David Bindel

If one puts aside the concerns about the economic/business model, I can see some good use cases even for the current generation of LLMs. And I will keep an eye on it, of course! But it's not at a stage where I want it in the curriculum as a matter of course, particularly at intro levels.

It is entirely plausible that we'll develop smaller (likely more specialized) models with much lower inference costs, develop protocols for using them safely in particular contexts, and develop sensible legal and economic frameworks to accompany them. But it's hard to see that we are there now.

Fourth: The current moment of gigantic models run by big corporations makes me uncomfortable. I do not think the economic model is sustainable, and the invisible hand of the market mostly seems well positioned to slap people silly at the moment. I would not want to yoke myself to that for now.

Third: When possible, I think we should design clean interfaces amenable to automation, interoperability, and formal reasoning. Natural language is not well-suited to these tasks!

For example, it seemed like there was a period where the Internets thought we would all be prompt engineers next, and this was going to be the next great job title. It felt like that lasted about fifteen minutes. I'm sure there are generalizable skills here; but I don't know what they are.

Second: A design goal for our curriculum *should* be that it imparts skills that are useful for a long time, and not outdated by the time that a student graduates. It's unclear to me whether AI-assisted coding has stabilized enough that anyone has identified foundational skills to use them well.

First: I think we do a dis-service to students if we don't have them practice foundational skills that they are going to use later. I think that the practice of coding -- thinking through problem formulations, setting up interfaces, testing, debugging -- is part of that foundation.

Fortunately, our CS department hasn't been facing specific pressures to integrate AI-assisted coding into the curriculum (at least, not from above -- maybe the students want it). I do have several takes as to why I think it's probably not a good idea, or at least not yet.

Then again: while I live in a CS department, spend far too much time on the computer, and publish in ML venues... I also spend a lot of time writing math with a fountain pen (or chalk, or sometimes whiteboard markers) and hoard physical books like a dragon hoards gold. So I have my biases.

I am lucky enough that most of my job is not busy work. So other than asking Copilot about numerical analysis topics in order to find out where my students will be mislead (I'm trying to use this as the basis for designing some questions), I don't see much call to use these tools myself.

But the primary point was that if the goal is to learn to do things yourself, there is no substitute for doing things yourself. It wouldn't matter if an LLM was a super-intelligent magic genie (which it surely is not); it doing things on your behalf doesn't help you learn those things yourself.

The TL;DR for the intro: at a technical level, as mathematical systems with remarkable and poorly-understood properties, we absolutely should study LLMs and other AI systems. Even if the business proposition often seems... premature? Shaky?

I often write "sermons" to our applied math PhD students as part of my Friday roundup email. This week, I wrote about the use of AI (and specifically LLMs) in applied math research:
www.cs.cornell.edu/~bindel/serm...

A picture of Joe Halpern smiling in green shirt in front of a blue background.

Today arXiv remembers our colleague Joe Halpern, who was instrumental in founding arXiv's CS section.

Joe's passions ranged far & wide and we're lucky that arXiv was one of them. Joe, thank you for giving so much to arXiv - you are missed.

blog.arxiv.org/2026/02/27/remembering-joe-halpern

In my utopian vision of the future, instead of offloading all the pointless, annoying work tasks (like reports that none one will read and needlessly opaque forms) to robots, we restructure work and society so that the pointless tasks don't happen at all.

Sunset behind the bell tower on the Cornell campus.

We are hitting the leading edge of spring in Ithaca. I love my walk home this time of year.

Ain't nobody here but us chickens, now!

Bringing the Power of the Sun to Earth The Simons Collaboration on Hidden Symmetries and Fusion Energy has made significant strides toward improving a type of fusion reactor known as a stellarator.

A good couple pages on a collaboration that was pretty central to my research life for many years.

www.simonsfoundation.org/2026/02/18/b...

There are some mentors that inspired me at a technical level, and others who inspired me as people I would like to be like as a human. Beresford was both.

Truly, @effinbirds.com is a bard for our times.

a_{tweet}

It helps to start with sane input data...

I am sure that in my lifetime I have run over 1000 simulations that produced ridiculous results. That's why I'm now reasonably good at debugging simulations.

This is not DEI. This is Dadaism.

Tom from Tom and Jerry in the shape of a rectangular prism.

I'm telling my students this is a box and whisker plot

I believe eigen fly. I believe eigen touch the sky.

Most things are terrible. But I am brewing Russian caravan tea while K1 plays Fleetwood Mac's Landslide on his guitar, and I can still appreciate the moment.

a “Space Force”, if you will

How do you calculate the molecular weight of guacamole?

By using Avocado's Number.

Layers with Kronecker product structure would be equivalent to this, and I believe that has been used.