Xianzi Insights｜Field Notes on AI Video & Indie Development｜TangYi Studio

Hands-on notes on generative AI, film production and solo development. Every number is measured, every pitfall is one we actually fell into.

• GPT-6 Astra Made Me Realize I’d Been Settling

  I thought a new AI lamp would catch a few more bugs. Then GPT-6 Astra uncovered issues in our website and apps and brought my Blender scene closer to its design. Suddenly, I was questioning how quickly I had accepted “good enough.”

• A Party on the Edge: One Person, One AI, Shipping a Party Game to Cloudflare Workers

  A shipping app with 1,428 cards, 17 tables and real-time multiplayer, built on Cloudflare Workers with Claude. Four traps nobody warns you about, and where the AI confidently gets it wrong. Every number measured, not estimated.

• AI Built Me an App, and Got It Wrong Dozens of Times

  This isn't a complaint post. Those dozens of mistakes prove exactly one thing: in this era, the most valuable skill isn't "knowing how to code" — it's "being able to spot what's wrong." And chances are, you already have it.

• Battle of the AI Lamps: Fable 5.1 vs GPT-6 Astra — Which One Should I Rub?

  Two new AI lamps launched at once, and the real question isn't who scored higher. Claude Code or Codex? How far does $20 actually get you? And the upgrade order nobody tells you about.

• OpenRouter's Numbers: AI Agents Now Use 5× More Tokens Than Humans — What 7.3 Trillion Tells Us

  In six months, tokens burned by agents went from 0.51 trillion to 7.3 trillion — five times direct human use. You think you're using AI; you're just the one who pressed the first button. Why agents burn so much, and who defines AI's direction once AI is its own biggest user.

• Qwen3.8-Max vs Qwen3.8-27B: A 25× Gap, and We Picked Wrong Once

  Running a model locally comes down to two numbers: the file's GB and your card's GB. Starting from the day we nearly told readers to download a 2.4-trillion-parameter model — which 27B quant to pick, how to set up RAG, and how to tell the AI isn't making it up.

• From Copyright Suits to Studio 2.0: Suno Is Rebuilding Itself from a Gacha Machine into a Workstation

  Warner settled, Sony and Universal are still suing, GEMA won in Germany, BMG signed an alliance — that's Suno in August 2026. The same month Studio 2.0 shipped: MIDI, stems, automation, a session-aware assistant. Here's how people who make video should use it.

• We Built a Scroll-Driven Frame-Sequence System — and Open-Sourced It for Free

  The background that scrolls with you on the tangyi.mx homepage isn't a video. It's scroll-frame-sequence: 535 lines, zero dependencies. How to budget mobile memory, order three-tier loading, and fix AI footage that won't loop — all measured numbers.

• How One Person Built and Shipped a Card Game App from Zero in Three Months

  A deck of question cards made seven years ago so interviews wouldn't go quiet became a game that passed App Store review. No engineer in between. Why an app instead of merch, how the three months went, and the holes — including three Apple rejections.

• Qwen3.6-Plus vs Seedance 2.0: One Writes Code, One Makes Video — Stop Putting Them in the Same Table

  The first version of this piece compared a language model with a video model in one table. Wrong. Rewritten to say what actually helps: how a small Taiwanese business sorts its work into "text-brain" and "image-brain" piles — confidential work runs at home, good-looking work rents from outside.

• Seedance 2.5: 30 Seconds of Native 4K from One Prompt, and 50 References to Feed It

  ByteDance's Seedance 2.5, announced in late June: 30 seconds of native 4K, up to 50 reference inputs, 10-bit colour, sound generated in sync with picture. For people who make video the point isn't resolution — it's that characters finally stay consistent.