Posts on Marginalia

Five Kettlebell Moves on Three Axes

Thu, 12 Nov 2026 11:30:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. Five movements land on three axes. The hips produce the power, the shoulders buy the range, the diagonal carries force across the trunk. The snatch sits where all three meet, which is why it is the one you cannot swap out. The film runs 26 seconds with no voice-over. The labels on screen are the language of this page.

No Vector Store: a 25.87 MiB SQLite Index Reaches 86.62 Recall@10

Wed, 11 Nov 2026 14:12:52 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. The semantic call can wait until after the question arrives. Keep words and counts in the index, and MuSiQue R@10 moves from 54.41 to 86.62 on a 2.36-second build with no model calls. The paper is SAG2: Embedding-Free Multi-Hop Retrieval with BM25 and JEV, by Yuchao Wu at Zleap AI, arXiv 2610.

Designing a Coding Agent Without a KV Cache

Wed, 11 Nov 2026 07:16:58 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. A design note from September 2026 gathers six coding-agent complaints under one sentence. Suppose a language model had no KV cache. What would these agents look like then. The answer contains no stronger model. It contains one thing, someone deciding what goes in front of the model on every turn.

Multi-teacher distillation is a token budget

Tue, 10 Nov 2026 07:02:51 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. TRL’s multi-teacher on-policy distillation example ran 60 steps. Instruction following held 20 percent of the prompts and got 1 to 3 percent of the trained tokens, and it was the only domain that barely moved. Open-MOPD calls this token budget misallocation, and fixing it raised the recoverable headroom from 35.

The Ontology Divide Is Not Model Size

Mon, 09 Nov 2026 11:08:09 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. Between December 2025 and November 2026, six papers closed in on one conclusion from three directions. The bottleneck in ontology construction moved from who can write an ontology to who can prove the written thing is right. Every number below goes back to a named section or table. Two minutes twenty-three, voiced in English, captions burned in and switchable in the player.

Evals: Four Numbers That Tell You Whether an AI System Actually Works

Mon, 09 Nov 2026 07:36:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. An eval spends most of its budget on reading outputs and labelling cases. Code is the cheap part. This piece splits the work into four parts and gives each part a number you can recompute. The film runs one minute forty-four, voiced in English, with the captions burned in and switchable in the player.

The firm grows a basement

Sun, 08 Nov 2026 12:40:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: two people of twenty-four start managing teams whose staff are machines. All three arrangements are already happening, and nobody has filled in the column for the extra floor. 1. Two people of twenty-four# Every org chart I have seen is printed on A4, landscape, three ranks of boxes squeezed into the height.

Retrieval as Navigation: DecisionRAG Asks a Tree Instead of Scoring a Corpus

Sun, 08 Nov 2026 12:06:43 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. The bottleneck in retrieval is not the size of the candidate pool. It is what you let the model see, and in which order. DecisionRAG replaces scoring with asking, and returns a 65.4 top-1 hit over 2.68M passages in 1.7 seconds. It also loses three cells, plainly. English voiceover, 3 minutes 12 seconds.

One Graph, Fifteen Spellings, a 28-Point Difference

Sun, 08 Nov 2026 12:01:41 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line. How a graph is spelled into a prompt is a measurable system parameter. It is worth 1.3 to 3.9 points on ontology extraction and 28.4 points on structural reasoning, and the three reasoning tiers of a single model family each prefer a different format, with the most accurate tier burning the fewest tokens.

The lamp in the stairwell

Sat, 07 Nov 2026 18:02:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: attention is the one thing nobody can print more of, so watching became a trade, and both the watcher and the watched end up wondering how much of themselves is still theirs. The stairwell in my building has a lamp that answers to noise. On the way up, at the third landing, you stamp a foot or clap once and it comes on.

The Bytes Moved, the Compute Didn't: Ten Years of Path Dependence in Decentralized Databases

Sat, 07 Nov 2026 12:31:38 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: the field’s unsolved problem is not that consensus is too weak, it is that compute never followed the bytes. A database is not one thing — it is four: write order, byte location, execution location, read permission. Six routes, and each one bought exactly one of them.

Project EVE: What Happens When Agents Enter an Economy That Already Has One

Sat, 07 Nov 2026 11:25:06 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: this is not a review of “AI playing a game”. It is the first experiment that splits an agent economy into three billable layers — execution, tooling capital, and the cost of trust — and quotes a price for each: four months of model progress buys 3× the return; swapping the author of a scaffold moves a fleet from losing ships daily to zero losses in twelve hours; and a saboteur who is told to sabotage is almost never caught.

Allegedly the singularity has begun: three sums from an old-fashioned man

Sat, 07 Nov 2026 09:40:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: allegedly we are already inside the singularity. I will not argue about the word. I want to redo the three sums the argument rests on — who writes the code, how cheap intelligence has become, and what is left in the human column. A note on method: this is a reading note.

The four-layer memory stack, and the dreaming loop beside it: reading Anthropic's memory research

Fri, 06 Nov 2026 22:12:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: memory is not a transcript you keep, it is cargo moved between four layers — the low ones are true and hard to use, the top one is easy and easy to get wrong. Dreaming is the worker that moves it while nobody is watching. Note on figures: the diagrams here are redrawn from the article’s own data and structure; its fifteen-odd figures and screenshots are at the source linked at the end.

One hour a day, 365 hours a year: which half of Dan Koe's deep focus actually lands

Fri, 06 Nov 2026 11:52:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: the useful thing in the essay is not “use your phone less” — it is a budget you can compute: one hour a day, 365 hours a year, 45 full working days, all of them spent on your own project. Scope first. The source is Dan Koe’s long post How to unlock extreme focus on command, published on X on 19 January 2026; I read the full body.

The Database Got a New User: Record Stores and Context Layers for Agents

Fri, 06 Nov 2026 11:30:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: the database’s main user stopped being a person, so it grew a second leg — one that holds state and provisions on demand, one that supplies context and compiles it ahead of the query. What really changed is the definition of retrieval: from a recall algorithm to an orchestration frame plus a compiled supply.

Seven layers between the model and the result: reading the guide against its own evidence

Fri, 06 Nov 2026 09:22:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. Abstract. Most production agent failures do not happen in the model layer. Split the distance from model to result into seven layers — model, instructions and context, tools, memory, loop, evals and guardrails, human signature — and the most common reflex (swap the model, rewrite the prompt) acts on the one layer least likely to be broken.

How one prompt grew into a company brain: Recall in four generations

Thu, 05 Nov 2026 22:42:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: it turns “remembering” into a retrieval contract — the source is stored upstream and never edited, the index and sandbox are rebuildable projections, and the server owns recall while precision and citation go back to the frontier model that asked. Source material: the author’s long post, plus the open skill repository (the engine is not open source; the skills are, MIT, 13 of them).

Reading DSec section by section: what a 3-million-sandbox-a-day platform says about agent training

Thu, 05 Nov 2026 15:28:40 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: the contribution is not another sandbox runtime. It is a measurement of what agent-training sandboxes actually look like — requests arrive in tens of thousands, CPU usage sits at 5% of what was requested, state lives for minutes to hours, and images are barely reused. Once all four hold, the platform cannot be a container runtime; it has to be an elastic execution platform.

Open Responses: one contract instead of N×M adapters

Thu, 05 Nov 2026 08:22:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: it swaps “messages” for bidirectional items, swaps token streaming for state-machine events, and gives extensions a namespaced side channel — so the endpoint and model name become configuration, and the cost moves from adapters to the semantics of individual fields. Read from three places: the specification itself (both the 2026-01-15 and 2026-04-24 versions), its TypeSpec schema and repository, and the docs of two live implementations at Databricks and Vercel.

Life is meaningless, the things you love are a con, and anger takes strength

Wed, 04 Nov 2026 21:50:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: meaninglessness is not bad news, and neither is the con. What has to be budgeted is strength — anger is physical work, and there is less of it every year. 1. Uncover the floor first# tap to enlarge Life has no meaning, and it arrived by accident.

A chat box is not an agent OS: five places a 30-person company breaks it

Wed, 04 Nov 2026 21:32:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: a chat box on a single-user harness moves the interface and nothing else. The difference is five things — who is speaking, what it remembers, what survives the process, which machine it acts on, and whether the reply arrived. First, the material boundary. The article’s body is behind an X login; what I could read is its own claim plus the opening paragraph.

Automatic Ontology Engineering: Reading NVIDIA's Auto-Ontology at the Source

Wed, 04 Nov 2026 20:41:07 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: Auto-Ontology automates the most expensive month of ontology engineering — it reads your metadata and query history, compiles a business ontology, and hands it to agents through nine read-only tools. What it hands over is a draft: in this codebase, certification is a badge rendered for display, not a gate on the query path.

One person as a creative team: what transfers from Dan Koe's letter

Wed, 04 Nov 2026 07:20:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: the part worth carrying away is the job split. It cuts the creative director role into brand strategy, content strategy, and writing plus ops. The price tags, the product plugs and the growth numbers belong to the sales side of the letter. Dan Koe published a long letter in November arguing that the most lucrative skill has always been capturing attention.

Cognitive surrender: hand off the work, not the judgement

Wed, 04 Nov 2026 06:54:07 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: handing over the work is fine. Handing over the judgement is a different act, and the slope between the two is quiet. English narration, 2 min 41 s. The Chinese cut runs 2 min 33 s with the same scenes and the same numbers. Addy Osmani published a long piece on X in November about keeping critical thinking while working with agents.

Think Before You Act: How JevAct Gates LLM Agent Actions

Tue, 03 Nov 2026 12:27:14 +0800

中文版是这篇文章的默认版本。阅读中文版 This English page stays available at its own URL. The Chinese version is the one maintained first. English readers land here directly. To pin English after switching languages, use ?lang=en on this URL, e.g. this link. An agent’s action should pass through a gate that can check, compute, and send back before anything executes. JevAct splits that gate in two. code settles what tool records decide, a decision model answers only “what did the conversation actually say”, and a boolean rule combines them.

SecLog, read closely: letting a delivery leave its own receipt

Mon, 02 Nov 2026 20:30:28 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: “the user cannot deny receiving the data” needs no trusted API and no data on chain — only a design where the last half of the decryption key can appear through one path, and that path leaves a record. Just over two minutes: how a key becomes a log entry → the two events and what each one proves → the 8,000-gas gap, and why I could not reproduce the absolute number.

Why Go Is the Default Language of the Agent Era

Mon, 02 Nov 2026 10:00:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: once AI drives the cost of writing code to near zero, a language’s value moves from how it reads to how it is reviewed — and every trade-off Go made in its first fifteen years sits on the review side. The two source posts landed in the same month, from different authors, different starting points, and the same conclusion.

Super-agent: DeepSeek Takes Sessions Away From the User

Mon, 02 Nov 2026 06:30:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: the interesting claim in DeepSeek Harness’s architecture note is not its five modules — modules change with deployment. It is one sentence: sessions should be an implementation detail of the system, not a mental burden on its user. Everything below is that sentence, taken seriously. Published 2026-10-31, the note opens by telling you what it is not: an architecture discussion, not a description of shipping features.

A Decentralized, Trusted MCP Runtime: Architecture and Implementation

Sun, 01 Nov 2026 12:00:00 +0800

中文版是这篇文章的默认版本。阅读中文版 → This English page stays available at this URL; the Chinese version is the one we maintain first. In one line: treat a tool as a content-addressed artifact, run it inside a sandbox, and keep an identity behind it — only then does “trusted” mean anything. That is the architecture; the second half of this article is the implementation. 1. The problem: an agent’s tools are black boxes# An agent’s capability is not bounded by its model.

The Architecture of Trust: A Research Report on Web3.0 Database and Storage Systems

Thu, 18 Sep 2025 23:04:12 +0800

tap to enlarge Section 1: Introduction - The Paradigm Shift from Centralized to Decentralized Data# The evolution of the internet can be demarcated into distinct architectural and philosophical eras. The current dominant paradigm, Web 2.0, is defined by interactive, user-driven platforms that have connected the globe. However, this model is predicated on a fundamental asymmetry: while users create the content and data, centralized corporations own the platforms, control the data, and capture the vast majority of the economic value.

The Vector Database Landscape: An In-Depth Analysis of Architectures, Technologies, and Strategic Applications

Mon, 25 Aug 2025 22:05:12 +0800

tap to enlarge Section 1: The Rise of Vector Databases in the AI Era# 1.1 Beyond Traditional Databases: The Unstructured Data Challenge# The landscape of data management is undergoing a fundamental transformation driven by the proliferation of artificial intelligence and machine learning. For decades, the dominant paradigm has been the relational database, which excels at storing and querying structured data—information neatly organized into tables, rows, and columns with a predefined schema.

A Research Report on the Golang Green Tea Garbage Collector

Sun, 17 Aug 2025 14:15:12 +0800

tap to enlarge I. Executive Summary The Green Tea garbage collector (GC) is an experimental, memory-aware parallel marking algorithm for the Go runtime. It represents a fundamental redesign of the garbage collection process, shifting from an object-centric to a memory-block-centric approach to combat performance bottlenecks caused by poor memory locality in modern many-core systems. This report analyzes its design, performance characteristics, and strategic importance. The core finding is that Green Tea successfully addresses its primary goal: reducing GC CPU costs and improving scalability by enhancing memory access patterns.

WarmStorageChain: Revolutionizing Decentralized Data Retrieval

Mon, 15 Jul 2024 13:15:12 +0800

tap to enlarge I. The Problem with Decentralized Storage Challenges in Decentralized Storage Systems 1. Increased Latency 2. Cold Storage Issues 3. Data Integrity and Security 4. Transparency and Trust 5. Scalability and Performance The Need for an Enhanced Solution Addressing Increased Latency Overcoming Cold Storage Issues Ensuring Data Integrity and Security Enhancing Transparency and Trust Achieving Scalability and Performance Enhancing Transparency and Trust Achieving Scalability and Performance Comprehensive Benefits II.

2022 Wanxiang Blockchain Spring Hackathon - IceFireDB Won The First Prize

Tue, 14 Jun 2022 12:55:18 +0800

tap to enlarge 2022 Wanxiang Blockchain Spring Hackathon Wrapped up on June 12th,We are very lucky and grateful to Protocol Labs for giving us this honor. In addition to joy, we are more moved. Our efforts and innovation are recognized. Next, we will work harder to incubate our projects. 2022 Wanxiang Blockchain Spring Hackathon, an immensely successful hackathon packed with over 20 teams of passionate builders and developers wrapped up yesterday, under the theme of “Metaverse: A Shared Future on Blockchain”.

Open source: NetHandle, a high-performance network listener for Go

Wed, 02 Jan 2019 13:39:18 +0800

NetHandle A library for building high-performance TCP servers, standing on the shoulders of Go. Repository: https://github.com/gitsrc/NetHandle Features# High throughput, low resource use. A very small API surface to learn. TCP, UDP and UNIX sockets. Install# go get -u github.com/gitsrc/NetHandle Benchmarks# 50 threads × 10,000 requests# tap to enlarge tap to enlarge 50 threads × 20,000 requests# tap to enlarge 100 threads × 10,000 requests# The harness and the raw output are in the repository; the numbers above are from a single machine, so treat them as a relative comparison rather than a capacity claim.

高级编程：PHP扩展的 INI 配置文件操作

Thu, 22 Nov 2018 10:24:18 +0800

Extension INI Entries (扩展INI配置项) Defining php.ini directives (i.e., INI entries) in an extension is easy. Most of the work involves setting up the global struct explained earlier in Section 14.10.3 Each entry in the INI structure is a global variable in the extension and thus has an entry in the global struct and is accessed using FOO_G(my_ini_setting). For the most part you can simply comment out the indicated sections in the skeleton created by ext_skel to get a working INI directive, but we will walk through it here anyway.

Nginx FAST-CGI caching, and what it bought

Sun, 24 Jun 2018 20:42:18 +0800

Background In a classic LNMP stack, PHP-FPM parses every PHP script and Nginx forwards each PHP request to it. For a busy PHP site that puts a lot of pressure on PHP-FPM — so a FastCGI cache at the Nginx layer removes most of that pressure and improves response time, at the cost of serving slightly stale pages. Nginx cache configuration 1. HTTP block# fastcgi_cache_path /tmp/cache/fastcgi_cache levels=1:2 keys_zone=fastcgi-cache:10m inactive=2d max_size=512m; fastcgi_cache_key $scheme$request_method$host$request_uri; proxy_buffering on; proxy_cache_valid any 10m; proxy_temp_path /tmp/cache/tmp; proxy_buffer_size 4k; proxy_buffers 100 8k; fastcgi_cache_lock on; fastcgi_cache_use_stale error timeout invalid_header updating http_500; fastcgi_ignore_headers Cache-Control Expires Set-Cookie; keys_zone=fastcgi-cache is the zone the server block below refers to.

源码阅读：C语言epoll模型

Fri, 08 Jun 2018 18:02:18 +0800

背景 epoll是Linux内核为处理大批量文件描述符而作了改进的poll，是Linux下多路复用IO接口select/poll的增强版本，它能显著提高程序在大量并发连接中只有少量活跃的情况下的系统CPU利用率。另一点原因就是获取事件的时候，它无须遍历整个被侦听的描述符集，只要遍历那些被内核IO事件异步唤醒而加入Ready队列的描述符集合就行了。epoll除了提供select/poll那种IO事件的水平触发（Level Triggered）外，还提供了边缘触发（Edge Triggered），这就使得用户空间程序有可能缓存IO状态，减少epoll_wait/epoll_pwait的调用，提高应用程序效率。 #模型代码今天给大家分享一下EPOLL的相关模型代码：服务端代码# #include <sys/types.h> #include <sys/signalfd.h> #include <sys/epoll.h> #include <errno.h> #include <poll.h> #include <signal.h> #include <limits.h> #include <stdio.h> #include <stdlib.h> #include <unistd.h> #include <fcntl.h> #include <string.h> #include <assert.h> #include <sys/ioctl.h> #include <sys/socket.h> #include <netinet/in.h> #include <arpa/inet.h> #include "utarray.h" /***************************************************************************** * This program demonstrates epoll-based event notification. It monitors for * new client connections, input on existing connections or their closure, as * well as signals. The signals are also accepted via a file descriptor using * signalfd.

Manjaro 安装 LNMP环境

Sat, 24 Mar 2018 13:03:18 +0800

安装PHP、PHP-FPM ; 编译安装 Nginx、MYSQL; 详细配置信息。#

1. 编译Nginx#

1.1 创建用户及用户组#

sudo groupadd -r www
sudo useradd -s /sbin/nologin -g www -r www

1.2 下载Nginx源码包并解压#

wget http://nginx.org/download/nginx-1.14.0.tar.gz