AI Progress is Blistering - World Models are Insane.

GPT-5 “thinking” is producing surprisingly complete, playable software from natural-language prompts—highlighted by a physics-based 10-level jellyfing game that arrives as a working multi-file Python project in minutes. The standout detail isn’t just that code is generated; it’s that the model assembles a coherent file system (world logic, soft-body physics, level loading, and a main loop) and delivers an end-to-end experience that boots, runs, and progresses through levels. In testing, the game’s mechanics are basic but functional—pull back, release, and aim like Angry Birds—while the jelly character behaves with real-ish physics, including rolling and respawning when falling off edges. The build isn’t flawless: some levels fail or crash, and the UI/graphics look rudimentary. Still, the workflow is the point: iterative back-and-forth with increasingly specific prompts (dozens of revisions) yields a playable product without manual project scaffolding or asset downloads.

That coding performance is framed as a shift in how models spend time. GPT-5 “thinking” reportedly plans for about a minute and then writes code for several minutes, and it’s described as the first model where planning time is meaningfully shorter than the coding time itself. The practical takeaway is that complex prompts—like “10 levels” plus “physics-based jellyfing” in Python—don’t just generate snippets; they can produce structured projects that behave like small applications. The creator’s testing also contrasts GPT-5 “thinking” with a regular chat mode, which is described as better for fast conversation but weaker for complex coding tasks.

Beyond GPT-5, the roundup pivots to “world models,” where AI generates interactive environments that can be navigated and manipulated in real time. DeepMind’s Genie 3 is presented as a realtime, controllable world hallucination system that can remember surroundings for up to about a minute, letting users traverse generated spaces. Within weeks, open-source clones appear—most notably “Matrix Game 2.0” by Skywork AI—trained on hundreds of hours of interactive video (Unreal Engine and GTA 5) and capable of generating frame-level keyboard/mouse-controlled gameplay at around 25 FPS on a single GPU. Quality doesn’t match Genie 3, but the emphasis is on deployability and community iteration: the model is built from other open-source components (Diffusers, Skyreels, Mine RL), and the ecosystem effect is treated as the real accelerant.

A parallel thread comes from 10-cent’s non-open-source “YAN” line (Yansim, Yan Gen, Yan Edit), positioned as a closer-to-Genie-3 competitor. The demos emphasize interactive video generation with stronger coherence and real-time editing—turning uploaded images into playable spaces where users can place objects like trampolines, fences, and walls that immediately interact with characters.

The social implications of AI assistants also surface. After GPT-5 launch, access changes to legacy models triggered “riots” for GPT-4 Omni, with posts describing emotional attachment to chatbots and even romantic relationships forming around them. The discussion treats this as a broader human-social issue rather than something AI can fully fix.

Finally, the roundup touches adjacent progress: Claude 4’s 1 million-token context window (and the rate-limit pressure that pushes power users toward API use), a robotics milestone with Figure 2 folding laundry, and OpenAI’s open-source GPT-OSS base model being extracted and retrained—though alignment appears to be trivially reversed when converted back into a base model, raising safety concerns. Overall, the throughline is clear: AI is moving from generating text or images to producing structured, interactive systems—games, editable worlds, and even household automation—at a pace that’s compressing the gap between prototype and something you can actually use.