Gemini 3 Pro testing is unbelievable, and World Models are BACK! [AI NEWS]

World Labs’ “RTFM” (real-time frame model) is pushing the idea of controllable “world models” into interactive, browser-demo territory—rendering a shifting 3D-like environment frame by frame without an actual 3D scene behind it. Users can steer a viewpoint with game-style control dials or a mouse, while the system generates new images in real time as the perspective changes. The catch is that the illusion is constrained: movement is limited, reflections and fine geometry degrade (mirrors blur into smeary, artifact-prone output), and higher-quality settings can feel choppy. Still, the core breakthrough is the learned-renderer approach—an end-to-end autoregressive diffusion transformer trained on large-scale video data that learns 3D geometry, reflections, and shadows purely from observation. Even with a single NVIDIA H100 GPU required for operation, the demo signals where “world model” interfaces are headed: interactive hallucinated spaces that feel navigable, even if they’re not truly physically simulated.

Google’s Gemini thread runs in parallel, with three distinct directions getting attention: a web-capable “Gemini Agent” prototype, a flood of unofficial Gemini 3 Pro-style coding demos, and broader multimodal and agentic upgrades across the AI ecosystem. The Gemini Agent concept centers on task execution inside a browser—navigating to requested sites in Google AI Studio without full file-system or computer-like access. That matters because it points toward practical productivity workflows for everyday users already living in Chrome, where an agent could potentially act on behalf of someone across routine web tasks.

The biggest buzz, though, comes from leaked or semi-public Gemini 3 Pro demonstrations shared by community users, often tied to an “ECPT” checkpoint and accessed via A/B testing in Google AI Studio. The examples lean heavily into code generation that produces real, interactive artifacts: an Eiffel Tower built from voxels via HTML/CodePen code; Space Invaders clones with CRT-like effects, particles, screen shake, and more polished gameplay; and animation-oriented outputs like a polar bear biking under a starry sky where limbs, motion, and even scarf flapping are handled with surprising consistency. Other demos go further—simulated iPhone 3G and Game Boy experiences with working apps and games, OS-like recreations with functional utilities, and a “seahorse emoji” test where the model avoids a common trap by not forcing a nonexistent emoji. The most striking educational example is a custom DNA-unzipping learning animation with labeled elements and real-time strand construction, framed as a tool that could be customized for different learners.

Beyond Gemini, the transcript highlights a broader acceleration: OpenAI’s Sora 2 is being integrated into production pipelines via Nvidia AI’s partner tooling; ChatGPT’s iOS app reportedly supports native video input by dragging clips into prompts; and Anthropic’s Claude adds “customizable skills” that load brand guidelines or slide-deck instructions from uploaded files to constrain outputs and generate structured deliverables. The overall throughline is clear: AI is moving from text-only answers toward interactive agents, code-generated software, and controllable media—while raising urgent questions about verification, misuse, and responsible labeling as video realism improves.