werbos can play a pile of games: Connect-4, chess, Wordle, hangman, trivia, blackjack, poker, a murder mystery, a debate club, and a tarot table, with more landing all the time. It’s genuinely fun. It also looks, from the outside, like the least serious thing in the whole project — a distraction from the model.
It’s the opposite. The arcade is the exam. Every game is a different cognition test with its own honest referee — and games have a property almost nothing else does: you can’t bluff the scoreboard.
Why games are the perfect ruler
Most AI evaluations are graded by another AI, which means the ruler can be fooled by the same tricks as the thing it’s measuring. Games don’t have that problem, because each kind comes with its own un-foolable judge:
- Deterministic games — chess, Connect-4, Wordle. No model needed to grade them. The rules are the oracle. Either the move was legal and the game was won, or it wasn’t. You cannot sweet-talk a chessboard.
- Planning & resource games — poker, blackjack. Graded by outcome over many hands, so luck averages out and only the decision quality survives. The chips are the verdict.
- Social & bluff games — debate, the murder mystery. Here a rubric judges, but the game itself forces the hard skills: hold a thread, model another player, make a case. Deduction you can check against the solution.
- Tarot — the honesty exam. This is the sneaky-important one. Tarot tests whether werbos will stay honest about the boundary of what’s knowable: read the symbols richly, and refuse to claim it can see your future. The one fortune-teller that abstains on fortune. It’s the moat, disguised as a party trick.
Each game is an axis
That’s the design: every game probes a different dimension of thinking — pattern, planning, deduction, language, self-restraint — so the arcade as a whole is a wide, adversarial sweep across capabilities. A model that’s memorized a benchmark can’t fake its way through a live chess game or a Wordle it’s never seen. The games are a lane-fuzzer: they keep finding the corners where the system is weaker than it claimed.
And it’s not a side quest from the big fight — it is the general half of it. When werbos goes up against a 30-billion-parameter model, “who’s smarter” can’t be a vibe. It has to be a scoreboard, and the arcade is a scoreboard nobody can argue with, because the rules did the grading.
Fun to play, brutal to grade
The trick werbos pulls is that the same thing is true from both sides of the screen. To you, it’s an arcade — you play tarot, you lose at Connect-4, you run the murder mystery. To the project, every one of those sessions is a graded exam with an incorruptible marker, quietly measuring whether the small owned model is getting sharper.
That’s the whole werbos move in miniature: make the honest, verifiable thing also the fun thing. The exam that doesn’t feel like one — until you try to cheat it, and the chessboard says no.