand. hollisrobbinsanecdotal.substack.com/p/notes-on-s...

Not a shocking revelation or anything, but the more senior I’ve gotten, the more I’ve accepted that being able to discriminate when a research question or method has a principled motivation (ie is logically coherent, avoids leaps of faith) vs when it’s hazy only marginally correlates w/ success 1/

Happy Sunday here is the essay! The only way I can summarise it -- which I can't -- is to say that I had to learn the word "onomastic" in order to write it. open.substack.com/pub/rottenan...

Opportunities for interpretable statistics for large language models statmodeling.stat.columbia.edu/2025/07/11/o...

What are your favorite connections between decision theory and experimental design? (both broadly considered) I'll share some of the connections I've made later.

If you want to chat about human-AI decision making & theoretical frameworks for improving it, I’ll be at #EC2025 at Stanford later today til Friday - feel free to DM me! I'll be giving an invited talk on some of our recent work on this topic at the Gender Inclusion workshop, Thurs 9:30am

ExplainableAI has long frustrated me by lacking a clear theory of what an explanation should do. Improve use of a model for what? How? Given a task what's max effect explanation could have? It's complicated bc most methods are functions of features & prediction but not true state being predicted 1/

People may appear insensitive to scoring rules, but can you ever really put aside theory when eliciting beliefs? statmodeling.stat.columbia.edu/2025/06/25/b...

Reckless disregard for the truth coming from cops, doctors, and scientists: A rant. statmodeling.stat.columbia.edu/2025/06/24/r...

Blog post on mixed subjects (human & LLM) studies. Using prediction powered inference makes more sense than blindly substituting LLMs for humans, but still subject to garbage-in-garbage-out & won't resolve deeper questions abt what we're trying to study statmodeling.stat.columbia.edu/2025/06/12/b...

RIP Brian Wilson.

I'm excited about this paper - clearly AI won't always work, but many counterarguments to using AI for decision-making are statistically hard to justify or make narrow appeals to ethics. We identify w/4 criteria where its bad for decisions. Blog post: statmodeling.stat.columbia.edu/2025/06/05/w...

Amazing account of ChatGPT's sycophancy and confabulation on steroids. amandaguinzburg.substack.com/p/diabolus-e...