Seed-2.1-pro did something this week that recalibrates expectations: an indie developer used it to build a browser-based 3D battle game with five skills and enemy AI from scratch, and the path from a stick-figure prototype to a playable version spanned multiple iteration rounds. The question worth asking isn't what the model can code, but whether it can manage a project.

What this is

The project uses Three.js (the standard library for running 3D graphics in the browser) for rendering and scene work, with a stick-figure protagonist. Five skills — charged fireball direct fire, lightning strike from above, ice spike zone with freezing rain, parabolic water projectile, and meteor with a massive explosion preceded by a warning indicator — each carry distinct cast logic and hit feedback. Enemies operate on a "wander—aim—draw bow—fire arrow" state machine (a script that switches actions based on context), and both sides have HP, win/loss resolution, and instant rematch.

This wasn't generated in one shot. The developer first had the model validate gameplay with the stick figure, then progressively swapped in four modules: scene, skills, enemies, and character. The fireball evolved from a blob of light into something with speed feel and screen-shake on hit; performance issues from uncapped rapid-fire were contained with lifecycle and count guardrails. The full project only took shape across multiple iteration rounds.

Industry view

Supporters see this as the real watershed for AI coding. Previous AI excelled at writing single-file code snippets, not at multi-module projects that require continuous maintenance. Models like Seed-2.1-pro can hold the same project context across many dialogue turns and attach new features to existing interfaces — that's the leap from demo to production.

But we have reservations. The source is the developer's curated success case, and the real pain points of 3D debugging (fireballs blowing out to white, geometry clipping, inverted orientation, bows flying into the sky) are brushed off with a single "problems kept cropping up one after another." Whether AI can reliably locate bugs in graphics projects that lack clear error messages, or whether the developer is mostly reading code and making judgments, still lacks systematic comparison. Running a demo isn't the same as shipping one.

Impact on regular people

  • For enterprise IT: The barrier to internal tools and small games keeps dropping, but teams still need people who understand fundamentals like Three.js and state machines — otherwise they can't judge whether AI-generated code actually runs.
  • For individual careers: For programmers, the AI coding competition is shifting from "can you write code" to "can you break down requirements, read AI output, and locate the errors AI can't solve."
  • For consumer markets: The cost of building browser-based 3D small games continues to fall, and more indie developers and small teams will emerge. Networking, save systems, and commercial-grade art — the pieces required for full commercialization — remain hard for AI to replace in the short term.