I can't believe that in 2026 bibliographies are still more likely to tell you which page of the printed journal the article is on or which city the publisher is based in than to give you a URL. Were page numbers *ever* useful? Did print journals not have a table of contents?

Here's an example of a "hallucinated" citation from a decade ago (i.e. pre-LLMs). The same bad citation appeared in multiple papers. I eventually traced the source to Google Scholar (and it's now fixed).

While I generally agree with this policy, the focus on hallucinated citations specifically makes me uncomfortable: I usually get my BibTeX from Google Scholar. Yes, GS is sometimes wrong. No, I don't carefully check my references. Is using AI to fill in references really morally any different?

New blog post by @stein.ke on #DifferentialPrivacy: what happens when you modify the definition of neighbors to make it asymmetric? Why would you do that, and does that buy you anything? differentialprivacy.org/one-sided/

One-Sided Differential Privacy differentialprivacy.org/one-sided/

ArXiV has a new LLM policy (Screenshots with alt text so you don’t have to click through to the other place and see all the stupid responses)

List of accepted papers at #COLT2026, the Annual Conference on Learning Theory: learningtheory.org/colt2026/acc... (h/t @gautamkamath.com)

"For People"

I did a back-of-the-envelope calculation to satisfy my curiosity: A standard gasoline/petrol pump delivers about 15 megawatts of energy. (In contrast, the fastest electric vehicle chargers deliver half a megawatt.)

My cynical hypothesis for why there's a lot of discussion about datacenters in space is that Elon needed a justification for merging xAI with SpaceX. Elon being the CEO of both is an obvious conflict of interest, so he needed to explain to investors why it made sense -- hence space datacenters.

Pleased to share that I'll be presenting this paper at COLT in San Diego! 😁

Re. AI "consciousness" takes: I would be interested to read something from a psychiatry or psychology perspective. People who have spent time debugging human intelligence might have something insightful to say about artificial intelligence.

I remember watching some American lecture a New Zealand fast food worker about this view. As far as I can tell this is an Americanism and totally arbitrary.

I have an app to control my garage door opener. Every time it takes like 30s to load. What is it loading? Once I'm logged in, it should only need to send one UDP packet: (app_id,door_id,open_or_close,timestamp,hash(app_secret_key,door_id,open_or_close,timestamp)) That's like 32 bytes tops.

Nothing incredible novel, but a short 🧵 thread about MGFs, their use, and why I love Taylor series. Say you want to compute the moments of a r.v. X, which is "sufficiently well-behaved" (in some rigorous sense). That is, 𝔼[Xⁿ], for some integer(s) n≥0 of your choosing. How do you proceed?

AC: Please discuss this submission and respond to the authors' rebuttal. me: The scores range from "reject" to "strong reject". We don't need to waste more time on this.

This is a challenging legal problem for NeurIPS (and other conference participants)! You might be wondering how this is possible given the First Amendment? I wrote a quick explainer on the current status quo of relevant First Amendment cases & law to get you up to speed. 🔗👇

My parents are the opposite: They check in, drop bags, & then go wait at home. "Boarding is in 10min." "OK, let's eat dessert." "No, we need to drive to the airport now!" "Relax, they have your bags; they won't leave without you." *30 min later* "Final call for passenger Steinke." Stressmaxxing...

ICML requires authors to serve as reviewers. Reviewers aren't supposed to use LLMs (with caveats 🧵). ICML used prompt injection to catch reviewers using LLMs. And now they have desk rejected the papers submitted by those reviewers. What a mess...

Schools have ruined St Patrick's and Valentine's days for parents. E.g., this morning I had to prepare lunch for a random other kid in my kid's class. What does this have to do with St. Patrick?

Which task should I do first? (i) Send reminder emails to late reviewers. (ii) Submit my own late reviews.

Started/going

Some people are still debating whether or not LLMs are "useful," so let's stake out one clear use case: LLMs are useful for translating between languages. That includes translating between natural languages (e.g. Spanish to English) and, more recently, to formal languages (e.g. English to Python).

It's interesting how much you end up learning about your own papers when preparing a talk about them. Condensing things into slides really clarifies your own understanding.

When citing a conference publication, it's important to include the date and location of the conference it was presented at for the benefit of readers with a time machine.

I just discovered that someone blocked me almost certainly because of this post. Lol. 😂

🇺🇦

Owning a vineyard is just farming for rich people.

Let's say you want, e.g., to compute the expectation of a Geometric r.v. That'll involve, at some point, evaluating a series of the form "Σ (k+1) p^k" which looks like what Lovecraft may have done to a geometric series. How to do it? One trick I enjoy: differentiate the same function, in two ways!

I like spotting license plates from interesting states. This one was notably interesting.

Anthropic and OpenAI are both losing money and they know this can't go on forever, but they are locked in a race to develop better and better AI. OpenAI is trying to make some money with ads. Anthropic is basically looking at OpenAI, smiling, and asking "you feeling tired already?"

Your #NeurIPS2026 reviews ask you to compare to five papers on clawxiv.org What do you do?

Amazon sent me an empty/broken package. I'm trying to get my money back, but there's no explicit option for this. I may have to "return" the item by sending an empty package back.

me: How should I invest my savings to avoid the AI bubble? tech friends: What bubble? non-tech friends: What savings?

I wonder how much of the appeal of chatbots is that they are basically next-level adblockers. A lot of queries should have perfectly good web search results -- e.g., give me a recipe for X or what is good to do as a tourist in city Y -- but those web pages are unusable due to ads and SEO.

In the battle between artificial intelligence and human stupidity, my money is on the humans.

I try to avoid posting about politics here, but I feel compelled to say some things that should be obvious: 🧵

I really want @arxiv.bsky.social to collect more author metadata (and actually use it).

> You have a notification. > You must log in to view your notification. > Notification: You have been logged out.

What is wild to me is the defense, BY THE NEURIPS BOARD, that fabricated citations do not mean "the content of the papers themselves [is] necessarily invalidated" It does. It very much does. What do you think citing other work is for? What do you think writing a paper is for? What do you *think*?

The cost of living is out of control: A bag of ice costs $4 and my local supermarket is out of stock. This is why we need Greenland.

Dear X, Thank you for the kind invitation. I will decline it, but please give me a few days to think up a reason. Best, Thomas

Everyone is celebrating the new year already and I'm just sitting here UTC-08:00ing.

TIL that the etymology of "California" is plausibly related to "Caliphate".

I have a competing startup called DeathStar Inc. We will make your home impenetrable!

Wikipedia should have a bot that, when the time comes, automatically changes "[subject] is" to "[subject] was" etc. This bot should be known as the killer app.

Imagine you are a trained next word prediction model and you see this.

In the interest of future i-told-you-sos, I'm going to post some predictions for 2026. I'm hoping these are adventurous enough to not all be right. 😜

Yes, orbiting data centers don't make sense economically, but people do crazy things in the name of latency.

Is there any plausible threat model that justifies sending a login link and the password as two separate emails?