TODAY · AI DAILY
No. 115 · 2026.11.09智能体改走潜空间,微软研究组告终
多篇新论文让智能体绕开文本,在潜空间压缩上下文、并行协作,线性注意力与 MoE 可解释性也有进展。英伟达董事会调查黄仁勋,微软 AI Frontiers 创始人离职。
Read today's issue (in Chinese) ↗Things I've made24

Prompts are picked from the public LLM Arena dataset and handed to Claude, GPT, Grok, Kimi and others for a single shot, with no manual edits. Tasks include a lava-lamp physics sim, a Windows 98 Paint clone, a MacBook Air comparison page, a rotating kebab and a flight through Van Gogh's Starry Night. Every entry opens full screen, and the arrow keys flip between models.
- Prompts
- 05
- Models
- 09
- Entries
- 45
LLM evals · React 19 · Express · Sandboxed iframes

Each prompt is reusable: a steampunk clockmaker's workshop, a cherry-blossom railway stop, a tidal research station. Every build records the model and coding agent behind it, so a prompt page can run all of its entries side by side. Scenes are self-contained single-file HTML, sandboxed twice (an opaque-origin iframe plus CSP), and are added through an authenticated API and a skill without rebuilding the image.
- Prompts
- 21
- Builds
- 33
- Model × agent
- 02
WebGL · Next.js · Coding agents · Sandboxing

A Go daemon on the dev machine and the browser both dial out to a self-hosted relay; pairing takes one 6-digit code. The web UI has a real PTY terminal, a file browser, Git diffs and local port forwarding, talks directly to ACP agents such as Devin and cursor-agent, and can hand out temporary read-only links when you want to ask someone for help.
Go · React · WebSocket · ACP

Upload one or more receipt photos and a vision model reads the items, unit prices and totals. Receipts in other languages get a Chinese translation alongside, and included tax, added tax and discounts are reconciled arithmetically so everyone's shares add up to exactly what was paid. Bills and photos live only in the browser's IndexedDB; the server keeps no user data.
Next.js 16 · Vision models · IndexedDB · Mobile web

No database, no admin panel, no build pipeline. A scheduled job files each issue into year/month folders, or pushes it idempotently through a key-protected POST endpoint. The home page shows today's issue (UTC+8) and falls back to the latest one. On top of standard GFM there are seven layout blocks, including metric cards, comparisons and timelines.
Next.js 16 · Markdown · Automated publishing

It reads the response of the send-message API to tell whether a message is visible only to you. Swallowed messages turn red in the chat list with a "banned" badge; a floating panel on the stream page covers the current room, and the toolbar popup gathers every room with filters, search and CSV import and export. Records sync across computers with your Chrome account. The sync area is packed deterministically from newest to oldest, so devices converge on one copy instead of overwriting each other. The only permission left is storage.
Chrome MV3 · JavaScript · storage.sync · Bilibili

v0.2 goes beyond a read-only Apple Health summary (sleep, HRV, resting heart rate, steps, active energy, exercise minutes) and adds eating and training. Snap a meal and AI breaks it into items with nutrition; tick items and drag a slider for how much you actually ate. Packaged food is read straight from the nutrition label, with kJ converted to kcal. Plan workouts by week, log sets and weights, and review history. The raised button in the tab bar is press-to-talk: say "a bowl of beef noodles for lunch" or "500 ml of water" and it's logged. Meals, water, weight and goals stay on device, export to JSON, and can optionally be written back to Apple Health; only photo recognition and voice commands go to my own server, and only after you allow it.
Swift · SwiftUI · SwiftData · HealthKit · FastAPI · Claude
Meal photo in the cover: Jjj84206 / Wikimedia Commons, CC BY-SA 4.0In progress · not released yet
Recording something and finding that one sentence later should both take two taps or less. The app opens on a big red record button, and you can also start recording from the Lock Screen, Control Center, the Action Button or Apple Watch without unlocking; a Live Activity lets you pause and stop. Recordings stay on device and anything over 30 seconds is transcribed automatically, either fully on iPhone with no upload and no account, or in the cloud via ElevenLabs Scribe with speaker diarization after Sign in with Apple. Transcripts are timestamped subtitles: tap any line to play from there, the current line follows playback, with speed control, ±15 s and speaker renaming. Organize recordings into projects, search titles and full transcripts, and sync across devices with iCloud.
Swift · SwiftUI · SwiftData · CloudKit · Speech · ElevenLabs
In progress · not released yet
Paste a link or some text, or upload a document. An LLM distills it into a structured summary, which is then laid out as a long read with sections, data cards and pull quotes, optionally translated into Chinese. A companion Chrome extension generates one from any article or Bilibili video page in one click.
- Articles generated
- 99
React · Python · Chrome extension · LLM

A pixel engine, a general character system, YAML or Python scripts and chiptune music in one framework. Every frame is a pure function of time and renders in parallel across processes, so an 8-core machine turns out a 33-second short in 15 seconds. It can also turn the subject of a video into a pixel character. The framework was extracted and generalized from a 93-second MIT-licensed animation.
Python · NumPy · ffmpeg · Pixel art

Every server runs a small Python agent that reads IPMI inlet temperature, GPUs, disk SMART data and NIC counters; the Next.js server polls them and writes to SQLite. The UI is styled like an industrial HMI and fits a single 1920×1080 screen. Sensors that cannot be read show "—" rather than 0, and alert thresholds can be changed at runtime.
Next.js · SQLite · IPMI · Wall display
Personal use · not public

Live sites2026● Live
Frontend Arena
Prompts are picked from the public LLM Arena dataset and handed to Claude, GPT, Grok, Kimi and others for a single shot, with no manual edits. Tasks include a lava-lamp physics sim, a Windows 98 Paint clone, a MacBook Air comparison page, a rotating kebab and a flight through Van Gogh's Starry Night. Every entry opens full screen, and the arrow keys flip between models.
- Prompts
- 05
- Models
- 09
- Entries
- 45
LLM evals · React 19 · Express · Sandboxed iframes

Live sites2026● Live
Miniature Scene
Each prompt is reusable: a steampunk clockmaker's workshop, a cherry-blossom railway stop, a tidal research station. Every build records the model and coding agent behind it, so a prompt page can run all of its entries side by side. Scenes are self-contained single-file HTML, sandboxed twice (an opaque-origin iframe plus CSP), and are added through an authenticated API and a skill without rebuilding the image.
- Prompts
- 21
- Builds
- 33
- Model × agent
- 02
WebGL · Next.js · Coding agents · Sandboxing

Open source2026● Open source
Coding Anywhere
A Go daemon on the dev machine and the browser both dial out to a self-hosted relay; pairing takes one 6-digit code. The web UI has a real PTY terminal, a file browser, Git diffs and local port forwarding, talks directly to ACP agents such as Devin and cursor-agent, and can hand out temporary read-only links when you want to ask someone for help.
Go · React · WebSocket · ACP

Live sites2026● Live
AA Bill
Upload one or more receipt photos and a vision model reads the items, unit prices and totals. Receipts in other languages get a Chinese translation alongside, and included tax, added tax and discounts are reconciled arithmetically so everyone's shares add up to exactly what was paid. Bills and photos live only in the browser's IndexedDB; the server keeps no user data.
Next.js 16 · Vision models · IndexedDB · Mobile web

Live sites2026● Live
AI Daily Briefing
No database, no admin panel, no build pipeline. A scheduled job files each issue into year/month folders, or pushes it idempotently through a key-protected POST endpoint. The home page shows today's issue (UTC+8) and falls back to the latest one. On top of standard GFM there are seven layout blocks, including metric cards, comparisons and timelines.
Next.js 16 · Markdown · Automated publishing

Apps · Extensions2024● Open source
Banned Danmaku Marker
It reads the response of the send-message API to tell whether a message is visible only to you. Swallowed messages turn red in the chat list with a "banned" badge; a floating panel on the stream page covers the current room, and the toolbar popup gathers every room with filters, search and CSV import and export. Records sync across computers with your Chrome account. The sync area is packed deterministically from newest to oldest, so devices converge on one copy instead of overwriting each other. The only permission left is storage.
Chrome MV3 · JavaScript · storage.sync · Bilibili

Apps · Extensions2026● In progress
VitaLoop
v0.2 goes beyond a read-only Apple Health summary (sleep, HRV, resting heart rate, steps, active energy, exercise minutes) and adds eating and training. Snap a meal and AI breaks it into items with nutrition; tick items and drag a slider for how much you actually ate. Packaged food is read straight from the nutrition label, with kJ converted to kcal. Plan workouts by week, log sets and weights, and review history. The raised button in the tab bar is press-to-talk: say "a bowl of beef noodles for lunch" or "500 ml of water" and it's logged. Meals, water, weight and goals stay on device, export to JSON, and can optionally be written back to Apple Health; only photo recognition and voice commands go to my own server, and only after you allow it.
Swift · SwiftUI · SwiftData · HealthKit · FastAPI · Claude
Meal photo in the cover: Jjj84206 / Wikimedia Commons, CC BY-SA 4.0In progress · not released yet
Apps · Extensions2026● In progress
Taptio
Recording something and finding that one sentence later should both take two taps or less. The app opens on a big red record button, and you can also start recording from the Lock Screen, Control Center, the Action Button or Apple Watch without unlocking; a Live Activity lets you pause and stop. Recordings stay on device and anything over 30 seconds is transcribed automatically, either fully on iPhone with no upload and no account, or in the cloud via ElevenLabs Scribe with speaker diarization after Sign in with Apple. Transcripts are timestamped subtitles: tap any line to play from there, the current line follows playback, with speed control, ±15 s and speaker renaming. Organize recordings into projects, search titles and full transcripts, and sync across devices with iCloud.
Swift · SwiftUI · SwiftData · CloudKit · Speech · ElevenLabs
In progress · not released yet
Live sites2025● Live
Infographic Renderer
Paste a link or some text, or upload a document. An LLM distills it into a structured summary, which is then laid out as a long read with sections, data cards and pull quotes, optionally translated into Chinese. A companion Chrome extension generates one from any article or Bilibili video page in one click.
- Articles generated
- 99
React · Python · Chrome extension · LLM

Open source2026● Open source
pixtoon
A pixel engine, a general character system, YAML or Python scripts and chiptune music in one framework. Every frame is a pure function of time and renders in parallel across processes, so an 8-core machine turns out a 33-second short in 15 seconds. It can also turn the subject of a video into a pixel character. The framework was extracted and generalized from a 93-second MIT-licensed animation.
Python · NumPy · ffmpeg · Pixel art

Personal tools2026● Private
Control Grid
Every server runs a small Python agent that reads IPMI inlet temperature, GPUs, disk SMART data and NIC counters; the Next.js server polls them and writes to SQLite. The UI is styled like an industrial HMI and fits a single 1920×1080 screen. Sensors that cannot be read show "—" rather than 0, and alert thresholds can be changed at runtime.
Next.js · SQLite · IPMI · Wall display
Personal use · not public
MORE · 13 more projects

It holds handy tools such as an LLM API playground, Anthropic and Gemini API testers, a JWT decoder and a timestamp converter, plus a pixel Mario remake and Flappy Bird. Snippets can be pinned and previewed, and any single one can be shared with anyone. The backend is wired to an LLM, so one sentence is enough to create or rewrite a snippet.
- Snippets
- 34
Toolbox · AI-generated · Flask · Self-hosted
Built for two or three people polishing one deck together. opencode runs in a container as the agent engine; you can paste screenshots, upload assets, import HTML or PPTX, and export HTML, ZIP, PDF or PPTX. Sign-in is a shared password, and the model provider can be switched at will.
opencode · Node.js · SSE · Self-hosted
Personal use · not publicRequests are load-balanced across several Modal accounts. Requests that carry a session id stick to one account to keep the prefix cache warm, and failures switch over automatically with a cool-down. /v1/messages is translated to chat completions and back, with reasoning mapped to thinking blocks, so Claude Code plugs straight in. All traffic can also be mirrored to Langfuse.
Python · Anthropic API · Modal · Langfuse
A small Go service: for each chat completions request it runs the matching agent CLI in one-shot mode and returns the result in OpenAI format. /v1/models lists only the CLIs that are actually installed, a colon in the model name passes an inner model through, and any client or SDK that accepts a base URL works out of the box.
Go · OpenAI-compatible · Agent CLIs
At the end of each turn a Stop hook asks a separate evaluator to check the condition. If it isn't met, the stop is blocked and the evaluator's reasoning becomes the next instruction; once it is met, the goal clears itself. A Python installer adds it to Devin's global config, and it works on Windows too.
Devin CLI · Hooks · Skills · Python
It lives in the same extensions repo as the danmaku marker. The manifest pins a public key, so the extension ID is the same on every computer and settings sync automatically through chrome.storage.sync. Setting up a new machine takes one script that pulls and builds.
Chrome extension · TypeScript · LLM translation
Personal use · not publicEach user gets their own working directory, API key and model settings. Conversations stream over WebSocket with tools such as Bash, file read, write and edit, Glob and Grep, and sessions can be renamed and revisited. One-command Docker Compose deploy, persisted in SQLite.
TypeScript · WebSocket · Docker
Models and tasks are organized by dimension (code quality, reasoning, frontend generation, summarization and more) and scored by an LLM judge, a script, or both. The results page previews HTML output in a sandbox, takes star ratings and notes, and compares different models' answers to the same task side by side.
LLM evals · TypeScript · LLM-as-judge
A FastAPI job API wraps it: upload an image, get a job id, and the model is built asynchronously, with orbit videos and depth maps rendered on CUDA machines. There are CPU and CUDA images, Bearer-token auth, and jobs and files tracked in SQLite.
Python · FastAPI · 3D Gaussian splatting · CUDA
A FastAPI backend connects to the chat servers over WebSocket, refreshes chat stats every 10 seconds and stores history in SQLite; a Next.js frontend handles the stream list and sending. Optional site-wide password, one-command Docker start.
FastAPI · Next.js · WebSocket · Bilibili
FastAPI with JWT sign-in, direct uploads to Alibaba Cloud OSS, straight-from-camera uploads on the phone, history and file management, all running on my own server.
FastAPI · Direct OSS upload · Redis
Personal use · not publicThe backend proxies the Gemini image API without exposing any ports, the frontend goes through a Vite proxy, and there's a login, since it's only for me and friends. I'd written similar test tools for Gemini and Sora before.
Gemini · Vite · Self-hosted
Personal use · not publicpandas rebuilds every position and computes win rate and P&L distribution, with charts written straight into Obsidian notes. Backtests cover ADX + Stoch RSI, EMA channels, MACD and more, split by year for in-sample and out-of-sample checks.
Python · pandas · Backtesting
Personal use · not public
Toys
- My Friend Dabuniu, episode 01An AI-made cartoon about not leaning on AI too much, with visuals and music all generated in code2026
- Super Mario, Worlds 1–2A pixel remake of eight levels, all beatable, with an auto-play demo2026
- Carousel in an Empty FunfairA little 3D scene from Miniature Scene: six horses bobbing slowly under a candy-colored canopy2026
- Flappy BirdThe classic, playable right in the browser2026
- SBTI Personality AvatarsNano Banana draws an avatar for each MBTI or SBTI personality type2026
- Map is Not the Territory9 Claude skills that help close knowledge gaps2026
- Screen FlashbangClick or press space and the whole screen flashes white, ear-ringing sound included2025
- Christmas Avatar MakerUpload a photo and add a Santa hat in one click2025
- LLM API PlaygroundPoke at /v1-compatible APIs: streaming, multimodal, raw logs2025
- Model Price Ratio CalculatorEnter input and output prices to get New API model ratios2025
- how2cmdSay what you want to do and an LLM tells you which command to type2025
- GitHub Stars AnalyzerAn LLM sorts and summarizes the repos you've starred2025
Timeline
- 2026.10VitaLoop v0.2 reached TestFlight with photo meal logging, training plans and a press-to-talk assistant; started Taptio, with on-device and cloud transcription and recording from the Lock Screen and Apple Watch
- 2026.10Started VitaLoop, with v0.1 reading Apple Health; this site became a personal homepage
- 2026.10Banned Danmaku Marker 1.1 landed on the Chrome Web Store, with the new source open too
- 2026.09Launched Miniature Scene; open-sourced the pixtoon pixel animation framework
- 2026.08Launched AA Bill for splitting receipts; built PPT Copilot for editing slides by chat
- 2026.07Launched Frontend Arena and AI Daily; open-sourced agent-provider and modal-provider
- 2026.06Remade eight Super Mario levels (Worlds 1–2) in pixels; ported /goal to Devin CLI
- 2026.05Open-sourced Coding Anywhere for controlling a dev machine and AI agents from a phone
- 2026.03Built the LLM Arena eval bench and the multi-tenant Claude Code Hub
- 2026.01Built the mlsharp single-image-to-3D service and Bilibili Danmaku Helper
- 2025.12Open-sourced Infographic Renderer with its Chrome extension; launched the Christmas Avatar Maker
- 2025.11Launched a link hub for my services, the ancestor of this site
- 2025.09Launched the self-hosted CodePen, which has since collected 30-odd snippets
- 2025.07Built Clipboard Studio, a cross-device clipboard
- 2024.11Released the extension that spots swallowed Bilibili live-chat messages
- 2021–2025Wrote Go backends and Web3 tools: Ethereum development, concurrent Etherscan calls, a private Avalanche L1 chain
Visitors
WORLD · 0 countries & regions
—
Each IP counts once a day · only country and province are kept, never the IP · IP data: ip2region
About
More than classic software engineering, I care about fitting AI into real-world problems. From large language models, AI agents and workflow orchestration to computer vision, NeRF, digital humans, speech-driven animation and real-time audio and video, I like to understand how things work first, then turn research into products people can actually use.
I'm used to carrying a project end to end on my own: choosing the stack, designing the architecture, writing the code, then deploying and running it. I'm at home with Docker, Linux and cloud servers. This site and most of the projects above run on machines I look after, and I enjoy tinkering with and polishing my own dev setup.
Beyond the tech, I care just as much about product design, user experience and human–computer interaction. I love trying new tools and new ways of working with AI, and I think a lot about how AI will change the way software gets written. Lately I've been building things around health tracking, knowledge processing and AI tooling.
I believe technology is worth more than the problems it solves: it's about opening up new possibilities.
For me, programming isn't just a job. It's a way to explore the world, express ideas and make things.
STACK · Day to day
Python · Go / TypeScript · Next.js / Docker · Linux
SELF-HOSTED · 12 services
- TRAEFIK
- NEW API
- OPEN WEBUI
- CLIPBOARD STUDIO
- CODEPEN
- NANO BANANA
- FRONTEND ARENA
- MINIATURE SCENE
- AI DAILY
- AA BILL
- INFOGRAPHIC
- CODING ANYWHERE
Fin is Intercom's AI agent. Intercom only loads once you open it, and it handles the conversation.
2026 · MuserQuantity