从一个概念，或一件想完成的事开始Start with a concept or a task you want to finish.

8 个内容站Eight content sites ，按「先懂再做、按需深挖」排好；另有一个去别的 AI 网站的出口。卡在哪儿就翻哪儿，不卡的地方跳过也完全没问题。 , ordered from first steps to deep dives, plus one gateway to other AI sites. Look up whatever you are stuck on; skip the rest, that is fine.

名词看不懂，我们先给你铺个路Terms made no sense — here is a path in

236 个词按「先学什么后学什么」排成六站，每个配一句大白话和官方出处。今天挑用得上的看两三个就够，不用一次看完。236 terms ordered by what to learn first, each with a plain-language explanation and its primary source. Read two or three today; you do not have to finish it in one go.

学 AILearn AI 236 词 · 六站236 terms · 6 stations

工具挑不动，我们摊开给你对比Too many tools to choose from — laid out side by side

262 份主流工具用同一套标准过一遍：各自能做什么、装在哪、权限多大。分数我们不替你排，你按自己需求挑就行。262 mainstream tools run through one yardstick: what each can do, where it installs, how much permission it wants. We do not rank them for you — pick by your own needs.

工具Tools 262 份 · 同一套标准262 entries · same yardstick

想在本地跑模型，我们帮你算清楚Want to run models on your own machine — here is the arithmetic

11 个系列、16 个模型仓、129 个量化仓、1262 个量化档。先按文件体积筛候选，再核对运行器、上下文、并发和可用内存；许可回到发布页确认。11 series, 16 model repos, 129 quant repos and 1262 quantised builds. Filter file-size candidates, then check the runtime, context, concurrency and available memory. Confirm licences at their sources.

模型Models 1337 个 GGUF 文件1337 GGUF files

想去别的 AI 网站？从这里按类目出发Want to visit other AI sites? Start by category

这里收录的是其他 AI 网站，不是 Specul 自己产出的内容。592 个 AI 网站按 11 类归好，每条一句话说清「做什么」。链接核验快照（2026-10-29）覆盖 700 个独立域名：608 个确认可达（539 直接 200、62 个反爬拦截但站点响应、7 个改址仍可达），92 个因探测环境限制未能确认，0 个确认失效；之后新增的条目尚未纳入该快照，卡片上标为待核验，不冒充全部可用。下方 11 个类目可以直接点进去。This directory lists other sites, not Specul's own content. 592 AI sites filed into 11 categories, one line each on what it does. The link-check snapshot (2026-10-29) covers 700 unique domains — 608 confirmed reachable (539 direct 200, 62 anti-bot but responding, 7 redirected and reachable), 92 unconfirmed because the probe environment was blocked, 0 confirmed dead. Entries added after that snapshot are marked as unchecked rather than passed off as working. The 11 chips below jump straight to a category.

外部站点 ↗External sites ↗ 592 条 · 11 类592 links · 11 cats

前沿资料站：规格、论文与链上凭证Frontier reference sites: specs, papers and on-chain credentials

作品库Demos→ 策展作品与许可台账→ Curation and licence ledger 硬件算力Compute & hardware→ 设备规格与选型路径→ Device specs and selection path AI 论文AI Papers→ 主张边界与复现状态→ Claim boundaries and reproduction status 链Chain→ 凭证、溯源与授权计算→ Credentials, provenance and authorised compute

还是对不上？换个说法找找看Still nothing matching? Try another way of asking

「训练」和「推理」老是分不清I still cannot tell training from inference → 学 AI→ Learn AI 在用了，说不清它到底动了我的什么I use one but cannot say what it touches → 工具→ Tools 怕东西传出去，或者账单按住不住Worried about data leaving, or a runaway bill → 模型→ Models

按主题串起来Explore by topic一页把全站资料排成可走的顺序One page puts material from every site in order

用 AI 写代码Write code with AI→ 选工具、改代码、复核结果→ Choose, change, check 用 AI 做游戏Make a game with AI→ 五阶段与素材授权→ Stages and asset rights 用 AI 做视频Make video with AI→ 分镜、素材与导出→ Shots, media, export 用 AI 写音乐Make music with AI→ 音频工具、接口与许可→ Audio tools, APIs, licences 用 AI 做研究Do research with AI→ 提问、溯源、逐条核对→ Question, source, verify 用 AI 处理办公任务Handle office work with AI→ 样例、工作表与留痕→ Samples, sheets, records 用 AI 做机器人Build robots with AI→ 软件栈、模型与目录→ Software, models, directory 写提示词与工作流Prompts and workflows→ 指令、上下文与复核→ Instructions, context, review

按任务直接进Jump in by task按任务读流程、看例子、核对结果Follow a workflow, read an example, check the result

也可以按标题、别名与任务搜索全站 →Or search titles, aliases and tasks across the whole site →

把 PDF / 扫描件转成表格Turn a PDF or scan into a table→ 样例与校验→ Samples and checks 合并多张表，找重复与冲突Merge tables and find conflicts→ 输入与异常单→ Inputs and exceptions 审阅文档的版本变化Review document changes→ 原文与问题表→ Sources and findings 把录音整理成纪要与行动项Turn audio into notes and actions→ 合成录音与核对→ Synthetic audio and checks 修一个真实 bugFix a real bug→ 六步流程→ Six steps 做一条 60 秒视频Make a 60-second video→ 五段流程→ Five stages 做一份可核查的研究简报Produce a checkable research brief→ 五步流程→ Five steps 搭一个客服 AgentBuild a service agent→ 六段流程→ Six stages 在本机跑一个模型Run a model locally→ 五步流程→ Five steps 做一次上线前安全与合规评审Run a pre-launch safety and compliance review→ 十三条路线之一，12 步→ One of thirteen routes, 12 steps 搭一条能自己复跑的评测流程Set up an evaluation flow you can re-run→ 十三条路线之一，11 步→ One of thirteen routes, 11 steps

去别的 AI 网站 · 按类目Other AI sites, by category11 个类目，点进去直接看那一类11 categories, each jumping straight to that section of the external directory

大模型与对话LLMs & chat48 AI 编程与 AgentAI coding & agents43 多模态生成Multimodal generation64 AI 办公与垂直应用Applications47 模型与权重Model providers32 开发框架与工具Developer tools47 算力与芯片Infrastructure55 机器人与具身智能Robotics & embodied87 学习资源Learning55 社区与资讯Community25 开发者工具Tools & utilities89

以上类目来自外部站点目录，收录的是其他网站，不是 Specul 自己产出的内容。These categories come from the external directory: other people's sites, not Specul's own content.

最近新增什么What has been added2026-11-06

四类任务有文件可练Four tasks with downloadable practice

PDF 转表格、多表合并、文档审阅与会议整理，附虚构输入、参考答案和验收方法。先处理输入，再对照结果。PDF extraction, table merging, document review and meeting notes include fictional inputs, designed references and acceptance checks. Try the input before opening the answer.

下载样例，选一个开始 →Choose a sample task →

练习卡住，按原因继续Continue from the point that failed

六个情境区分概念、输入与验证问题，展开后看判断理由和下一步。打开页面不算学会，练习不作能力认证。Six scenarios separate concept, input and verification problems, with reasoning and next steps. Opening a page is not completion; practice does not certify competence.

做一个判断练习 →Try a diagnostic scenario →

预算列清假设与单位Budgets with explicit assumptions and units

成本与容量可填写、复算、下载假设。文件大小不等于运行内存；页面功能检查与文档核对不代表工具或模型能力实测，各页单列核验范围。Fill, recalculate and export cost/memory assumptions. File size is not resident memory. Page checks and documentation review are separate from product/model capability tests; each page states its scope.

算容量 →Budget memory → · 算费用 →Estimate cost →

安全与评测单独成类Safety and evaluation, its own category

工具站新增「AI 安全与评测」大类：护栏与内容过滤、评测与观测、红队与注入检测、网关与访问控制、内容溯源与标识。每份档案只写官方在册的能力，取不到的写未知，不排总分。Tools gains an "AI safety & evaluation" category: guardrails and filtering, evaluation and observability, red-teaming and injection scanning, gateways and access control, provenance and labelling. Entries record only what official pages state; gaps stay marked unknown, and there is no composite score.

看这一类 →Browse the category → · 上线前评审路线 →Pre-launch review route →