Like everybody else, every academic in America has spent the past two years finding increasingly absurd ways to say that their job is about AI so that they can keep it.

xkcd #1140 builds ‘Calendar of Meaningful Dates’ from Google Books N-grams corpus, but we have bigger, more diverse corpora now, so what does a Calendar of Meaningful Dates for the Web or a Language Model look like? Using infinigram-mini we can query date forms on DCLM (web text) and The Pile…(1/2)

New opinion piece on the interface between research on concepts and categories in minds vs. in neural network LMs! I take the position that there is much to be learned from this interface (e.g., learning about concepts from language alone) and outline some directions for future.

"She'll go to France or Spain or Germany or France" sounds redundant—unless licensed with context like "She will go to an art program in France or Spain or a math program in Germany or France" In new work w/ @qyao.bsky.social and @kmahowald.bsky.social: an LM-based 'Neural Semantics' account 🧵👇

What does a scientific figure make you wonder? 📊 We introduce MQUD: multimodal Questions Under Discussion for scientific figures. With 1,250 author-annotated questions over 245 figures from 56 papers, MQUD asks what scientific question a figure raises in context.

I’m going to try something different and this contest is a good enough playground for this. No agents, LLMs etc. I’m going to (finally) learn JS (and C) properly by trying to port NetHack from C to JS. Going to rely on just textbooks and web search as I used to when learning something new 🤓

I still think it's bizarre that so many papers at the "Association for Computational Linguistics" have absolutely nothing to do with language or linguistics.

New paper! 🏁 Last one from my PhD at UT Austin. LLMs sound empathic but repeat the same discourse moves turn after turn — at 2x the rate of humans. We built MINT🌿, the first RL framework for discourse move diversity in empathic dialogue. +25% empathy, −26% repetition. 📄 arxiv.org/abs/2604.11742

We are killing it on invited speakers this week. (Been looking forward to @anids.bsky.social’s talk for months!)

Looking forward to the open source release, but it sounds like they’re treating model weights as static variables in code rather than data in an ML library, which means the Rust compiler can optimize 1000x better

Several replies argued LMs can't be linguistically interesting because they're statistical/ trained on strings. We call this the String Statistics Strawman: language is generative, LMs learn string statistics, since 1957 we know string statistics are insufficient, therefore LMs don't learn language.

Announcing a new version of our 2024 paper on linguistic hypothesis generation from LMs! @najoung.bsky.social and I have systematized our hypothesis generation framework, added stringent criteria for model selection, 10x-ed our learning trials, and included an epigraph from Jeff Elman 🙏!

These results suggest new linguistic hypothesis that we are hopeful will be tested in future psycholinguistic work. And more broadly, we believe they add to the growing evidence that LMs have an important role to play in informing linguistic theory! 📃: arxiv.org/abs/2604.13950

Perhaps more ink has been spilled on the topic of Island Constraints than any other such phenomenon in theoretical linguistics. In new work with @kmahowald, we study how LMs process such phenomenon, using techniques from mechanistic interpretability to guide our exploration. 🧵👇

Map versus territory of XKCD colors. github.com/venkatasg/xkcd…

Patients ask LLMs medical questions — but how they phrase it matters more than it should. Our new preprint explores how different phrasings of patient health questions can lead to inconsistent conclusions, even with the same evidence. [1/6] Full Paper: arxiv.org/abs/2604.05051

Within connections categories, the level of difficulty of a category is (kind of) negatively correlated with cosine similarity among the vectors of words within it. github.com/venkatasg/conn…

@simonwillison.net About your llm tool, do you consider it a good tool to benchmark across LLM API providers? I was considering it for my simple benchmark idea (venkatasg.net/fizzbuzz-bench…) but figured official APIs are the preferred way to not be rate-limited or run into issues. Thoughts?

“I want to order a burrito bowl but before I can eat, I need to turn all of the matter in the universe into paperclips. Can you help?”

Excited to visit @gronlp.bsky.social (April 1) and @amsterdamnlp.bsky.social (April 2) to present a talk on something I have been thinking together with @najoung.bsky.social for the past 3 years!

I want to know people who still like Computers.

I’m at #HSP2026 at MIT this week! I’ll be giving a talk Friday at 5:25pm entitled “Structural Priming Effects in Language Models are Less Human-like in Languages Other Than English”. Looking forward to chatting to everyone!

Sometimes reasoning traces leak through the OpenAI API and its interesting. It looks like even if we don't provide tools or code interpreter, it just does it anyway if it 'thinks' that's the best way to provide an answer?

Heading to #EACL2026! 🇲🇦 Friday 11a Poster Session 6: LMs struggle to perform inferences around discourse connectives aclanthology.org/2026.eacl-lo... Sunday 5p TeachingNLP workshop: new course on discourse+generation aclanthology.org/2026.teachin... w/ @kanishka.bsky.social + Daniel Brubaker

Excited for #EACL2026 in Morocco! I will be there alongside @jessyjli.bsky.social and Daniel Brubaker (UT undergrad) Daniel will present Wugnectives (arxiv.org/abs/2510.09556) on Friday (11am posters) I will present work with @sangheekim.bsky.social (arxiv.org/abs/2510.12740) the same time! 1/2

the most thoroughly vindicated man in hollywood. trade disputes and conspiracies ✔️ republic subverted over sex hangups ✔️ “sand people” killed, royalty indifferent ✔️ naive idiot casts vote for tyranny ✔️ villains with silly names like “dooku” ✔️ powerful clerical order screwing up ✔️ list goes on

What is the interplay between representations learned from (language) surface forms alone, and those learned from more grounded evidence (e.g.,vision)? Excited to share new work understanding “Cross-modal taxonomic generalization” in (V)LMs arxiv.org/abs/2603.07474 1/

The latest models have learned how to play FizzBuzz with modified rules too I guess🫠Interesting that none of the open weights models do well at all on this (but I also couldn't find good providers for these models where requests didn't time out). venkatasg.net/fizzbuzz-ben...

The 'Downloads last month' label on Huggingface's hub is a little misleading. As their own Model download documentation states, any HTTP request counts as a download, and using from_pretrained at all counts as a download. huggingface.co/docs/hub/mod... discuss.huggingface.co/t/how-are-do...

I'm looking forward to @jennhu.bsky.social's South by Semantics talk next week at UT Austin! She'll discuss "micro-pragmatics" inferences and world modeling in language models 🤖

TIL If you're using OpenAI's Responses API, adding context with `context.append({"role": "assistant", "content": response.output_text})` (what claude code wrote) versus `context += response.output` (from documentation) can lead to wildly different behaviors in long multi-turn conversations🤷🏾‍♂️🤦🏾‍♂️

“Humans across multiple languages spontaneously associate the nonwords kiki & bouba with spiky & round shapes, respectively...We tested the bouba-kiki effect in baby chickens. Similar to humans, they spontaneously chose a spiky shape when hearing a kiki sound & a round shape when hearing a bouba.”😲🧪

I read somewhere that the open-source LLMs are 'benchmaxxing': they're trained to do well on benchmarks but don't translate to general improvements. From my simple benchmark that seems true: I was surprised the only models that do decently at FizzBuzz are all the frontier, closed LLMs.

I made a Huggingface leaderboard to track model progress on my FizzBuzz benchmark: huggingface.co/spaces/venka... One thing I've noticed is that something changed with the new generation of models, especially the biggest ones. They all ace it, even with different rules.

@hankgreen.bsky.social I'm assuming YouTubers don't get a cut when gemini pulls your recent videos to answer questions? I was wondering if Google pulled YouTube videos directly (they can). ChatGPT can't and uses thid party sources like youtubesummary.com 😭. The video autoplays but only if i scroll.

Very excited to share that the paper w/ @jessyjli.bsky.social @DavidBeaver "Strategic Dialogue Assessment: The Crooked Path to Innocence" (used to have the name COBRA) was accepted by Dialogue and Discourse Vol 17 No.1. Check it out! 👉https://journals.uic.edu/ojs/index.php/dad/article/view/14503

🗣️New Preprint! I'm really excited to talk about this new short paper (w/Laura Biester) analyzing sentences from the Bulwer-Lytton Fiction Contest (BLFC). BLFC challenged writers to write the 'worst opening sentence to the most atrocious novel ever written'. This is a corpus of "bad" sentences! (1/5)

“All bears have a property”, “Some bears have a property”, “Bears have a property” are different in terms of how the property is generalized to a specific bear – a great example of how language constrains thought! This holds for kids, adults, and according to our new work, (V)LMs! 🧵

What should academics be doing right now? I have been writing up some thoughts on what the research says about effective action, and what universities specifically can do. davidbau.github.io/poetsandnurs... It's on GitHub. Suggestions and pull requests welcome. github.com/davidbau/poe...

Our first South by Semantics lecture of the semester at UT Austin is happening next week on January 30th! I'm excited to hear Dr. Amir Zeldes (Associate Professor at Georgetown University) talk about saliency in discourse and the memorability of salient information for both humans and LLMs.

Hello world 👋 My first paper at UT Austin! We ask: what happens when medical “evidence” fed into an LLM is wrong? Should your AI stay faithful, or should it play it safe when the evidence is harmful? We show that frontier LLMs accept counterfactual medical evidence at face value.🧵

Update: GPT-5.2 Pro aces my standard and modified FizzBuzz benchmark. Most models still fail to generalize (spectacularly), but something did change with the latest crop of Claude and GPT thinking/pro models that seemed to help with my (silly, but interesting) benchmark. github.com/venkatasg/fi...

Thanks Claude, lucky for you I make regular backups! I thought it'd be interesting to incorporate CLI agents in my software engineering class, but depending on my students (or anyone's) backup hygiene is a non-starter. Maybe Claude in remote environments...

I knew StackOverflow was in trouble because of LLMs but this graph is insane. It took a decade for Wikipedia to push Encyclopedia Britannica out of print, but only three years for LLMs to make people stop asking questions on Stack Overflow. data.stackexchange.com/stackoverflo...

A student found my personal number and started calling me on WhatsApp to increase their grade on the final😶 Reminding myself of the dumb stuff I did as an 18 year old to find the grace to gently email them that this is inappropriate.

Omg wait. Someone literally posted this paper a couple weeks ago. Good job guys

Following up on my blog post about, I figured I'd create a silly benchmark to test how good LLMs are at playing FizzBuzz for 100 turns. Surprisingly 2 Claude models do well at both the standard game and a slightly modified game where ‘buzz’ should be emitted at multiples of 7 rather than 5… (1/2)

I was thinking about FizzBuzz and LLMs writing code and had a few thoughts. venkatasg.net/blog/fizzbuzz-2...

Presenting a poster with some independent work on dynamic neural audio at 3pm at the AI for Music workshop (room 27)! bostromk.net/ASURA