⚡Megatron LLM Hub
←Back to search
Website

October Lingers 秋 · 五味杂陈 5 reads translated from Simplified Chinese ↗ Dictation · Assisted · Title · Assisted English Synopsis open...

October Lingers 秋 · 五味杂陈 5 reads translated from Simplified Chinese ↗ Dictation · Assisted · Title · Assisted English Synopsis open yohaku ↗ It's already been over two weeks since my last recap, so it's time to catch up on what's been happening lately. Nyatabi Hits the App Store Last time, I mentioned running into so many headaches after a trip that I decided to build an app. At the end of my last post, I wrote, "Musubicho is still in development and not finished yet." Less than two weeks later, I renamed it Nyatabi (喵旅) and officially launched it on the App Store. That said, I haven't even been the first one to really test it out in the wild. I'm such a homebody that even though a trip sparked the project, I haven't had the chance to use all its features myself yet. I just love building products—created purely out of passion. It's live on the App Store now, and I put together a decent landing page for it. Feel free to check it out! (Though it's iOS-only for now.) Nyatabi — your AI travel companion for iPhone Plan with Musubi, explore together, split spending fairly and turn your route into memories. Your whole journey, in one local-first iPhone app. Nyatabi Beyond helping you map out trips, it lets an agent handle expense tracking and bill splitting on the go (either by snapping receipts or just telling it what you spent). In group chats, the agent can scout nearby restaurants or pull up your friends' locations. And once the trip wraps up, you can replay your route, generate an auto-edited highlight video, or export a detailed travel footprint report. Before the official release, I shot a 4-part short video series walking through the entire flow: pre-trip planning → snapping receipts to split expenses on the go → settling balances with minimum transfers → everyday roommate bookkeeping. On October 16, I posted a full 2-minute demo. Along the way, I also open-sourced the rounded-line icon set used in Nyatabi at icons.nyatabi.app . Engineering shortcuts vs. engineering rigor: refusing to be a human test runner, I wired up XCUITest across two physical iPhones so agents could handle the entire closed loop of "implement → verify → fail → fix" on their own (check out my technical post, " My AI-Native iOS Development & Verification Workflow ", for details). It could be handy whether you're traveling or just going about everyday life: I've always found life-logging footprint apps a bit awkward to use, so I baked my own footprint features right in to track daily routines as well. I don't want to turn this into a full-on product pitch, so I'll stop here. If you're interested, feel free to give it a spin. Nyatabi: AI Travel Planner App - App Store Download Nyatabi: AI Travel Planner by Moeru AI on the App Store. See screenshots, ratings and reviews, user tips, and more apps like Nyatabi: AI Travel Planner… App Store AI in the Trenches A Love-Hate Relationship with Claude I had a backup Claude alt account that managed to hold out for over two weeks, so I decided to upgrade it. Went from Pro to Max, only to get slapped with an instant ban—no refund, either. I was so anxious that night that I stayed up by myself until 5 a.m. Eventually, I couldn't help setting up yet another account, and that one's somehow still alive. On top of that, my $DRAM shares kept tumbling before I could break even, so those days felt like a brutal double blow from Anthropic and $DRAM. It also got me calculating the real cost of compute inflation: last year, a $20 Claude subscription felt like it could do so much; now, even the $200 Max tier barely scratches the surface. By comparison, Apple's $99/year developer membership is so cheap it's almost moving. OpenAI's "God Tier" Account and Kimi Falling Flat on Its Face There was a funny little detour along the way. Around the 20th, I noticed a bug in OpenAI: an old 20x subscription was still usable even after downgrading or canceling. I honestly thought I'd struck gold with some chosen "god-tier" account. But by day five of freeloading, the bug got patched and my quota went straight to zero. Then on the 23rd, I dropped $19 on Kimi, wanting to see how it compared to $20 Claude. I asked it to write a website footer—and before it even finished, barely ten minutes in, I slammed right into a 5-hour rate limit. Might as well just call yourselves QueueShot AI. When it comes to domestic models, everyone should just use DeepSeek—what are the rest of these garbage knockoffs even doing? DeepSeek is the only one that can actually play: no token sales, no bullshit. The API pricing is right there: take it or leave it. The TPS is right there too—DeepSeek 4.1 Flash pushes over 200 TPS, and at those prices, burn through as much as you want. Unlike Kimi—their greed this time is honestly shameless: zero compute power, but pulling that classic Lei Jun artificial scarcity playbook and taking users for fools to the extreme. Ryugi (流儀) — Open-Source Claude Artifacts After getting jerked around like that, I thought: why not just build my own version of Claude Artifacts? I genuinely love using that feature. For docs, artboard designs, and architecture sketches, it's just so damn good—a product genuinely made with care. But being locked into Claude's proprietary subscription was a dealbreaker. So what choice did I have? Build it myself. Claude Artifacts support five types, and the ones I use most are Artboard, Document, and HTML. Sick of taking Anthropic's nonsense, I cobbled together a standalone workbench called "Ryugi" ( ryugi.innei.dev ). Building this brought an unexpected bonus too: it completely changed my view of Codex's design skills. I used to think Codex was garbage at design, and just on the 10th I had tweeted about how miserable it was trying to write UI with it. But once I plugged it into Ryugi's Artboard workflow, it turned out to be surprisingly solid—I guess I was just using it wrong before. To my surprise, that tweet blew up, racking up 93K impressions and over 700 bookmarks. I've been using it for a few days now, putting it through heavy daily use and constantly refining it. The cloud environment is live too, so now it helps me brainstorm, work on designs, hash out architecture docs, and even knock out small feature requests on the side. The experience genuinely captures the vibe of Claude Artifacts now. Hooking this workflow up to other harnesses—like DeepSeek or Codex—makes them dramatically more useful. If you haven't tried it yet, you really should. Repository · Innei 75 ryugi 流儀 — a development workflow for coding agents: brainstorm with mockups, approve a spec, build it under TDD TypeScript Here are a few UI designs and animations I generated using DeepSeek, Codex, and Grok: Document · 历史回看 · Ryugi View “Document · 历史回看” on Ryugi Cloud, an artboard artifact. Ryugi 简单模式 · 描述显隐 · Ryugi View “简单模式 · 描述显隐” on Ryugi Cloud, an artboard artifact. Ryugi LobeHub · AI Provider 接入 · Ryugi View “LobeHub · AI Provider 接入” on Ryugi Cloud, an artboard artifact. Ryugi 室内乐与夏日小品 · Nine Pieces · Ryugi View “室内乐与夏日小品 · Nine Pieces” on Ryugi Cloud, an HTML artifact. Ryugi LobeHub Bids Farewell to AntD On the work front: I've been here for almost a year now. Back when I first joined, I declared that I was going to ditch AntD once and for all. That turned into a ten-month marathon, but on the 9th of this month, I finally pulled it off—migrating everything over to @base_ui + StyleX (meaning lobe-ui v6). Looking back at the numbers, the engineering scale was pretty wild: lobe-ui rewrote 54 components on top of Base UI, and AntD dependencies in lobe-editor , lobe-charts , and lobe-icons were completely eliminated. The main repo has over 11,000 TypeScript files. Replacing the underlying component library spanned 28 PRs and touched 3,242 files, with the single largest PR modifying 1,040 files. Along the way, we also migrated our CSS-in-JS setup to StyleX, slashing the original 179 emotion files down to just 3. The most satisfying part was running 5 agents across parallel branches: in three and a half hours, they plowed through 84 commits. We then ran regression tests across 45 routes using 1,000 screenshot comparisons and computed-style diffs, reliably catching 4 regressions. Getting rid of AntD paid off right away. Bundle sizes shrank dramatically, and pages feel noticeably snappier. Thinking back, before AI agents existed, a core refactor of this scale would have easily taken an entire team a full quarter of grinding. But with today's models being so capable, I managed to knock out this massive project almost entirely on my own. The Story Ends Here Throughout October, blind dates from family and friends were nonstop, but most of the time chatting was pretty draining. Earlier in the month, I tweeted a rant saying some WeChat chats feel like "an LLM running on a potato server, taking hours of inference just to spit out a few characters, and if you don't send a request, you'll never get a response." That tweet unexpectedly racked up over 70,000 views, and predictably enough, I got hit with the "you're a nice guy" card in real life soon after. But just a few days later, through another blind date, I unexpectedly met someone I could genuinely click with—something that hadn't happened in a long time. I find myself wanting deeper conversations lately—talking about fundamental values, life outlooks, and thoughts on the future. In day-to-day life, finding someone who can actually match that energy is rare. But that night—which also happened to be our last conversation—we talked about so much. Beyond having plenty of shared interests, what hit hardest was a real resonance in who we are deep down. Just knowing someone on that exact same wavelength actually exists was a wonderful, lucky surprise. Unfortunately, she felt our personalities were just too similar to be a good match. In the end, we parted ways gracefully, and that was where the story ended. Turning My Resume into a Canvas Another project I spent a lot of time on was completely rebuilding my personal resume page for Programmer's Day on October 24, launching it at resume.innei.in/lab/ . The spark was simple: surfing online recently, I saw a few personal sites and resumes that looked amazing, so I wanted to bring my own ideas into the mix and polish up something of my own. A traditional single-page format can't hold very much, so I turned it into a canvas you can freely pan, zoom, and explore—connecting all the projects I've worked on and milestones I've reached over the six or seven years since college into one continuous timeline. Working on this canvas also ended up unlocking some new skills. To give the page some fitting background music, I dove into arranging with realistic virtual instruments, tweaking synthesis, and fine-tuning the details (huge thanks to @wibus_wee for walking me through Ozone mastering), and I even built an interactive sheet music display into the canvas. Cramming a bunch of music arrangement and audio mastering just to build a resume was an unexpected bonus. Once I get around to picking music theory back up properly, I really should build an app for instruments and music theory. Thanks for reading this far. Also, I recently updated the background on this blog with dynamic nature scenes that change with the seasons and time of day. I even split the effect out into its own standalone preview page ( innei.in/dev-demos/backgrounds ); if you're interested, feel free to click in and play around with it.