June 4, 2026
Bot bluffs, internet scoffs
Ask HN: Gin rummy strategies
AI keeps getting wrecked at gin rummy, and the comments are loving the chaos
TLDR: A developer tried to make AI build a gin rummy coach, but the “smart” bots keep losing to the easy one. Commenters say the likely problem is bad move-scoring, while others insist the real path to victory is tracking discards like a suspicious card shark.
A Hacker News user showed up with a wonderfully relatable problem: they asked artificial intelligence to build a gin rummy training game, and somehow the so-called “medium” and “hard” bots keep getting embarrassed by the easy one. That alone was enough to wake up the card sharks in the comments, who immediately turned the thread into a mix of strategy clinic, bug hunt, and gentle roast session.
The strongest opinion? This is probably not a “strategy” problem at all, but a busted scoring brain. One commenter basically said if the smarter bots are still losing, the real villain is likely the system that decides what counts as a good move. Another big camp argued that the true secret sauce is not fancy card math, but discard drama: stop handing your opponent the exact cards they obviously want, throw away high-penalty junk early, and pay attention to what they pick up. In plain English: don’t just play your own hand, spy on theirs.
Then came the old-school flex. One user casually dropped the ultimate “read the sacred text” move, pointing to Michael Sall’s famously pricey book as the definitive guide. It’s giving “your AI needs homework.” The vibe of the thread was half serious coaching, half amused disbelief that a machine can crunch numbers all day and still lose at a classic card game. Honestly, the comments read like a table full of smug uncles, math nerds, and one librarian yelling, “There’s a book for this!”
Key Points
- •The article is a request for advice on improving AI strategies for a local Gin Rummy trainer.
- •The author reports that medium and hard bots are still losing to the easy bot despite repeated testing and modification.
- •One proposed strategy is to choose draws and discards by minimizing the expected number of unseen cards needed to complete the hand.
- •The article says possible melds can be fully enumerated and that the draw-pile search space is computationally feasible on modern hardware.
- •Suggested improvements include heuristics for knocking, defensive discard logic, opponent-hand inference, and win-rate logging to detect scoring-function problems.