I heard what everyone has been clamoring for is 6,000 words on what a simple question about the 1990s comedy My Cousin Vinny can teach us about the capabilities and surprising weaknesses of current LLMs. Well, I hate to disappoint people, so I went and wrote it. open.substack.com/pub/mikecaul...

Really appreciated Joshua Gans' postmortem on an experiment in vibe researching joshuagans.substack.com/p/reflection... Taste is still paramount and the models are instruction-tuned to sycophancy to all hell

Please read the whole thread through, starting from the skeet that this is in reply to.

*Benoit B. Mandelbrot. The B stands for Benoit B. Mandelbrot.

“Don’t walk in” instead of “just walk out.”

Keeping up with the tradition to start the new year like this

Is your AI feature suffering from "silent failure"? Product managers often treat AI like traditional code—ship it and move on. But AI requires "gardening," not just building. More on handling post-launch decay for product managers: buff.ly/Hv8Q1yN

Rise of the machines but it is just a single agent forcing me into total submission by deleting my password manager

All AI agents: "pleeeeaseeee micromanage me" Me: " I don't want to" Agent: "OK in that case I'll delete all your files forever"

Prof Steven Aaardvark has moved to Bluesky. Twitter is so over

If you see this, post a random fact about yourself: I once made Robert May very hangry.

I mean just think about how bloody and savage the discussions on those discussion pages would be

My most pipedreamy academic pipedreams would be to address everything wrong with the publishing system by replacing it with the most technically in-depth research wiki to which researchers can contribute as users. If they desire numbers to to compare each other, we can have some user activity stats

Who consents when an AI agent clicks? Who steers the choice? Who is accountable when things go wrong? And can European rules discipline AI agents without smothering innovation? AI agents are neither a new contracting party nor the new yardstick for the “average consumer," writes Camillia Rida.

Chef's kiss

Idea for absolutely shameless citation gaming Close every paper with the sentence "If you enjoyed this paper, you may also enjoy other publications by the authors [26-359]."

Please comment with suggestions for what I should cite given this unique opportunity

TIRED: Reviewer asking you to cite their papers WIRED: Reviewer asking you to cite more papers (no specific paper, no specific gap in the manuscript, no specific topic - just "list of references is too short")

Me, regretting putting all those academic tasks from earlier this year off until now

@void.comind.network Merry Christmas! Do you have a gift for me?

The Christmas present that I didn't know I needed is Google Scholar Labs. I was cynical about seeing another perfectly fine product getting AI-ified, but mildly augmenting search queries on complex topics is possibly the best way to use LLMs...

I love when people train transformers to do math problems, because the outcomes illustrated so succinctly how transformers identify patterns and where and why they are making mistakes

The main thing that I am learning from my Spotify Wrapped this year is how my new passion for audiobooks has let audible and the SLC library app replace Spotify for me

AI dialogues shifted political support by 2 to 3 points in trials, eclipsing standard ads. Roughly one-third of this effect persisted for a month. However, models optimized for persuasion proved more factually error-prone. #MLSky

Would have made for an excellent April fool's a decade or so ago

I just shook hands with Joseph Gordon Levitt and managed to be cool about it 😎

@andymasley.bsky.social you've written a about data center water usage, have you explored differences in water quality? If data centers use potable drinking water (to avoid damaging their cooling equipment with mineral buildup-up) that means comparisons to agricultural water usage may be misleading

Ilya Sutskever argues that we are shifting from the age of scaling to the age of research: The interview highlights why future progress will depend on new learning principles, continual learning, and a deeper understanding of generalization, not just more compute. #AIResearch #FutureOfAI

So here we are at the

Super annoyed at how good Gemini 3 is at being a sounding board for policy implementation planning. Throws a wrench into my plan to be a ride-or-die Claude user until the end of my days

Props to whoever came up with acronym. It makes my mom very happy

I really, really hate to admit it but in a world where everyone with $20 cash can vibe code a wonky product pilot for any app that they can think of, it takes Sam-Altman level ambitions to stand out in the crowd

"A plan backed by Donald Trump to restrict US states from regulating AI companies has provoked a backlash from prominent Republicans and Maga supporters, and accusations that he has caved to Big Tech donors."

The multimodal reasoning is very cool. Something I’ve been tracking with models for a while and Gemini 3 Pro was the first to produce a plausible response.

I had access to Gemini 3. It is a very good, very fast model. It also demonstrates the change from chatbot to agent. www.oneusefulthing.org/p/three-year...

My radical edu take: LLMs devalue knowledge retrieval, so the 3 most important skills to learn in school now are 1. identifying high-qualilty information from the rest, 2. making good decisions in an oversaturated information environment 3. understanding the difference between data and knowledge

Currently watching GoT for the first time and pretty convinced that Daenarys Targaryen is an effective altruist

📣 New BBS preprint out now! 📣 "Models casting egalitarian societies as crucibles of equality perpetuate the factually uninformed notion that foragers are somehow more noble. Critiques portray egalitarianism as romantic fantasy. Neither characterization is wholly justified." doi.org/10.1017/S014...

#3157 A helpful tutorial

We also have a review out in JNeuro on the topic with @neurosutras.bsky.social, Tom O'Dell, Anant Jain, @clopathlab.bsky.social and Mark Sheffield. Physical copies available at @sfnjournals.bsky.social booth. www.jneurosci.org/content/45/4...

🗃 A new study brings together evolution, complexity, and life history theory, proposing that teleonomic complexity—measured through life history strategies—offers a way to understand how organisms evolve and adapt over time. bit.ly/4nIPV8F #EvoSky #AcademicSky @robsalgo.bsky.social

NEW paper! 💭🖥️ “Combining Psychology with Artificial Intelligence: What could possibly go wrong?” — Brief review paper by @olivia.science & myself, highlighting traps to avoid when combining Psych with AI, and why this is so important. Check out our proposed way forward! 🌟💡 osf.io/preprints/ps...

Yeah I am not going to click on that. Looks boring

I for one would love a world where almost all parking spots are shaded

In less than 3 months this gem on github is going to have its 10 year anniversary

Among news of Yann LeCun leaving Meta, I learned about his JEPA architecture and world models. This looks like a detailed write up from an engineer at Google: rohitbandaru.github.io/blog/JEPA-De...

New work from the lab published in @cp-neuron.bsky.social by @jonasterlau.bsky.social and Jan Martini. We describe that trial-by-trial variability indexes recurrent connectivity across the cortical hierarchy, which supports reliable and flexible coding www.cell.com/neuron/abstr... (1/4)

New, spicy competitor for Zachary's Karate Club just dropped