I watched a 25-year-old founder type “Day 1 of building GTA 6” and my feed hiccupped. You feel the gamble the moment the repo goes public — ambition and embarrassment in the same click. I keep asking myself: is this bold engineering or a fast way to waste a week of cloud credits?
He pushed a repo and a tiny blue oval to the feed — then promised to race Rockstar
On X, Ziwen Xu announced he was using Anthropic’s Claude to “vibe code” a GTA-scale game and linked a public GitHub. You can go read the commits yourself: early proof-of-concept assets, test scenes, and a steady stream of videos showing incremental progress. I followed because as a reporter I want the smell of the project — not the press release.
Xu is the founder of Hyperecho, which sells AI workers, so you should expect him to run the numbers around model usage and dev velocity. He upgraded to Claude Max 20x and admitted he burned a chunk of the weekly quota within hours. That confession is a tiny authority cue: this is not a casual tinkering session but a focused, expensive experiment.
Can AI build a full-scale video game?
Short answer: parts, fast. You can get characters, collisions, and rough city blocks from large models and prompt chains, but a cohesive open-world with missions, reliable AI, and tight controls is an engineering mountain. What you see in early clips — a 3D blue oval hopping between gray boxes — is a scaffolding stage. The climb from scaffolding to a believable, playable city is measured in weeks of human iteration and an army of tests.
A skyline mistake exposed a bigger truth — the AI follows prompts, not context
Xu posted a clip where the agent “built downtown LA skyscrapers” even though the game should be Florida. That error is funny on the surface, but it’s a reminder: the model stitches together training data and heuristics, often ignoring local logic. You end up with plausible geometry, wrong geography, and a lot of patching.
I saw the same pattern when founders vibe coded apps last year — the AI can generate an impressive fragment but can’t yet keep a multi-part narrative consistent without tight human direction. Think of it like an orchestra led by an AI conductor who sometimes reads the wrong score.
What does vibe coding mean?
Vibe coding is heavy reliance on natural-language prompts and AI assistants to generate and debug code. Tech figures such as Jack Dorsey have tried it publicly, and Andrej Karpathy famously warned that autonomous agents can be “net unhelpful” in some workflows. The method speeds prototyping, but it also produces brittle systems that need constant oversight.
The project moved from blue ovals to NPCs, cars, and weapons within days — momentum matters
By day three the feed showed pedestrians, traffic, and even weapons. Those are milestones, not proofs. NPCs that walk and cars that roll are visible wins, but they don’t guarantee emergent gameplay, mission design, or bug-free multiplayer.
Xu posted the repo publicly and has been updating daily. That cadence creates a momentum loop: followers, critique, and quick fixes. If you watch the timeline you can almost see the product evolve — which is addicting in the same way a reality show is addictive.
When will GTA 6 actually arrive?
Rockstar’s official schedule still controls the real release date, and no amount of hobbyist coding will legally or practically make that irrelevant. Xu’s stated goal is to beat Rockstar to launch — an audacious PR angle. You should read that as a hook, not a forecast.
Community prompts and public experiments are shaping how code gets written — and tested
A private investor’s idea to set up a community-funded Fable run with a loop prompt inspired Xu. People suggested a continuous prompt like “loop until you’ve created a GTA-VI-caliber open-world game” using Anthropic’s Fable 5 or Mythos variants. What you end up with is a social experiment: code by committee, accelerated by paid model calls, visible to the world.
There are ethical and practical risks. Model hallucinations, license friction with Rockstar IP, and runaway cloud spending are real hazards. I worry less about the mockups and more about the incentives that make founders chase attention. Sometimes a prototype is like a kid with a power drill in a museum: curious, destructive, and very public.
Will Xu beat the official launch or make something useful to the community? Either outcome will teach us where AI actually helps in building complex systems — and where human craft still wins. Are you rooting for the founder, the model, or the studio that actually owns the IP?