Upgrading the Sage

The Sage from Always Never, wearing a dark green pointed hat and robe and holding a staff

Always Never is my tiny game about inventing proverbs a word at a time. It's inspired by word-at-a-time improv games such as 3-Headed Expert. It's still evolving, but it's live here.

I created it in a week or so in July for OpenAI Build Week. I had to freeze my work on it after that for a while until the competition results were announced (no surprises, I didn't win!).

In recent weeks, I've been working on some improvements.

First, I wanted to give it a visual makeover. I liked the SVG Sage which got created as part of my initial prototype, but I wanted more sophisticated animation.

I got GPT Image to render a series of images of the Sage in different poses to reflect his mood (pleased, frustrated, excited, etc.). I then fed these into Claude Design to work out how to integrate them into the site.

Mood sheet showing the Sage from Always Never in twelve poses, including waiting, thinking, offering, worried, delighted and bowing — See the full-size mood sheet

Secondly, I really wanted to work on the quality of the Sage's word selection and choices. This is where the game really lives or dies — it has to be fun, and preferably a little wacky. The proverbs that get created need to make sense grammatically, but also have a certain playful, whimsical humour to them. That means improving the selection of words that the Sage offers each round, choosing a good follow-up from the Sage, and better identifying when the proverb might be complete.

This has taken me into interesting terrain. Here's a summary of some things the algorithm is now doing, explained by the model in simple terms ("ELI5, without resorting to inventive metaphors").

From the model

The engine still enforces grammar first: a small state machine tracks which word class can legally come next, so every proverb parses. On top of that, recent work borrows from three fields.

From distributional semantics: every word now carries a compact word embedding derived from ConceptNet Numberbatch, compressed with principal component analysis and shipped with the game. Cosine similarity between these vectors lets the Sage measure how far a candidate word sits from the words already on the table, with no live AI call.

From computational humour: twists are modelled on incongruity-resolution theory, using a local-versus-global surprisal test. A deliberately odd card is only offered if a two-word lookahead finds a way to make it fit, and each twist is labelled with a named rhetorical move such as personification or literalisation.

From search and ranking: word choice uses Gumbel-top-k sampling instead of a fixed argmax, so good words rotate rather than repeat. A regression harness plays hundreds of seeded games, and a Bradley-Terry model ranks policies from players' pairwise votes.

In testing, distinct words offered rose by about a third and hands containing a resolvable twist rose from 86 to 97 percent, with no loss of grammaticality. The next gap is selectional preference, meaning which objects each verb can plausibly take, which is why it can still suggest cracking a rope.

I find it exciting, and a little odd, to be able to develop things like this without needing to master the underlying algorithms and research that they depend on.