Money-Burning-alarm-colok
2026
Burn Alarm · 会烧钱的闹钟
Burn Alarm · The Alarm That Spends Your Money
闹钟响后 60 秒内解不出三道心算题，就用你的 LLM API 额度烧钱
Fail three mental-math problems within 60 seconds and your own LLM API starts burning money
- 闹钟响起锁定屏幕，60 秒倒计时
- 未解出则调用你自己的 LLM API，以最大
max_tokens 持续请求，实时显示累计花费 - 多种防护：额度上限、失控保护、幂等计费
- 中英双语界面，适配 Android 15 后台启动限制
- Full-screen lock-screen alarm with a 60-second countdown
- On failure it hammers your own LLM API at maxed-out
max_tokens, showing the running cost - Multiple guards: spend caps, runaway protection, idempotent billing
- Bilingual UI, adapted to the Android 15 background-activity limits
把一个行为经济学问题（起床拖延）直接翻译成经济损失。
Translates a behavioural problem (not getting up) directly into a financial loss.
Cabinet_BNN
2026
Cabinet-BNN · 从语义码生成权重
Cabinet-BNN · Generating Weights from a Semantic Code
二值激活语言模型，每个 token 的权重由 64 位语义哈希码生成，而不是从全局权重张量里查表
A binary-activation LM whose per-token weights are generated from a 64-bit semantic hash code instead of looked up from a global tensor
- 每个 token 一个 128 位权重码：
hash(64) 取自 HSH-64，param(64) 承载该 token 的独有信息 - 权重由共享低秩基矩阵合成：
W_t = Σ_j s_j(t)·(U_j V_jᵀ)，不存在全局权重 - 二值激活
sign(x) + STE，权重保持浮点 - MoE 稀疏 FFN：码驱动路由，零路由参数、负载天然均衡
- 码即索引：码的汉明距离与权重余弦相似度相关系数 −0.99，无需训练即可传导几何
- 如实记录负面结果：双向关系门实测有害
- A 128-bit weight code per token:
hash(64) from HSH-64, param(64) carrying token-specific information - Weights are synthesised from shared low-rank bases:
W_t = Σ_j s_j(t)·(U_j V_jᵀ) — no global weight tensor - Binary activations
sign(x) with STE; weights stay floating point - MoE-style sparse FFN with code-driven routing: zero router parameters, structurally balanced load
- Code-as-index: Hamming distance correlates with weight cosine at −0.99, transferring geometry with no training
- Negative results documented: the bidirectional relation gate is measured to hurt
把 HSH-64 的码从「用来检索」推进到「用来生成参数」——这是整条线的下一站。
Takes the HSH-64 code from retrieval to parameter generation — the next stop on this line.
Cabinet_hsh64
2026
HSH-64 · 64 位可学习语义哈希
HSH-64 · 64-bit Learnable Semantic Hashing
把语义码从 20 位扩展到 52 位，单个 u64 存储、单次 popcnt 比较
Pushing the semantic code from 20 to 52 bits — one u64, one popcnt
- 结构化 64 位编码：
feat(4) + sim(52) + abs(8)，单 u64 存储 - 三阶段端到端训练：连续预训练 → STE 离散精调 → 召回导向贪心后处理
- 自适应多索引哈希（MIH），动态半径扩展
- 非对称距离评分，缓解符号量化的信息损失
- 纯 CPU 运行，训练与推理均无需 GPU
- 有论文与 Zenodo DOI
- Structured 64-bit code:
feat(4) + sim(52) + abs(8), one u64 - Three-stage training: continuous pretrain → STE fine-tune → recall-oriented greedy post-processing
- Adaptive multi-index hashing with dynamic radius expansion
- Asymmetric distance scoring to reduce sign-quantisation loss
- CPU-only — neither training nor inference needs a GPU
- Paper and Zenodo DOI available
本仓库是整条技术线的顶端：把前作的 20 位方案扩展到 64 位。
The top of this technical line: extends the earlier 20-bit scheme to 64 bits.
hakimi-codex
2026
Hakimi Codex · 多智能体编程助手
Hakimi Codex · Multi-agent Coding Assistant
基于工具调用架构的 CLI 编程助手，能读写文件、执行命令、协作完成开发任务
A CLI coding assistant built on tool calling — reads, writes, executes, collaborates
- 严格的工具调用协议（JSON over Markdown code blocks），保证稳定性
- 可配置 LLM 后端：切换 API 端点、模型与认证信息
- 流式响应，实时输出处理进度
- 多工具集成：读写文件、目录列表、文件搜索、命令执行
- Strict tool-calling protocol (JSON over Markdown code blocks) for reliability
- Configurable LLM backend: swap endpoint, model and credentials
- Streaming responses with live progress
- Multi-tool integration: file I/O, directory listing, search, command execution
把 LLM 的「说」变成「做」——这是 Agent 与聊天机器人的分界线。
Turns what an LLM says into what it does — the line between an agent and a chatbot.
Cabinet
2026
Cabinet · 面向 AI Agent 的语义记忆检索
Cabinet · Semantic Memory Retrieval for AI Agents
用 20 位结构化整数替代 768 维浮点向量，纯 CPU、可解释、可增量更新
Replacing 768-dim float vectors with 20-bit structured integers — CPU-only, explainable, incremental
- 针对现有 RAG 的三个结构性缺陷：不可解释、更新成本高、依赖硬件
- HSH-20 编码：
feat(4) + sim(8) + abs(8)，检索路径完全可审计（类别→簇→词） - 仅追加写入 + 后台 LSM 合并，新增文档无需重建索引
- 索引体积缩小约 1000 倍，笔记本即可运行
- Rust 核心 + Python 绑定 + 独立 GUI
- Targets three structural flaws in current RAG: opacity, expensive updates, hardware dependence
- HSH-20 code:
feat(4) + sim(8) + abs(8); the retrieval path is fully auditable (category → cluster → word) - Append-only writes plus background LSM merge — no index rebuild on insert
- Index size reduced by roughly 1000x; runs on a laptop
- Rust core, Python bindings, standalone GUI
这是 HSH-64 的前作，提出了层次语义哈希这套编码思路。
The predecessor of HSH-64, where the hierarchical semantic hashing idea was introduced.
Lixivium
2026
Lixivium · 可视化数据清洗
Lixivium · Visual Data Cleaning
拖拽连线配置清洗流程，自动生成 Pandas / SQL / PySpark 代码
Drag, connect, configure — auto-generate Pandas / SQL / PySpark code
- React Flow 画布，拖拽式节点编排
- 一次编排，输出三种代码：Pandas / SQL（CTE）/ PySpark
- DuckDB 驱动的实时预览
- Monaco 编辑器 + 50 步撤销栈
- React Flow canvas with drag-and-drop node orchestration
- One graph, three outputs: Pandas / SQL (CTE) / PySpark
- DuckDB-powered real-time preview
- Monaco editor with a 50-step undo stack
把重复的清洗脚本变成可视流程——写一次，出三种引擎的代码。
Turns repetitive cleaning scripts into a visual graph — write once, emit three engines.
MindMath_Studio
2026
MindMath Studio · 科学计算平台
MindMath Studio · Scientific Computing Platform
数据清洗 + 方程建模 + 数值求解 + 可视化，一站式 Web 平台
Data cleaning, equation modelling, numerical solving and visualisation in one web platform
- 代数方程 / ODE / PDE 交互式求解（有限差分法）
- 50+ 预设方程库，支持自定义方程与参数映射
- 缺失值与异常值处理（IQR / Z-score）
- 磨砂玻璃质感界面，支持实时曲线预览
- Interactive solving of algebraic equations, ODEs and PDEs (finite difference)
- 50+ preset equation library, plus custom equations and parameter mapping
- Missing-value and outlier handling (IQR / Z-score)
- Frosted-glass UI with live curve preview
技术栈从数据处理起步的地方，也是后来 Cabinet 那条线的源头。
Where the data-processing stack started — and the origin of the line that led to Cabinet.
-
2026
简易数据建模软件
Simple Data Modelling Tool
导入 CSV、自定义方程、快速建模 —— 这条技术线的起点
Import CSV, define equations, model quickly — the start of the line
- CSV 导入与方程自定义
- 仓库名暂为
-，无 README
- CSV import with custom equations
- Repository is currently named
- and has no README
最早的仓库。名字还是个占位符，但它是后面所有工作的起点。
The earliest repository. Its name is still a placeholder, but everything after starts here.