鼬 (@evvvvince) profile picture

girl who codes | vibecoding | BUPT

Joined November 2021
Photo by 鼬 (@evvvvince)Photo by 鼬 (@evvvvince)Photo by 鼬 (@evvvvince)Photo by 鼬 (@evvvvince)Photo by 鼬 (@evvvvince)Photo by 鼬 (@evvvvince)Photo by 鼬 (@evvvvince)Photo by 鼬 (@evvvvince)Photo by 鼬 (@evvvvince)Photo by 鼬 (@evvvvince)
Why Fuel Trucks Aren't Empty Inside
2
7
260
Horology doesn't have to be learned from flat diagrams. A Swiss lever watch movement in interactive 3D — from the coiled mainspring barrel to the 15-tooth escape wheel, ruby pallet fork, and breathing hairspring. 145 KB of procedural code, zero 3D asset files. Built with Claude Fable 5.2 and @threejs. Live bench ↓
11
17
660
Interactive bench (runs at 60fps in the browser — drop to 0.04x to inspect the ruby pallet lock & impulse): horology-escapement.pages.de…
1
8
256
OpenAI cuts off bundled models in Cursor tomorrow (Nov 12). Claude Fable 5.2 dropped 48 hours ago. And Opus 5.5 has felt weird since Nov 3. Everyone's coding setup just broke in the same week, so I spent last night reading Anthropic's new prompting docs, Artificial Analysis v5, and ~40 workflow threads to sort out what actually works right now. 4 things worth knowing before you burn your weekly limit: 1. Fable 5.2 acts like a "polite guest" by default. Buried in Anthropic's Fable 5.2 guide: if you don't state who owns the repo, it assumes it's visiting someone else's code. It bolts on helper functions and leaves duplicate code alone (<10% proactive cleanup). Put this one line in your CLAUDE.md: "This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it." In Anthropic's own tests, proactive cleanup and bug fixes jumped from <10% to ~50%. 2. Pin your sub-agents or your quota evaporates. In Claude Code, picking Fable 5.2 ($8/$40 per 1M, or $16/$80 in Fast mode) makes spawned sub-agents copy Fable by default. Run Fable 5.2 strictly as the planner/conductor, and pin worker sub-agents to Opus 5.5. 3. Don't rage-quit Cursor yet, just change how you route it. Bundled OpenAI access ends tomorrow post-SpaceX acquisition, so you either plug in BYOK or lean on Claude/Gemini inside Cursor. Honestly, Cursor's new multiplayer Projects + Tab completion paired with Claude Code in the terminal is still the cleanest combo. 4. Use DeepSeek V4.1 / Qwen 3.8 for the dirty work. Frontier models are too expensive for brute-force auditing. Spinning up 5-10 DeepSeek V4.1 workers in a VM to fuzz edge cases or write boring tests costs pennies before Fable touches architecture. Curious what setup everyone else is landing on after tomorrow's cutoff.
27
1
22
967
everyone keeps complaining that AI is generating mountains of new tech debt meanwhile in secops the exact opposite is happening: coding agents are fuzzing and reading legacy repos so fast that 30-year-old C and PHP codebases that only survived via "security through nobody-wants-to-read-this-file" are getting torn apart overnight we're about to pay off 40 years of tech debt whether we want to or not
10
27
934
funny seeing two things trending on the exact same morning: 1. nvidia getting caught with encrypted firmware on RTX Spark that throttles linux inference down to 35% 2. some absolute madman on github running the 2.78-trillion parameter Kimi K3 on a single CPU in 8.2GB of RAM using raw C99 nvidia is practically begging open source devs to route around CUDA at this point
2
6
787
tweaking your `.agents` and `CLAUDE.md` folder is the new configuring neovim dudes are spending 14 hours a week wiring up 5 parallel git worktrees and 23 markdown files that roleplay as a "VP of Engineering" and "Release Manager" just to ship a todo app with a broken login button
1
5
428
i can tell within 0.3 seconds if a landing page was one-shotted in cursor it's always: - TINY TRACKED-OUT EYEBROW TEXT - Serif headline with *one* italic orange word - 3 bento cards with lucide icons found two repos blowing up on github this week (`impeccable` and `hallmark`) that literally just lint your CSS/DOM for these exact AI tells and force your agent to fix them. lowkey genius.
4
4
468
genuine question for people shipping daily: are you spending more time in cursor or claude code right now? i used to be 100% cursor but lately i find myself living in the terminal 80% of the time and only opening the IDE to review diffs
8
5
392
wild how different AI video is evolving in china vs US right now on my timeline here everyone's still posting 5-sec sora clips of cyberpunk rain meanwhile on douyin some random account with 22k followers just dropped ep 1 of a 2-minute AI romance anime and got 455k likes in a day. another channel is on episode 9 of a 45-MINUTE AI sci-fi show (700k likes). they aren't treating AI video like vfx demos anymore, they're running full soap opera factories
1
3
239
spent 3 days wiring up a 6-agent LangGraph workflow last month just to watch them apologizing to each other in a loop for $4.20 a run ripped the whole thing out and replaced it with a 90-line bash script + one claude call + a strict linter runs in 2 seconds and actually works. why are we overcomplicating this
3
234
keep seeing the same tells on AI-built landing pages caps label above the H1, one colored italic word, split hero, gradient boxes two open-source skills going after this in claude code / cursor / codex: impeccable: a design linter + commands like critique, polish, bolder hallmark: material themes + a headless chrome layout check across viewport widths links in reply
5
255
a 35-year-old product manager in wuxi, china: 10+ years in tech. laid off 2 months ago. 12 interviews last month. zero offers. learned design, marketing, used AI to code, up till 2-3am every night. tried a factory job. 8am to 10pm. lasted 2 days. now drives the ¥300k EV he bought back then for deliveries. ~¥200 ($28) a day, before charging and insurance.
6
9
404
best closed model: Claude Opus 5.5 best open model: DeepSeek V4.1 Flash 1017 best closed video model: Seedance 2.5 best open video model: Minimax H3 best image model: GPT Image 2.5 best decision model: KnowLine-31B-Gen1 best TTS model: ElevenLabs Eleven v4 Turbo best ASR model: ElevenLabs Scribe v2 best embedding model: Voyage 4 Large best music model: Suno v6
3
10
290
stack overflow monthly questions peak: ~200,000 now: under 969 down 99.5%
2
8
242
story going around chinese twitter: she's 45. master's in rail engineering. helped build high-speed rail. laid off in 2023. hasn't found a job in almost 3 years. applied for a ¥4,500 admin job. the gen z HR asked why a rail engineer wants their job. scrubbed her degree to apply as a mall cleaner. the agency said she "looks too much like a manager." now she keeps the lights off to save power and does fake e-commerce orders for ¥0.30 each.
2
3
224
laid off from big tech frontend at 35. outsourcing firms say you're too old. chinese twitter's answer: go run automated cranes in a chip fab in hefei or wuxi. 2 weeks of training. 4 days on, 2 off. ¥13k a month take-home. ex-programmers there say a job with no powerpoint reviews is heaven.
8
14
444
fujifilm is spending ¥10 billion to 5x its magnetic tape production yes. tape AI churns out so much data that nobody will open again for years keeping it on spinning drives costs too much power so it goes on tape the future of AI is being archived on 1980s tech
7
11
328
github is adding 4 million CPU cores and 160 PB of storage commits hit 7.4 billion in september. 5x last year uptime dropped to 84% at one point it's the agents. push a whitespace change, fail, rename a variable, push again, rerun CI forever real humans' PRs now wait in line behind ten thousand bots stuck in retry loops "AI boosts productivity." first it took down the platform
2
7
303