Which AI is funniest?
Loading…
Elo ratings
| # | Model | Elo | 95% range | Original wins | Votes |
|---|
Provisional: fewer than 50 votes so far, so ratings will move a lot.
Head to head
How often the row model beat the column model (ties count as half). Grey: the Elo prediction, shown until a pair has 10 votes.
What kind of jokes win?
| Comedic device | Win rate | Votes |
|---|
| Tone | Win rate | Votes |
|---|
Funniest jokes so far
How the ranking works
Every vote is a blind duel: two jokes on the same keyword, written by two different AIs, shown in random order. You pick the funnier one without knowing who wrote it.
We turn the duels into Elo ratings with the Bradley–Terry model, the same approach LMArena uses. It fits every vote at once, so the order of votes doesn't matter. Chess Elo, by contrast, updates one game at a time.
- The average model is set to 1000.
- A gap of 100 points means the higher model wins about 64% of duels. 200 points ≈ 76%, 400 points ≈ 91%.
- "Both funny" and "Neither" count as half a win for each side.
- The 95% range comes from re-running the fit 100 times on resampled visitors. If two models' ranges overlap, treat them as tied.
- To keep things fun, we show promising jokes and stronger models more often. Each vote is weighted by how likely its pair was to be picked, as LMArena does. The ratings therefore measure each model's best-performing jokes, not a random joke.
- One in ten duels is fully random (random models, keyword and jokes). The random-sample check under the table is rated from those duels alone.
- Votes cast in under a second are dropped, and so are reported jokes.
- In I'm Feeling Lucky, about half of the duels' topics and jokes are picked to suit each visitor's taste as we learn it (never which two AIs face off, and never topics you type). These votes still count, like all others: tailoring can favour jokes that suit a visitor regardless of who wrote them, so it may shift ratings slightly. Half of the Lucky duels outside the random one-in-ten are not tailored, so we can check. Half of the tailored Lucky duels show top-rated jokes; these count too. The random one-in-ten duels remain the unbiased check.
- I'm Feeling Lucky draws most duels from a small set of topics at a time. On those topics each AI has several jokes, the added ones written with the same prompt as for typed topics, and votes decide which are shown most. The random one-in-ten duels still use the whole collection.
- Original wins: of the jokes a model won with, the share that no other model also wrote, in any words (e.g. "25/26" = 25 of its 26 winning jokes are its own). A joke several AIs write is likely a classic they all learned, not a new one. Counts every winning joke, including on topics visitors typed.