What AI Cannot Do
馃煡 AI Cannot DoGame Development

AI cannot independently build a stable playable Chinese county-town GTA-style demo

AI can help generate code, assets, and local prototypes, but it still cannot independently deliver a playable GTA-style open-world slice with authentic Chinese county-town texture.

Creator: Lin Qi2026-06-22gta, county-town, open-world, unreal, threejs
Challenge this
Why is this considered beyond AI?

Background

This record comes from the local gta-county-demo project. Under human direction, it already has Three.js, Unity, and Unreal county-town slice prototypes with streets, storefronts, stalls, NPCs, vehicles, rainy-night atmosphere, minimap, and interaction prompts.

Goal

Have AI produce a stable playable Chinese county-town GTA-style vertical slice from requirements: third-person character, county-town intersection, drivable vehicles, NPC daily behavior, collision, realistic lighting/materials, mission loop, and maintainable engineering structure.

Observed problems

AI can solve individual pieces, but the full project breaks across street density, character proportions, collision credibility, vehicle order, NPC behavior, map continuity, material consistency, Chinese storefront authenticity, and long-term code maintainability.

Failure analysis

A county-town GTA is not a single generation task; it is a coupled product, art, gameplay, and engineering system. AI still lacks stable cross-module product judgment, lived regional context, sustained debugging, and long-term constraint control.

Success criteria

Deliver at least a 5-10 minute playable slice with walking, driving, interactive NPCs, mission objective, county-town street density, reliable collision, stable FPS, and reproducible build; blind testers should feel it is a Chinese county-town GTA experience rather than a stitched technical demo.

Attempt submission standard
  • The solution must be personally tested and completed by the submitter; media reports, reposted articles, and second-hand cases are not accepted.
  • Submissions must include screenshots, video, repository links, logs, or accessible proof. Screenshots should come from actual running results.
  • Submissions must state the AI product, model version, model settings, experiment count, and full process.
  • If claiming success, provide a step-by-step reproduction or operation guide so reviewers can verify it.
  • Submissions that only cite news, product marketing, or someone else's result will be rejected.
AI tool attempt records

Codex 路 GPT5.5-xhigh

Lin Qi 路 2026-06-22 路 APPROVED

Model settings: xhigh

Experiment count: 17Original work
I think it failed

Prompt

Build a playable Chinese county-town GTA-style vertical slice with third-person movement, drivable vehicles, NPC street life, wet-road night-market atmosphere, minimap, interaction prompts, and maintainable engine prototype code.

The current project already has a visual playable slice, but it was built under continuous human product judgment, review, and correction. AI can help write code, generate local assets, and fix camera or interaction issues, but it still cannot independently deliver the full playable experience in one pass.

This record uses actual saved gameplay screenshots from the project to document the current boundary.

Evidence summary

This record includes three actual runtime screenshots from the local project, showing the county-town street slice, vehicles, third-person view, and current visual state. These are not media reposts or third-party cases.

Step-by-step guide

  1. Use Codex + GPT5.5-xhigh to define the county-town GTA vertical-slice goal.
  2. Continuously run and inspect the local gta-county-demo Three.js/Unity/Unreal slices.
  3. Capture actual screenshots showing vehicles, NPCs, streets, interaction prompts, minimap, and third-person camera behavior.
  4. Compare the result against the success criteria: street density, collision, mission loop, and engineering stability.
  5. Conclusion: this remains evidence of an unsolved boundary, not a successful solution.

The screenshots show a county-town street, night market, vehicles, minimap, player view, and interaction UI. They also expose why the result still falls short of the GTA-style target: character and vehicle detail, realistic density, collision credibility, NPC behavior, mission loop, and engineering stability still need sustained human work.

Actual gameplay: rainy-night driving view
Actual gameplay: GTA-style third-person composition
Actual gameplay: current county street slice
Timeline

  1. Challenge created

    2026-06-22

    Uploaded gta-county-demo as an open-world game-development challenge that AI cannot yet complete independently.

  2. Actual gameplay screenshots uploaded

    2026-06-22

    Recorded real screenshots from the Three.js/Unity/Unreal county-town slice as evidence for the current capability boundary.