Research Preview — This is an early proof of concept. Expect bugs, rough edges, and incomplete content. Learn more →
Harold
0
0
0
0 review cards due
listening...
Quick check
What you learned
Changelog
v1.45 — 19 April 2026
- Omega Preview — a custom, in-house model that fuses a neural network with deep search. This preview phase focuses on validating the search component; pick “Omega” in the Create Game dialog to play against it. Leaderboard marks Omega games with Ω
- Early tests estimate Omega at ~3100 Elo running search-only (without the neural network fused in), measured at a 0.5s time budget per move with ~3.5M nodes/sec throughput
- Difficulty = search-node budget per ply (Casual 700k, Challenge 2.5M, Master 5M). Wall-clock is derived from CPU-calibrated NPS at boot, so Omega plays the same strength on any machine — slow boxes just take longer per move
v1.44 — 13 April 2026
- Harold Voice Mode — real-time voice coaching. Talk to Harold naturally, interrupt mid-sentence, ask questions by voice. Separate Voice button next to Play & Learn
- Voice visualizer — dual-layer audio waveform (your mic + Harold’s voice), large avatar, mute button
- Voice tools — Harold can make moves for you, undo moves, clear board marks, enter quiet mode, and inspect the board — all by voice command
- Play from Puzzle — take any puzzle position and continue playing against the AI
- Mobile layout — board + chat locked to viewport when coaching, eval rail fix
- Custom FEN support — Harold correctly coaches from any position (editor, puzzles)
v1.43 — 28 March 2026
- Play & Learn with Harold — AI chess coach now open to all players, no access code needed. Welcome modal with multimodal AI features showcase
- Multimodal coaching — Harold fuses a proprietary chess neural network, structured knowledge graph, and LLM into a single coaching experience with move-by-move annotations
- ElevenLabs v3 voice — upgraded TTS engine with fully spelled-out chess notation (knight eff three / skoczek ef trzy) for natural voice coaching in English and Polish
- Anti-oscillation engine — prevents repetitive knight retreat loops (Nf3→Ng1) across all difficulty levels; opening book now enabled for casual mode
- Live language switching — switch between English and Polish mid-game, Harold re-sends his message in the new language
- Coach prompt caching — static system prompt enables ~75% cached token hits on supported providers (Anthropic, OpenAI, Google)
- Security hardening — input sanitization, rate limiting (15 req/min coach), Secure cookies, client FEN validation, prompt injection protection
v1.42 — 25 March 2026
- Eval bar — vertical evaluation rail next to the board showing real-time position advantage, like Lichess and Chess.com
- ONNX Runtime backend — lightweight inference engine (~200 MB) replaces full PyTorch (~2 GB), enabling deployment on smaller servers with 2.5× faster CPU inference
- Minor bug fixes — puzzle AI Vision stability, mobile layout improvements
v1.41 — 25 March 2026
- 3 difficulty levels — Casual (~1500 Elo), Challenge (~2200 Elo), Master (~2700 Elo) with separate search depth, temperature sampling, and opening book settings
- Create Game modal — choose side (White / Black / Random) and difficulty before starting
- Lazy game start — server game created on first move, not on page load
- Difficulty badge — current level shown in the evaluation panel
- Leaderboard — difficulty column with sorting, difficulty shown in game replays
v1.4 — 23 March 2026
- Model V4 — 45M-parameter deep transformer with adaptive attention bias, nearly 5× larger neural network than V3. Trained on 2400+ Elo games, scores ~60% against Stockfish (~2700 Elo equivalent)
- Latent thought tokens — the model "thinks" before committing to a move, boosting tactical accuracy without external search. Visible in the AI Thinking panel
- 2-ply negamax search — minimax lookahead with quiescence picks the objectively best move, overriding raw policy when search finds a better line
- Learn platform (closed alpha) — AI-driven interactive course with 48 lessons, 172 board drills, piece tray exercises, and spaced repetition. Fully procedural — every session is unique
- AI Coach Harold — personal tutor with voice, memory across sessions, board visualization tools, and adaptive difficulty
- Puzzle AI Vision — LLM-powered tactical analysis with line visualization and candidate move trees
v1.32 — 20 March 2026
- AI move analysis — hover top moves to see arrows on the board, click to override AI's choice (analysis mode)
- Probability bars — visual move confidence bars with gradient highlight on chosen move
- PGN import — paste any PGN with headers into the editor or replay to view games
- AI puzzle descriptions — LLM-generated tactical briefings for each puzzle via Gemini
- Server-side game state — anti-cheat: server validates every move and determines results
- Replay upgrade — new Chessground board with arrows, last-move highlighting, check indicators
v1.31 — 19 March 2026
- New board — smoother animations, last-move & check highlighting, legal move dots, better mobile touch
- Puzzle Marathon — streak-based leaderboard, solve puzzles in a row without hints to rank
- Puzzle result modal — board snapshot, rating change, solution replay on demand
- Puzzles — tactical training with Lichess puzzle database, difficulty & theme filters, streak tracking, puzzle Elo
- About page — project story, goals, and technical details
v1.3 — 18 March 2026
- Model V3 — hybrid CNN + Transformer with latent thought tokens, ~2500 Elo
- 2-ply lookahead — negamax search with quiescence
- Security hardening — rate limit bypass fix, game result tampering prevention
v1.21 — 17 March 2026
- PGN notation — copy, load, and replay games in standard chess notation
- Smarter Edit — Edit passes current position, Play from here opens editor
- Mobile experience — tap-to-move with legal move dots, responsive board
- Replay improvements — controls under board, Moves/PGN tab toggle
- Better game stats — full move count, per-model win rates, time tiebreaker
- Draw details — stalemate, threefold repetition, 50-move rule
- AI thinking spinner — animated indicator during AI turn
- AI performance graph — AI vs human win rate trend over time
v1.2 — 17 March 2026
- Model V2 — trained on ~100M positions (vs 9.5M in V1), Elo ~2200 (vs ~1800)
- Multi-ply search — fewer blunders thanks to negamax with quiescence search
- Separate leaderboards per model version (V1 / V2)
- Player Elo — optional rating shown in leaderboard
v1.1 — 16 March 2026
- Board editor — set up any position and play vs AI
- Promotion picker — choose piece on pawn promotion
- Dynamic OG images — per-game board preview when sharing
- Leaderboard mini stats — total games, win rate, ranked count
- Replay “Play from here” — continue any position vs AI
- Game-over modal — SVG board with move count and stats
- Forfeit tracking — abandoned games marked as AI wins
v1.0 — 15 March 2026
- Initial release — play chess vs neural network
- 1-ply value search with policy head
- Leaderboard with game replays