Show HN: Noctie – Chess practice using humanlike AI

Hey HN, I posted a proof-of-concept of Noctie here some 17 months ago (”Noctie – A chess AI that predicts your rating”) and got really encouraging responses. Since then I’ve worked full-time on the project together with a couple of friends, with the ambition to develop it into a fully-fledged AI chess tutor.

Noctie is a React/RN frontend for web/iOS/Android + serverless Firebase/Firestore architecture + chess engine in Rust that runs deep learning models that try to emulate human chess cognition, running on GCE instances.

The basic idea with Noctie is to enable for chess what’s called “deliberate practice” which is considered the gold standard for learning a skill. Essentially, you should be able to perform targeted practice at an appropriate difficulty level, and get immediate, high-quality feedback. Both targeted practice and feedback is kind of lacking from online chess.

Noctie makes that possible by being a “humanlike” practice partner at your own level, where you can choose what to play, get hints, go back and repeat, and get feedback on each move. A great use case for Noctie is for example to play normal games in a specific opening and get a feeling for what the common responses are, traps to avoid, etc.

Features so far:

- Plays quite similar to a human, including move timings, from advanced beginner (800 Elo) to Super-GM (2700 Elo) level.

- Instant color feedback after each move which isn’t just Stockfish centipawn loss but more correlated to human intuitiveness of the move

- Custom exercises (”flashcards”) after each game, based on the biggest mistake. These you can then practice using spaced repetition. You can share the puzzles with others as well.

- Choose what opening to play, and the AI is forced to play it + you get hints for the common variations in that opening

- Rating progression, where you increase rating by beating harder versions of the AI, and get extra style points from “playing above your level” style-wise

- Drill scenarios / themes: get instruction in e.g. common theoretical endgames, then drill scenarios against the AI to make it stick. You can make your own themes as well, and share them with others.

Major future milestone I’m working towards is to provide deeper intuition on pros/cons of a move, based on recognition of tactical patterns and positional themes. So far I’ve kept that to a minimum due to many examples from competitors where trying to explain what the computer doesn’t actually understand just ends up looking like nonsense.

Happy to get any feedback on the progress we’ve made, and always up for discussion about the tech stack, the science of learning or chess AI!