Attention is really hard to get in the world. Online, there are kajillions of videos, posts, shorts, etc all screaming for you to click on them. I’ve seen a really frustrating trend where people are not willing to put in the effort to make great products that demand human attention. Writing, videos, and presentations are three examples where you expect someone to put in effort to consume the product. I am really sick of this trend and in this post I want to make a plea for presenters to put in more effort for their presentations.

In the first week of October I made the trek to Oakland to attend SPLASH. Yay! I was excited to splash around with my friends from around the world.

There are many things to like and dislike about conferences. One thing that I frequently dislike is how much effort people put into their slides. Not enough. Or, they didn’t practice the talk in front of third party who can catch obvious no-nos. This year I attended the three major PL conferences: POPL, PLDI, and OOPSLA. At all three, I saw lots of the following slides.

These slides look nice, sure. They’re even web based, cool. But they are obviously AI generated.Generated by Fable 5, to be precise. I gave it our paper, our blog post, and the constraints and intended audience. Agents are great at many things, and they are exceptionally terrible at knowing what humans do and don’t want to look at. There are teeming hordes of bad presentations in the world and, statistically, your AI-generated slides will be no different.

A common presenter misconception is that their audience is listening. Your audience is not listening. Your audience is looking for the first possible reason to stop listening. They have emails to respond to, texts to send, and Instagram to scroll. You’re competing for their attention and there is nothing more boring than an AI-generated presentation.

Use AI, but Better

You absolutely should use AI to make things. However, if it took you a week to make good slides before AI, then it should still take you around a week to make slides with AI. Just try and achieve more in that week.

For example. Matthew Flatt gives some of the best presentations.Having done my undergrad at the University of Utah I am probably biased here. Consider the following presentation, which you should watch in full. If you don’t, watch at least the first 5 minutes, and then again another minute around the 9:45 timestamp.

It’s even more impressive that Matthew programs all of his slideshows in Racket. Obviously, I am nowhere near as smart as Matthew, but with the power of AI, maybe I can get closer?

Here is my presentation from OOPSLA this year. Use left–right arrows to slide around, or click the navigation arrows below the slides. Apologies if you’re on a device that does not support WebGPU, they won’t render (more on that later).

I was particularly inspired by Matthew’s talk I linked above. Especially because his slides are 3D! Or at least they appear 3D.In a brief conversation with Matthew it sounded like he added some custom code for 3D rendering, which doesn’t exist in Racket’s Slideshow. Around the 9:50 mark when he “tilts” the code to show the layers of binding scopes…beautiful. But again, I’m not a good enough programmer like Matthew to build that. But I thought: What if the slideshow was already a 3D scene? Then this slide is a simple camera rotation with some varying z-indexes. Lightbulb.

For my OOPSLA talk I created my own slideshow software. It’s a DSL heavily inspired by Racket Slideshow, and the animation language is also inspired by Oliver’s latest ICFP paper. Written in Common Lisp, it’s a declarative Slideshow DSL that compiles into F-Reps, then rendered on the GPU.I’ll eventually get around to writing a full post about this. I may also open source the project at some point. If you click on my talk above, then hit the ‘c’ key, you detach the camera and can turn the slide around in space.

This post is not to boast or make myself look cool; I am not. This is to demonstrate what’s possible if you use the same amount of time to make things, but do it with AI. Since the presentation, I’ve iterated on the software: cleaned up the DSL, cleaned up the renderer, wrote proofs about the compilation process, etc. The project forced me to learn more about graphics, more about text on the GPU, and even more about Common Lisp and ACL2.

Anytime you make something that requires human attention to consume (e.g., prose, presentations, diagrams) make sure you put in enough human effort to produce it.