The talk, as given; watch on YouTube if the player does not load.
© 2026 Michael Todd Edwards · MIT License — copy it, change it, give it away.
Building the activity I wanted meant being a programmer, so for most of thirty years it stayed a description in my head. That constraint is gone, and these three are what came out of it — low floor and high ceiling, game-like, laboratory-style. Telefone has a model inside it, where the machine is the specimen rather than the author. Metaplete and Name Game have no AI in them at all, and both were built with it.
- Telefone
- Telephone, except the machine is every player. Hand it a page — a student's own writing is best — and ask it to clean it up. The result goes back in to be cleaned again, round after round. Each round shows what is left, what this round changed, and what is lost, struck through. Nobody decides to stop; the game ends when the machine has nothing left to correct. What it corrected away turns out to be the writer. It needs a model to play against — the free local one under Run a model yourself does fine, and nothing leaves the machine. Code.
- Metaplete
- Delete cells until every row and every column hits its target. The puzzle time-stamps each move as you go, so afterwards a student can scrub back through their own solve and annotate it: when they knew, when they were guessing, where they backed a choice out. No model, no account, nothing to install. One of three in Watch Your Own Thinking, alongside Metadoku for deduction and Metagons for working backward.
- Name Game
- Learn your students’ names from their photos before the first week is out. Study walks the roster with the names showing; Quiz hides them and asks, and you write your own hint for the faces that will not stick. Nothing is uploaded and there is no account: the photos sit in a folder on your own computer and never leave it, which is the only defensible way to hold a room full of teenagers’ faces. No model, no network, no AI in it anywhere. Downloads for Windows, Mac and Linux are on the releases page. It arrives with the fifteen most recent presidents already in it, so there is nothing to collect before you play; a second roster — writers, scientists and reformers is here if you would rather. Neither one is anybody’s student.
Showing any of this to a class over a video call: share your computer’s sound, not just the screen. In Zoom it is a tick-box in the share dialogue, easy to miss — and without it, anything an app plays reaches nobody.
© 2026 Michael Todd Edwards · MIT License — copy it, change it, give it away.
What ran during the talk was a 3-billion-parameter model, on an ordinary laptop, with the wifi off. Here is the whole recipe.
- Install Ollama — Mac, Windows or Linux, one download.
- In a terminal, fetch the model. About two gigabytes, once:
ollama pull llama3.2:3b
- Open any of the three apps and choose Local in its AI Setup. It finds the model by itself.
Any laptop from the last few years with 8 GB of memory will do. It is faster with a graphics card and fine without one. llama3.2:3b is Meta's, under the Llama 3.2 Community License; Ollama's library has plenty of others to try.
© 2026 Michael Todd Edwards · MIT License — copy it, change it, give it away.