LLMs are moving from generating artifacts to creating hyper-custom worlds on demand, but can't audit their creations because they can't natively perceive videos
We're starting to leave the territory where you'd test an LLM by e.g. “create an svg of pelican on a bicycle”.
@karpathyAndrej Karpathy
Context & Ripple Effects
This extends coverage of LLM progress from falling prices and widespread multimodal vision to a more demanding use case: generating environments tailored to an individual request. The intervening record also identified reasoning and useful coding agents as meaningful capability steps, including the rise of useful coding agents.
The limiting factor is no longer only producing an output; it is determining whether a generated environment is coherent and meets its requirements. That makes the gap between generation and native auditing central as LLM use moves beyond discrete artifacts.
First-order effects
Teams using LLMs to generate custom worlds must add external checks or human review, because the model cannot natively inspect and audit its own result.
The practical benchmark shifts from making a single artifact to reliably creating and validating a larger, internally consistent environment.
Second-order effects
Tool builders have an incentive to pair generation with inspection, testing, and evaluation layers rather than compete on generation alone.
Users seeking dependable outputs will favor workflows that expose failures and support iteration, echoing the hands-on experimentation associated with vibe coding as capability exploration.
Third-order effects
If on-demand world creation becomes common, the observation–synthesis boundary becomes a core competitive divide: systems that can verify outputs may be more useful than systems that merely generate them.
This points toward governed generation, where deployment value depends increasingly on controls and auditability alongside model creativity; the pace of that shift depends on whether native perception improves.
The trend:Generative AI is moving from producing isolated artifacts toward interactive, personalized environments, making verification a central product capability.
How about no? People routinely post AI-generated “tech demo” slop here with no gameplay and call it a game. Whatever “AI twitter” thinks about anything even remotely artistic is just information noise.
AI is starting to do amazing, superhuman things, and even the smartest AI experts are amazed “that it even does anything at all.” AI is supernatural magic.
I am starting to invest in more real world businesses because in the next decade the Dead Internet Theory will become true whether you like it or not. And more importantly, people will start seeking out reality again.
State exploration is a different set of skills than state generation. LLMs are excellent state generators. Not so much state explorers, especially for modalities outside of text space. About time somebody fixed that... :)
The model's limited visual capability is the biggest bottleneck when building games with AI. It's probably the one thing preventing AI from recursively building and refining games at speed. Once that's unlocked, we will likely see a cambrian explosion of games and interactive exp…
The exciting part to me isn't even the gameplay itself but more how believable and explorable these generated worlds are becoming. spatial experiment: https://robotroommates.com/
Same experiment, different book: I gave Opus 5 the first paragraphs of Harry Potter and asked for a Three.js render. It built a 4-minute procedural film, 7.6k lines of TypeScript. Zero assets. Every texture drawn on a 2D canvas at runtime, every character assembled from boxes and…