Back to home

Compare

Comparing: Qwen Plays WoW Solo — A Weekend Agent Project More Interesting Than You'd Think & 通义千问能自己打魔兽 — 业余项目里的 Agent 实践,比想象的有意思

AEN
Tongyi QwenMCPAnthropic·

Qwen Plays WoW Solo — A Weekend Agent Project More Interesting Than You'd Think

What this is

A Reddit user used AI-written code to put Alibaba's Tongyi Qwen in control of a World of Warcraft character — all text-based decisions, with output above 50 tokens per second. The game is open for public testing at jankcraft.xyz. The story isn't the demo itself; it's that this is a personal weekend project.

Three steps: first, an Agent (an AI program that autonomously executes multi-step tasks) stood up a private WoW server; second, a browser-based client eliminated the need to install the game; third, MCP (Model Context Protocol, Anthropic's open standard for "AI tool-calling" interfaces) drives the game client directly, with finer granularity than a generic browser agent.

Almost no traditional code was written. The author's method: vibe coding — telling AI in natural language what to build and letting it write the code.

Industry view

Supporters see this as a meaningful expansion of the MCP ecosystem. Since MCP went open-source in late 2024, an increasing range of tools — databases, browsers, IDEs — have plugged in. This project proves MCP can drive a complex game, suggesting "tool-calling" is becoming a general capability rather than a vertical-scene exception.

Skepticism is warranted: a WoW server is a closed environment with enumerable states and clean interfaces. Real business has dirty data, broken workflows, and an order-of-magnitude higher complexity. Others point out that vibe coding suits prototypes but is far from production-ready. The author himself admits he's "still vibing" — this is a weekend project, not a product.

Another angle we think is being missed: game AI isn't new — past efforts used purpose-trained small models. This time, a general large model plus a standard protocol pulled it off, meaning the same capability can transfer to customer service, scheduling, ops, and other long-chain tasks.

Impact on regular people

For enterprise IT: protocol-level standards like MCP for "AI tool-calling" are becoming infrastructure. When IT departments evaluate new systems, "does it support MCP?" may become as standard a question as "does it support REST API?" once was.

For working professionals: vibe coding lets non-programmers build usable systems. This project took one person a few days; a traditional team would need weeks. Employees who can write strong prompts will gain more influence inside their organizations.

For consumers: a signal that AI assistants are crossing the threshold from "can chat" to "can act." A model that can play WoW solo is, technically, not far from an assistant that books your flights and handles your spreadsheets.

BZH
通义千问MCPAnthropic·

通义千问能自己打魔兽 — 业余项目里的 Agent 实践,比想象的有意思

这是什么

一个 Reddit 用户用 AI 写代码,让阿里通义千问自己玩魔兽——输出速度需超过每秒 50 token,全程纯文本决策。游戏已对外开放试玩(jankcraft.xyz)。这件事的看点不在演示效果,而在于它是个人的周末项目。

具体三步:第一步,让 AI 代理(Agent,能自主执行多步任务的 AI 程序)搭了一个私有魔兽服务器;第二步,做了浏览器客户端,免装游戏;第三步,通过 MCP(Model Context Protocol,Anthropic 开源的“AI 调用工具”标准接口)让模型直接驱动游戏客户端,颗粒度比通用浏览器代理更细。

整个项目几乎不靠传统编程,作者用的方法是 vibe coding——用自然语言让 AI 自己写代码。

行业怎么看

支持者认为这是 MCP 生态的一次有意思扩展。MCP 自 2024 年底开源以来,从数据库、浏览器到 IDE,越来越多工具接入。这个项目证明 MCP 能驱动复杂游戏——意味着“工具调用”正在变成通用能力,而不是某个垂直场景的特例。

质疑声同样存在:魔兽服务器是封闭环境,状态可枚举、接口明确,真实业务里数据脏、流程断,复杂度高一个数量级。还有人指出,vibe coding 适合做原型,离生产部署还远。作者自己也说“还在 vibing”,这是周末项目,不是产品。

另一个被忽略的角度:游戏 AI 不是新鲜事,过去都是专门训练的小模型。这次是通用大模型 + 标准协议跑通,意味着同样的能力可以平移到客服、排程、运维等长链路任务。

对普通人的影响

对企业 IT:MCP 这类“AI 调用工具的标准协议”正在变成基础设施。IT 部门评估新系统时,“接不接 MCP”可能会像当年问“支不支持 REST API”一样变成标配问题。

对个人职场:vibe coding 让非程序员也能搭出可用的系统。这个项目一人几天搞定,传统要团队几周。提示词写得好的员工,在企业内部的话语权会变高。

对消费市场:AI 助理从“会聊天”走向“能动手”的拐点信号。一个能自己打魔兽的模型,距离能帮你订机票、处理表格的助理,技术上不再遥远。