Making games in Go: 3 months without LLMs vs. 3 days with LLMs

Original Article ↗ Hacker News Discussion ↗

“Where are all the LLM-made games?”

  • Some argue that if a solo dev can build a game in 24h, LLMs should enable polished Steam-ready games in days, yet there’s no visible explosion of quality titles.
  • Others counter that ~50+ games already release on Steam daily; the bar for visibility and success, not raw output, is the real constraint.

What’s actually hard about making games

  • Strong theme: “code is not the bottleneck.” The hard parts are:
    • Fun and novel mechanics, balance, pacing, and content.
    • High-quality, coherent art, animation, sound, and UX.
    • Marketing, discoverability, risk, and post-launch support.
  • Counterview: for many non–game dev engineers, coding is a bottleneck; LLMs help them cross engine/graphics learning curves.

Impact of LLMs and Steam release stats

  • Some see 2024 Steam releases as noticeably above trend and attribute some of that to AI, especially cheap NSFW/shovelware.
  • Others say growth is modest vs pre-AI trajectory; if LLMs were truly 10× multipliers, releases would spike far more.

LLMs as coding assistants, not designers

  • LLMs excel at:
    • Refactoring or re-targeting existing code (e.g., cloning one card game backend to another).
    • Boilerplate, glue code, and exploring unfamiliar languages.
  • They struggle with:
    • Greenfield, ambiguous design.
    • Deep gameplay iteration and debugging without strong human guidance.
  • Comparison in the article is criticized as unfair: the “3‑day” version reused code and learnings from a 3‑month first attempt.

AI for assets and playtesting

  • Image models are widely seen as useful but inconsistent for reusable assets (e.g., sprite sheets, consistent characters, multiple poses).
  • Many consider current AI art “cheap” looking, but note that many low-budget games look bad anyway.
  • Prejudice and potential backlash against AI art still deter some devs.
  • Idea of AI playtesters sparks debate: some think data-driven models could help with balance and engagement; others doubt AI can judge “fun” or fear it will optimize for bland, hyper-engaging designs.

Go, WASM, and architecture

  • Several question using a Go “backend” compiled to WASM for a purely client-side card game, calling it overengineered versus plain JavaScript.
  • Discussion notes that static typing (Go, Rust) tends to work better with agentic LLM tools than dynamic languages, due to fast compile-time feedback.