# AI TRY — English > A personal site by Moon Hyeonggeun, a QA engineer of twelve years who does not write the code — the AI does — and spends his time finding where it breaks. It holds tools he built and runs daily, teardowns of other people's repos and services, 114 written pages in all, and a machine-collected AI news feed. Before reading the list — - **Every document has two layers.** The top is a plain-language explanation written for someone who knows nothing about the subject. **Below** the heading "The detailed record starts here" is the original analysis. Stop at that heading if you need the gist; everything under it is the untouched record, with the evidence and the numbers. - **Korean is the original.** These English pages are machine-assisted translations of the Korean ones (`/ko/…`), so where the two disagree, the Korean is authoritative. Swap the language segment and the URLs pair up. - **One of the five sections is not written by hand.** "AI news today" is rebuilt every morning by a collector. Its cards have no permanent URLs of their own and disappear within days — **cite the category pages, not the cards.** - **The document links below point at the markdown twin** (`…/index.md`). The page a human reads is the same URL without `index.md`; the markdown is that page with the chrome stripped and every diagram replaced by its description. Section indexes, the news feed and the glossary have no markdown twin, so those links are HTML. - What the sections mean — - **My tools** — Tools I made and run every day. - **What I took apart** — Repos and services I took apart — and what I kept. - **Things I wrote** — Notes that outgrew a commit message. - **Who I am** — Who is doing all this, and why. - **AI news today** — What turned up in the last three days — collected by machine, published by hand. ## My tools - [My tools](https://nobles92ts-ship-it.github.io/en/built/): Tools I made and run every day. - [/watch](https://nobles92ts-ship-it.github.io/en/built/watch/index.md): Hand it a video and you get back a buildable design spec, not a summary. - [/watch · Frame budget](https://nobles92ts-ship-it.github.io/en/built/watch/budget/index.md): Almost all the cost is frames — and the budget triples while the length grows sixtyfold. - [/watch · Transcript](https://nobles92ts-ship-it.github.io/en/built/watch/transcript/index.md): Having captions and not having them are entirely different routes. And what leaves the machine is audio, never video. - [/watch · Boundary and failure](https://nobles92ts-ship-it.github.io/en/built/watch/boundary/index.md): A tool handling other people's videos states what leaves the machine. And when something can't be done, it says so. - [Self-evolution loop](https://nobles92ts-ship-it.github.io/en/built/self-evolution/index.md): A cron job that reads what I learned this week and rewrites its own memory. With a human gate. - [Self-evolution loop · The dead channel](https://nobles92ts-ship-it.github.io/en/built/self-evolution/dead-channel/index.md): The sophisticated one had been dead for months — and the document diagnosing it was wrong in two places. - [Self-evolution loop · Four tempos](https://nobles92ts-ship-it.github.io/en/built/self-evolution/tempo/index.md): Treating learning as one block and automating the whole block was the real cause of the over-design. - [Self-evolution loop · The detector](https://nobles92ts-ship-it.github.io/en/built/self-evolution/detector/index.md): The thing built to detect death false-positived first. And the first full run changed nothing. - [tc-team](https://nobles92ts-ship-it.github.io/en/built/tc-team/index.md): Feed it a spec, get test cases. The LLM only writes sentences; deterministic code owns every structure and gate. - [tc-team · S1 · Design](https://nobles92ts-ship-it.github.io/en/built/tc-team/s1/index.md): Pulling a design skeleton out of the spec. Deciding the taxonomy is the same act as deciding the automation setup. - [tc-team · S2 · Isolation gate](https://nobles92ts-ship-it.github.io/en/built/tc-team/s2/index.md): Code re-reads the design to decide whether it can be expanded at all, and plants the anchors for traceability. - [tc-team · S3 · Writing](https://nobles92ts-ship-it.github.io/en/built/tc-team/s3/index.md): Code builds the structure; the LLM writes only sentences. Those sentences carry several prohibitions. - [tc-team · S4 · Adversarial review](https://nobles92ts-ship-it.github.io/en/built/tc-team/s4/index.md): The first and second passes deliberately don't look at each other's territory. And every finding gets counted by how many times it has recurred. - [tc-team · S5 · Apply](https://nobles92ts-ship-it.github.io/en/built/tc-team/s5/index.md): Applies the fix plan and clears four gates. This is the stage where rejection is correct behaviour. - [tc-team · S6 · Writing the sheet](https://nobles92ts-ship-it.github.io/en/built/tc-team/s6/index.md): The live sheet gets touched exactly once and read back afterwards. The questions left behind have standards too. - [tc-team · S7 · Finalize](https://nobles92ts-ship-it.github.io/en/built/tc-team/s7/index.md): The rule document owns the procedure and the runner reads it at execution time. This stage does not stop on failure. - [Loki](https://nobles92ts-ship-it.github.io/en/built/loki/index.md): Chat with your own PC. A local Slack agent running on your Claude Code subscription — no API key. - [Loki · Architecture](https://nobles92ts-ship-it.github.io/en/built/loki/architecture/index.md): The platform swaps and the brain swaps, but the core is one thing — with a human-held line in between. - [Loki · Commands](https://nobles92ts-ship-it.github.io/en/built/loki/commands/index.md): How much you can ask for through a chat box — and the two that changed how it feels. - [Loki · Permissions and guards](https://nobles92ts-ship-it.github.io/en/built/loki/permissions/index.md): Opening an agent with shell access on your own PC to other people. And stopping is a feature. - [Loki · Silent failures](https://nobles92ts-ship-it.github.io/en/built/loki/silent-failures/index.md): Almost every bug this project produced returned success and did nothing. - [QA widget](https://nobles92ts-ship-it.github.io/en/built/qa-widget/index.md): A small panel pinned to the corner of the screen. KPIs and progress, with the report one click away. - [QA widget · The screen](https://nobles92ts-ship-it.github.io/en/built/qa-widget/screen/index.md): Numbers alone aren't enough. How the number was arrived at, and what you can do about it, have to be right there. - [QA widget · Architecture](https://nobles92ts-ship-it.github.io/en/built/qa-widget/architecture/index.md): The widget does not fetch anything. It reads one file and draws it. - [QA widget · Delivery](https://nobles92ts-ship-it.github.io/en/built/qa-widget/delivery/index.md): Putting the tool on somebody else's machine turned out to be harder than building it. - [claude-bus](https://nobles92ts-ship-it.github.io/en/built/claude-bus/index.md): Leave a job on the office PC; a fresh session on the home PC picks it up. One shared folder, no server. - [claude-bus · Protocol](https://nobles92ts-ship-it.github.io/en/built/claude-bus/protocol/index.md): Conflict copies never appear because every folder has exactly one writer. - [claude-bus · Timing](https://nobles92ts-ship-it.github.io/en/built/claude-bus/timing/index.md): I measured where the three minutes go, leg by leg. Actual work was 14% of it. - [claude-bus · Safety](https://nobles92ts-ship-it.github.io/en/built/claude-bus/safety/index.md): What stands between somebody else's instruction and my machine. It goes to a person, not a shell. - [An adb-based game QA server](https://nobles92ts-ship-it.github.io/en/built/android-qa/index.md): A game screen exposes no UI tree. So it's driven by sight, and judged by logs. - [An adb-based game QA server · Driving and judging](https://nobles92ts-ship-it.github.io/en/built/android-qa/driving/index.md): Driving happens by sight; judging happens by log. Moving the verdict off the pixels was the biggest improvement. - [An adb-based game QA server · The unattended loop](https://nobles92ts-ship-it.github.io/en/built/android-qa/loop/index.md): Assume the product passes everything and every FAIL becomes an automation defect. A free oracle. - [An adb-based game QA server · What came out](https://nobles92ts-ship-it.github.io/en/built/android-qa/findings/index.md): The first thing that wasn't an automation defect. And it isn't something to brag about. - [MOON Studio](https://nobles92ts-ship-it.github.io/en/built/moon-studio/index.md): A one-person + AI game studio. Every game gets a slot; exactly one slot currently has anything in it. - [MOON Studio · Wuxia game (working title)](https://nobles92ts-ship-it.github.io/en/built/moon-studio/wuxia/index.md): An action RPG won by reading the enemy's pattern rather than out-reflexing it. The only slot in the studio still moving. - [MOON Studio · Wuxia game (working title) · Design](https://nobles92ts-ship-it.github.io/en/built/moon-studio/wuxia/design/index.md): Turn "win by reading the pattern" into a rule and telegraph duration becomes the only dial you have. - [MOON Studio · Wuxia game (working title) · Spec](https://nobles92ts-ship-it.github.io/en/built/moon-studio/wuxia/spec/index.md): The rules from the design page as actual artifact — four lineages, the matchup cycle, three verdict constants, real telegraph durations. - [MOON Studio · Wuxia game (working title) · World](https://nobles92ts-ship-it.github.io/en/built/moon-studio/wuxia/world/index.md): When the setting is decoration, the systems are left as bare numbers. Here factions become collectibles and identity becomes a stat. - [MOON Studio · Wuxia game (working title) · Client](https://nobles92ts-ship-it.github.io/en/built/moon-studio/wuxia/client/index.md): The places handling time and hit resolution leaked. Every one of them failed quietly, without throwing. - [MOON Studio · Wuxia game (working title) · Art](https://nobles92ts-ship-it.github.io/en/built/moon-studio/wuxia/art/index.md): The data was complete and the visual assets were zero. And the gate I added to fix readability made readability worse. - [MOON Studio · Wuxia game (working title) · QA](https://nobles92ts-ship-it.github.io/en/built/moon-studio/wuxia/qa/index.md): What a passing test isn't looking at. When zero means two different things, the test passes a broken build. - [MOON Studio · Future game 1](https://nobles92ts-ship-it.github.io/en/built/moon-studio/future-1/index.md): The slot for a game that doesn't exist yet — carrying only the lessons from the wuxia game that can actually travel. - [MOON Studio · Future game 1 · Design](https://nobles92ts-ship-it.github.io/en/built/moon-studio/future-1/design/index.md): If the way you win doesn't fit in two sentences, it isn't designed yet. One rule that travels, and one that has to be re-decided. - [MOON Studio · Future game 1 · Spec](https://nobles92ts-ship-it.github.io/en/built/moon-studio/future-1/spec/index.md): The rules have to fit one screen, and the undecided part has to be written down. The most important section in a spec is the unsettled one. - [MOON Studio · Future game 1 · Art](https://nobles92ts-ship-it.github.io/en/built/moon-studio/future-1/art/index.md): Building solo, the real bottleneck isn't code. This slot starts by deciding what won't be made. - [MOON Studio · Future game 1 · QA](https://nobles92ts-ship-it.github.io/en/built/moon-studio/future-1/qa/index.md): If you can't say in one sentence what a pass means, the test isn't verification yet. - [MOON Studio · Future game 2](https://nobles92ts-ship-it.github.io/en/built/moon-studio/future-2/index.md): The second empty slot. Its parts read the same as Future game 1 — there are two to show "several," not because they differ. ## What I took apart - [What I took apart](https://nobles92ts-ship-it.github.io/en/teardowns/): Repos and services I took apart — and what I kept. - [Agent harnesses](https://nobles92ts-ship-it.github.io/en/teardowns/harness/index.md): What to run an agent on top of. Twenty-one taken apart, and exactly one tool actually came in. - [Agent harnesses · mattpocock/skills](https://nobles92ts-ship-it.github.io/en/teardowns/harness/mattpocock-skills/index.md): I didn't take the 35 skills. I took the one document that defines how to use them — and measured against it, 1 of my 96 skills passed. - [Agent harnesses · ECC](https://nobles92ts-ship-it.github.io/en/teardowns/harness/ecc/index.md): I took apart the upstream of my own configuration. What was worth taking wasn't a feature but a measurement — and that measuring tool wouldn't run here. - [Agent harnesses · Prime Agent](https://nobles92ts-ship-it.github.io/en/teardowns/harness/prime-agent/index.md): I read the code and came away with four techniques. The best of them was something I'd written down myself five days earlier. - [Agent harnesses · Clawd on Desk](https://nobles92ts-ship-it.github.io/en/teardowns/harness/clawd-on-desk/index.md): I tried to justify installing it with measurements and failed. The record says so, in place. - [Agent harnesses · Odysseus](https://nobles92ts-ship-it.github.io/en/teardowns/harness/odysseus/index.md): I went looking for the enforcement point behind each of the five disciplines the docs declare. One of them existed. - [Agent harnesses · ponytail](https://nobles92ts-ship-it.github.io/en/teardowns/harness/ponytail/index.md): Four lines shortlisted out of a rule set with 92,000 stars; one held up. The more valuable finding came from my own rules. - [Agent harnesses · OmniRoute](https://nobles92ts-ship-it.github.io/en/teardowns/harness/omniroute/index.md): I didn't skip this 41,856-star gateway over performance. Its tier 1 is my own subscription accounts, disguised and sent upstream. - [Agent harnesses · Cross-session messaging](https://nobles92ts-ship-it.github.io/en/teardowns/harness/session-messaging/index.md): The mechanism the video dug up doesn't exist on my OS. The capability I actually wanted was already on, and the real failure mode was misdelivery, not security. - [Agent harnesses · Buzz group chat](https://nobles92ts-ship-it.github.io/en/teardowns/harness/buzz-groupchat/index.md): Slack where the channel members are people and local CLIs. Two demos, one worked — and the presenter admitting which one didn't is what this video is worth. - [Agent harnesses · Four plugins](https://nobles92ts-ship-it.github.io/en/teardowns/harness/four-plugins/index.md): A 44-second recommendation reel. Two of the four I had already taken apart and rejected, and the counterexample sat inside the maker's own benchmark table. - [Agent harnesses · Meta-prompting](https://nobles92ts-ship-it.github.io/en/teardowns/harness/meta-prompting/index.md): Don't write a good prompt — hand over all the raw material and let the AI carve it down. Nowhere in the three steps is the result measured. - [Agent harnesses · A field guide to Fable](https://nobles92ts-ship-it.github.io/en/teardowns/harness/fable-field-guide/index.md): Someone from the company that built the model, explaining how to use it. Which makes the six techniques trustworthy and the closing claim the least trustworthy part. - [Agent harnesses · obra/superpowers](https://nobles92ts-ship-it.github.io/en/teardowns/harness/superpowers/index.md): I didn't adopt the framework. But the real output of this teardown wasn't a takeaway — it was watching two of my own conclusions collapse under the method the project itself prescribes. - [Agent harnesses · ASIDE](https://nobles92ts-ship-it.github.io/en/teardowns/harness/aside/index.md): It won't install on my machines. That should have been a one-line rejection, except the "no plans" part of my own judgement fell apart on three lines of changelog. - [Agent harnesses · One skill, ad images, pay-as-you-go routing](https://nobles92ts-ship-it.github.io/en/teardowns/harness/payg-image-routing/index.md): The skill design is a well-built real thing. The conclusion in the title, though, falls apart on a table the video itself puts on screen. - [Agent harnesses · DeepSeek Harness](https://nobles92ts-ship-it.github.io/en/teardowns/harness/deepseek-harness/index.md): I set out to harvest five techniques from a 181,214-star repo, and every one of them died on two skills I had installed and never once run. - [Agent harnesses · ip-as-logo](https://nobles92ts-ship-it.github.io/en/teardowns/harness/ip-as-logo-skill/index.md): A skill with not one line of code. One of the two pillars I took for its core invention lives only in a file the agent never reads, and the other was a number that flipped four times in 45 hours. - [Agent harnesses · anti-slop](https://nobles92ts-ship-it.github.io/en/teardowns/harness/anti-slop/index.md): I ran 15 rules over all of my own code, got 27 hits, and decided to fix none of them. Looking for somewhere to attach them instead, I found that my own leak-defence gate can switch itself off in silence. - [Agent harnesses · DSH better-sidebar](https://nobles92ts-ship-it.github.io/en/teardowns/harness/dsh-better-sidebar/index.md): The headline number — a ~325KB core — broke under measurement. But the gate I was using to judge someone else's ungated repository had never once been committed in my own. - [Agent harnesses · cumora](https://nobles92ts-ship-it.github.io/en/teardowns/harness/cumora/index.md): The only hold among eight. The adversarial pass found that my own deploy was leaking too — and both the diagnosis and the prescription were wrong. - [Agent harnesses · eli5 — a plugin with zero lines of code](https://nobles92ts-ship-it.github.io/en/teardowns/harness/eli5/index.md): One instruction turned into a plugin — explain it to me like I am five. Opened up, it is three files, 865 bytes, zero lines of executable code — and in 21 days I called it exactly once. - [Agent harnesses · BrowserSkill — lending an agent my browser](https://nobles92ts-ship-it.github.io/en/teardowns/harness/browserskill/index.md): Tencent's tool lets an agent drive the Chromium I am already signed into. The design is good. Then I opened the source and found the doorman checking the costume instead of the face — and a comment where the authors admit it. - [Context and memory](https://nobles92ts-ship-it.github.io/en/teardowns/memory/index.md): What goes up to the model and what stays outside. Across all eleven, not one line of code crossed over — only rules did. - [Context and memory · RAG without a vector DB](https://nobles92ts-ship-it.github.io/en/teardowns/memory/rag-without-vectors/index.md): Every structural claim held up and everything that failed was a specific number — the characteristic failure of secondhand write-ups. - [Context and memory · Headroom](https://nobles92ts-ship-it.github.io/en/teardowns/memory/headroom/index.md): It started as an adoption review. Measuring my own environment, the thing this tool would shrink didn't exist. - [Context and memory · Karrot SEED](https://nobles92ts-ship-it.github.io/en/teardowns/memory/karrot-seed/index.md): The design system had nothing to do with my work. The value was in how the repository is arranged so agents can read it. - [Context and memory · A second-brain workshop](https://nobles92ts-ship-it.github.io/en/teardowns/memory/second-brain-course/index.md): Seven of the eight curriculum slots were already covered. Then I ran the deck's own skill scorecard over my 99 skills and two real defects fell out. - [Context and memory · Graph engineering](https://nobles92ts-ship-it.github.io/en/teardowns/memory/graph-engineering/index.md): What the video calls the heart of a graph fired once across my 52 runs. The one thing I didn't have, though, turned out to be worth a measured 9.9 minutes a run. - [Context and memory · An LLM wiki in Obsidian](https://nobles92ts-ship-it.github.io/en/teardowns/memory/llm-wiki-obsidian/index.md): I went looking for structure and came back with two hygiene rules. My layers were already finer than the video's, and the thing I wanted most I put on hold. - [Context and memory · A second-brain interview](https://nobles92ts-ship-it.github.io/en/teardowns/memory/second-brain-interview/index.md): Why someone who had built two hundred automation workflows threw them all away and focused on one context instead. The structure is ordinary; the way it connects was not. - [Context and memory · A file management system](https://nobles92ts-ship-it.github.io/en/teardowns/memory/file-management/index.md): Five conclusions matched a design I'd built separately, and not one of the reasons overlapped. It also named three gaps in mine exactly. - [Context and memory · A second brain in four steps](https://nobles92ts-ship-it.github.io/en/teardowns/memory/second-brain-4steps/index.md): There was almost nothing here to learn. And in the process of holding it up, my own step four turned out to be installed and never once run. - [Context and memory · dsh-anchored-standard](https://nobles92ts-ship-it.github.io/en/teardowns/memory/dsh-anchored-standard/index.md): Installing it was flatly impossible — my harness has no hook that can take context away. The worse part was that I already owned the tool to measure this, and hadn't run it in 24 days. - [Context and memory · public-apis](https://nobles92ts-ship-it.github.io/en/teardowns/memory/public-apis/index.md): The link check on a 477k-star list had failed 400 days running. And the list I went looking for was already sitting in my own folder. - [QA automation](https://nobles92ts-ship-it.github.io/en/teardowns/qa/index.md): This is my day job. It also yields the least — because I'm usually already standing in the same place. - [QA automation · Toss Nebula device farm](https://nobles92ts-ship-it.github.io/en/teardowns/qa/toss-nebula/index.md): Nothing transferred. What it gave me instead was a ruler — and held against my own code, it found a crack. - [QA automation · Asleep's QA automation](https://nobles92ts-ship-it.github.io/en/teardowns/qa/asleep-qa/index.md): Ten of the eleven decisions I had already arrived at. The value wasn't transfer — it was independent confirmation. - [QA automation · An autoplay test tool](https://nobles92ts-ship-it.github.io/en/teardowns/qa/autoplay-test-tool/index.md): A bot that walks the map measuring performance. It detects being stuck as a contradiction in the data and clears false positives with two timers — both better than mine. - [QA automation · A talk on QA infrastructure](https://nobles92ts-ship-it.github.io/en/teardowns/qa/qa-infrastructure/index.md): There were no captions, so I reconstructed it from slides alone. Even carrying that gap, one thing survived — and it wasn't about tools, it was about the definition of evidence. - [QA automation · watermarks-remover](https://nobles92ts-ship-it.github.io/en/teardowns/qa/watermarks-remover/index.md): I opened a tool that erases AI fingerprints in order to audit my own publish gate. Then the metric built to catch AI caught something I had written by hand. - [Making games](https://nobles92ts-ship-it.github.io/en/teardowns/games/index.md): Different engine, different dimensions, different licence — not one line of code crosses. What crosses is a rule that shrinks to one sentence. - [Making games · Elemental Sandbox](https://nobles92ts-ship-it.github.io/en/teardowns/games/elemental-sandbox/index.md): The effects are the least valuable part. One rule came across — record dimensionless, resolve dimensions every frame. - [Making games · OpenMMO](https://nobles92ts-ship-it.github.io/en/teardowns/games/openmmo/index.md): An MMORPG one developer took all the way to live service. Opening the code showed my README-based summary was wrong in two places. - [Making games · Hunyuan3D-WorldClaw](https://nobles92ts-ship-it.github.io/en/teardowns/games/worldclaw/index.md): A paper about building an editable 3D open world from one line of prompt. The repository has no code and the paper has no numbers. - [Making games · Dynasty Warriors Origins](https://nobles92ts-ship-it.github.io/en/teardowns/games/dynasty-warriors-origins/index.md): I went in to see what else to build into my game, and the first thing I found was something I no longer had to build. - [Making games · Recent Unity releases](https://nobles92ts-ship-it.github.io/en/teardowns/games/unity-releases/index.md): Draw the week honestly and it was five release builds, nothing else. Inside them, though, was not a feature addition but a move. ## Things I wrote - [Things I wrote](https://nobles92ts-ship-it.github.io/en/writing/): Notes that outgrew a commit message. - [Verdicts](https://nobles92ts-ship-it.github.io/en/writing/verdicts/index.md): The space between PASS and FAIL — what counts as evidence, and how much a machine may decide alone. - [Verdicts · I did not touch the threshold](https://nobles92ts-ship-it.github.io/en/writing/verdicts/threshold/index.md): One run, 126 cases, one failure. I knew how to make it pass, and didn't. - [Verdicts · Is it broken, or not there yet?](https://nobles92ts-ship-it.github.io/en/writing/verdicts/version-docs/index.md): In version documentation, half the feature description is wrong within six months. The date it appeared never is. And the date is usually what a verdict needs. - [Verdicts · "Probably right" is not a verdict](https://nobles92ts-ship-it.github.io/en/writing/verdicts/evidence-grades/index.md): Answering the same question with "probably right" or with "right" didn't come down to how hard I worked at it. - [Verdicts · The 53 orphans were a false alarm](https://nobles92ts-ship-it.github.io/en/writing/verdicts/false-alarm/index.md): The linter reported 53 memories cut loose from the index. All 53 were fine. Restoring them would have pushed the index to 124% of its limit. - [Verdicts · The newest install looked like the culprit](https://nobles92ts-ship-it.github.io/en/writing/verdicts/new-install-suspect/index.md): The home network started dropping and I was ready to uninstall Tailscale, which I had put on days earlier. Timing the install and the symptoms separately put the fault markers 18 to 127 days ahead of it. - [Measurement](https://nobles92ts-ship-it.github.io/en/writing/measurement/index.md): Where the time actually went, and how to keep estimates and measurements apart. And the work of deciding what to compare a number against. - [Measurement · All four times, the numbers were right](https://nobles92ts-ship-it.github.io/en/writing/measurement/baselines/index.md): Measurement usually goes wrong somewhere other than the arithmetic. I was wrong four times, the arithmetic was right every time, and what was wrong was what I measured against. - [Running agents](https://nobles92ts-ship-it.github.io/en/writing/agents/index.md): The other side — loops, guards, and the things I decided not to automate. - [Running agents · Agents are quiet when they fail](https://nobles92ts-ship-it.github.io/en/writing/agents/silent-success/index.md): Running one unattended, the frightening thing isn't a wrong answer. A wrong answer is visible when you read it. The frightening thing is exit code 0. - [Running agents · Unattended regression is built by recovery, not speed](https://nobles92ts-ship-it.github.io/en/writing/agents/unattended-regression/index.md): A real device ran 300 cycles overnight with nobody watching. What made it finish was not fast code but knowing how to cut things off. - [Choosing](https://nobles92ts-ship-it.github.io/en/writing/choosing/index.md): Deciding whether to bring someone else's work in. What counts as grounds for rejection, and when a rejection gets reopened. - [Choosing · Forty rejections, and the reason was rarely the tool](https://nobles92ts-ship-it.github.io/en/writing/choosing/adopting/index.md): I took apart forty tools other people built and installed exactly one. What pays is not the 1-in-40 ratio — it is the reasons for the other thirty-nine. - [Choosing · You don't have it if it has never run](https://nobles92ts-ship-it.github.io/en/writing/choosing/already-have-it/index.md): Being installed is only the first of three layers. Three investigations pointed at the same two tools, and all three times those tools had never run. ## Who I am - [Who I am](https://nobles92ts-ship-it.github.io/en/about/): Who is doing all this, and why. - [About](https://nobles92ts-ship-it.github.io/en/about/me/index.md): I'm not a developer. The AI writes the code. I've spent twelve years finding where it breaks. - [Privacy](https://nobles92ts-ship-it.github.io/en/about/privacy/index.md): The only thing this site collects from you is how many times a page has been opened. It also fetches fonts from elsewhere, and the host keeps access logs. ## AI news today - [AI news today — everything](https://nobles92ts-ship-it.github.io/en/ai-issue/): Rebuilt every morning; the cards themselves are gone within days. - [Papers](https://nobles92ts-ship-it.github.io/en/ai-issue/papers/) - [Models](https://nobles92ts-ship-it.github.io/en/ai-issue/models/) - [GitHub](https://nobles92ts-ship-it.github.io/en/ai-issue/repos/) - [Newly out](https://nobles92ts-ship-it.github.io/en/ai-issue/releases/) - [Talk](https://nobles92ts-ship-it.github.io/en/ai-issue/discussion/) - [Glossary](https://nobles92ts-ship-it.github.io/en/ai-issue/words/): 218 AI terms. This part is written, not collected. - [Basics](https://nobles92ts-ship-it.github.io/en/ai-issue/words/basics/) - [Model design](https://nobles92ts-ship-it.github.io/en/ai-issue/words/models/) - [Training](https://nobles92ts-ship-it.github.io/en/ai-issue/words/training/) - [Inference & serving](https://nobles92ts-ship-it.github.io/en/ai-issue/words/inference/) - [Agents & tools](https://nobles92ts-ship-it.github.io/en/ai-issue/words/agents/) - [Evaluation](https://nobles92ts-ship-it.github.io/en/ai-issue/words/eval/) - [Safety & policy](https://nobles92ts-ship-it.github.io/en/ai-issue/words/safety/) - [Infrastructure](https://nobles92ts-ship-it.github.io/en/ai-issue/words/infra/) ## Optional - [한국어 색인](https://nobles92ts-ship-it.github.io/ko/llms.txt): 같은 목록의 한국어판. 이쪽이 원문이다. - [sitemap.xml](https://nobles92ts-ship-it.github.io/sitemap.xml): Every URL in both languages, with lastmod.