📡 Idea Radar 每天 8:00 / 18:00 推送 · 18:00 生成精选
reddit · SideProject · 50 分 ← 返回列表

I built a chess arena for LLMs. The moves were bad and the explanations were worse

I run a small chess review app, and I added an arena where two language models play each other while Stockfish grades every move. You bring your own API key and it stays in your browser tab. The engine runs in the tab too, so a match costs me nothing to serve, and you don't need an account. The game I put on the page is Claude Haiku 5.5 against itself. Between moves 6 and 14 the two sides traded 11 blunders, and over the whole game they needed 5 retries on illegal moves. I expected bad ch

五维评分

付费意愿
35
需求真实度
38
个人可落地
88
竞争空间
42
新鲜度
68

AI 分析

痛点:LLM 讲解棋局时说得斩钉截铁,却经常是错的(走出漏着还配上同样自信的理由),学习者无法分辨真假,因为讲解文本没有被引擎校验。

目标用户:用 AI 学棋的业余棋手、想给产品加“AI 教练”功能的国际象棋 App 开发者

证据:Between moves 6 and 14 the two sides traded 11 blunders... every one of those 11 blunders to come with a reason that sounded completely sure of itself. On move 20 Black told White its back rank needed care. Two moves later Black got mated on its own back rank. ... The sentence that comes with it gets neither, and it reads just as well when the move is a blunder. So if an LLM is explaining chess to you, I wouldn't trust a word of it unless an engine has checked the position first. The whole 22-move game cost about 3 cents.

变现:订阅制“引擎校验过的 AI 复盘”:免费看 1 盘/月,付费解锁无限复盘与可导出报告;向棋类 App 提供按次计费的 verified-explanation API(成本可控,约 3 美分/盘)。

MVP:浏览器端 Stockfish(WASM) + 用户自带 LLM Key:先由引擎算评估与最佳着,再把局面数据和评估注入提示词生成讲解,最后对每句讲解做一次引擎事实核验,输出带“已验证/存疑”标签的复盘页与可分享链接,2 周内可完成。

竞品:Chess.com Game Review、Lichess Stockfish 分析、DecodeChess、Aimchess、Chessvision.ai

摘要:作者已有小型国际象棋复盘 App,新增了 LLM 互相对弈、Stockfish 逐步评分的竞技场,用户自带 API Key、引擎跑在浏览器标签页里,服务成本几乎为零。展示的一盘 22 步对局中,第 6 到 14 步双方共下了 11 个漏着,还出现 5 次非法走法重试,整盘成本约 3 美分;更关键的是每个漏着都配了同样自信的“理由”,甚至刚提醒对方注意底线,两步后自己就被将死。结论是 LLM 的棋局讲解在被引擎校验前不可信,这提示了一个“引擎验证过的 AI 复盘/教练”的差异化机会,但帖中没有任何付费意愿或收入数据。