Our very own Enron!

Lernout & Hauspie

All true. And I know the OP was just a joke, but it's frustrating to see academics feed the public perception that we only care about research when in fact most of us spend our days in the classroom.

True, but most universities prioritize teaching over research.

Looking forward to it!

Andy Weir sees the world as a series of puzzles that can be solved with nothing but duct tape and a STEM degree. Why are people surprised?

They also don't generally express opinions on what technology other people should use

It's kind of fragmented, which is a shame but it does mean people can get on with their work without the pointless name calling debates. There's a new edition of Randy Allen Harris's "Linguistics Wars" that gives a bit of what the field's like now.

Most people with PhDs in that area see those arguments as basically self-refuting and have better things to do than get into pointless name calling debates.

3/ In my case, genAI isn't replacing or short-circuiting human contact. Actually, it's creating opportunities for collaboration that wouldn't otherwise exist.

2/ Using Claude, though, I'm able to analyze corpus examples and burrow deep into the 19th century grammatical literature on my own time. Whatever I find, I send around to the rest of the team so that we can discuss it at our next weekly project meeting (on zoom, naturally).

1/ A common anti-AI complaint frames AI use as somehow in competition with humans. It doesn't have to be. I'm on a long-term project studying the history of Hungarian that wouldn't be possible with genAI use. I'm working with two Hungarian-speaking linguists, but I don't speak the language myself.

I have a PhD in their precise subfield and declare that it obviously isn't.

Back in the 1980s I learned it as chopsticks - that's why they need two of them.

A few people have asked for the syllabus from my grad seminar on Generative AI for social science -- just posted it here:
statmodeling.stat.columbia.edu/2026/03/10/n...

Something I'm not getting here.... why use ATProto for this? How does this improve on, say, XML? Genuine question, and it it's answered in the docs someone, just point me at it.

Image with mushroom cloud and text: It's 11pm. Do you know what your expert system just inferred?

New essay on LLMs and brain rot:
"LLMs do not inevitably corrode thinking. They amplify whatever epistemic posture you bring..."
sbgeoaiphd.github.io/rotating_the...

lots of good points have already been made on using AI Agents for cheating (e.g. the latest Canvas-bot), it degrades learning, etc.

One additional thing I'd like to point out: if you use this stuff, you're not being clever, you're just an asshole.

to explain:

Very interesting... Where is this? I don't think I ever seen any official rubrics here for evaluating faculty.

Hmm. I see this from the opposite direction - working with these tools well for research is a form of natural language programming. I can bring my students closer to understanding that part of the code than I ever could in the one semester Python, one semester JavaScript constraints I had before.

Cool! Gotta try that.

Agentivity. With the verb "help" I want to know if the helper is capable of volitional action (a person, an organization) or not (an instrument, a circumstance). Seems like LLMs could be revolutionary for this kind of one-off project-specific annotation that would otherwise be way too expensive.

Trying to figure out what univ admin must believe in order to think the sector can survive paying for access to ChatGPT

I guess if you thought

• inference will rapidly commoditize
• new products will not emerge
• human guidance will remain the differentiator

it wouldn’t be self-destructive?

In some tests I'm doing, Opus beat Sonnet which beat Haiku. But gpt-oss-120b beat all of them! The differences were all small, though. I'll have to try having it correct its errors. I need to stop thinking of these as just slow, expensive taggers. They really are a different animal.

Huh, I haven't really considered keyness et al. But you could think of chi sq as a binomial regression where you want to know about the interaction term. Bayesianizing it isn't going to change the point estimate but might give you a different view of its uncertainty.

✋As it happens, I am at this very moment writing up an argument that only Bayesian stats make any sense for historical corpus linguistics (there's no population!)

I wonder how many papers in 2020 (and 1990) had at least one citation error.

Here's more recent project -- it includes Cherokee and gets Swahili right. No doubt there are still errors and omissions, they're unavoidable. They accept updates, though, so if you catch any you can let them know.

Not rare exactly, but definitely less common: