大胆假设Bold bets, 并行推演parallel simulation, 固化交付shipped tools.
我们用投机策略让 AI 推演一切，把结果做成人人可用的工具。 像 speculative decoding 一样：先假设多条路径，并行验证，择优兑现。
We run speculative strategies so AI can reason about everything — then ship the results as tools anyone can use. Like speculative decoding: assume many paths, verify in parallel, keep the best.
SITES · 四个站
各有一件事要做
Four sites, four jobs
同一个团队，四种交付形态：一个规划中的基座、三个有证据的对比图谱、一份精选目录。
One team, four delivery forms: a base in planning, three evidence-backed atlases, one curated directory.
Keel
Keel
正在重新打造，与过去的 Keel 3D 完全无关。新方向确定后，会在这个页面更新。
Being rebuilt from scratch, unrelated to the old Keel 3D. This page updates as the new direction takes shape.
IDEAI 编程 IDE 图谱
IDE atlas
8 个编辑器与编码工具，按固定 8 维度对比。每条判断挂官方源，查不到就写「未知」。
Eight editors and coding tools on a fixed 8-axis schema. Every judgement links to a primary source; unknowns stay unknown.
CLI终端 AI 编码工具图谱
CLI atlas
6 个终端工具的非交互模式、CI 可用性与 Git 集成深度。
Six terminal tools compared on non-interactive mode, CI usability and Git integration depth.
MCPMCP 服务器图谱
MCP atlas
9 个 server 的权限范围、传输方式与输出可用性。第一个问题是安全，不是能力。
Permission scope, transport and output usability across nine servers. The first question is safety, not capability.
HARNESSAgent Harness 图谱
Agent harness atlas
要不要自己搭 Agent？10 套运行时的责任边界、状态持久化与权限模型。
Should you build your own agent? Ten runtimes compared on responsibility, state persistence and permission model.
MODELS量化模型图谱
Quantized models
同一模型几十个 GGUF 量化版，该下哪一个。按你的显存倒推，不做质量排名。
Dozens of GGUF quants per model. Pick by VRAM budget. No quality rankings.
一个词根，两种动作
One root, two moves
Specul 来自拉丁语「观察 / 推测」。在中文里我们故意保留张力：投机是策略，推演是方法。
From Latin specula — watch, infer. In Chinese we keep the tension: speculation as strategy, simulation as method.
投机
Speculation
借用大模型里的投机策略思路：在算力与时间有限时，并行押注多条假设，用后续验证淘汰错误分支。 我们用同样的方式选赛道、选技术、选产品——不是赌运气，而是结构化地下注。
Borrowed from speculative strategies in LLMs: when compute and time are scarce, bet on multiple hypotheses in parallel and let verification kill the wrong branches. We pick markets, stacks, and products the same way — structured bets, not luck.
推演
Speculate / Simulate
用模型、数据与可运行原型把世界往前推：推演玩法手感、推演工具形态、推演分发路径。 推演不是报告，而是能点开、能玩、能改的产物。
Use models, data, and runnable prototypes to push the world forward: playtests, tool shapes, distribution paths. Simulation is not a slide deck — it is something you can open, play, and fork.
从假设到工具
From bet to tool
投机假设
Speculative bet
用最小假设锁定问题：谁需要什么、哪里有结构性机会。
Lock a minimal hypothesis: who needs what, where the structural edge is.
AI 推演
AI simulation
让大模型与工具并行展开可能性，用可运行结果替代空想。
Let models fan out possibilities in parallel; replace opinion with runnable results.
固化交付
Ship the tool
把验证过的结论做成人人可用的基座与目录。
Freeze what survived verification into tools anyone can use.
现在可用的东西
What you can use now
Keel
Keel
Keel 正在重新打造，与过去的 Keel 3D 完全无关。新方向确定后，会回到这个页面更新。
Keel is being rebuilt from scratch, unrelated to the old Keel 3D. This page updates as the new direction takes shape.
看规划PlanAI 编程工具图谱
Tool atlases
IDE · CLI · MCP · Harness 四赛道 33 个对象，按固定 8 维度记录能力边界， 每条判断挂官方源，未知就写未知。
33 objects across IDE, CLI, MCP and Harness, recorded on a fixed 8-axis schema. Every judgement links to a primary source; unknowns stay unknown.
IDE 图谱IDE atlas CLI 图谱CLI atlas MCP 图谱MCP atlas Harness 图谱Harness atlas量化模型图谱
Quantized models
定了模型之后还有一关：同一模型有几十个 GGUF 量化版。 按你的显存倒推该下哪一档，体积是实测值，许可逐条标原文。
After picking a model, one more question: which GGUF quant to download? Resolved by VRAM budget, with measured sizes and per-quant licences.
Models 图谱Models atlas硅基导航
Directory
529 条精选：模型、编程、多模态、算力、机器人与开发工具，按 11 类组织。
529 curated entries across models, coding, multimodal, compute, robotics and dev tools, in 11 categories.
浏览目录Browse十一个品类，直达分类
Eleven categories, one click in
开始你的推演
Start your own simulation
从一个大胆的假设开始，让 AI 帮你并行验证。
Start with a bold hypothesis; let AI verify it in parallel.