Google’s SIMA 2 AI Plays Games! + Nano Banana 2 Absurd Demos!

Google’s SIMA 2 is being positioned as a step-change in “agentic” AI for virtual worlds: a multimodal system that can watch video, interpret images and audio, understand a user’s goal, and then navigate and complete multi-step tasks in complex game environments—including ones it hasn’t seen before. The demos emphasize more than scripted tool use. SIMA 2 can open inventories, follow instructions, and iteratively improve by playing on its own, transferring learned concepts across different games. In one example, a user draws a rough spaceship and the agent figures out what that sketch represents, then searches the world to find the corresponding object. In puzzle-style scenarios, it still needs occasional correction—such as being told to place the right piece—before it proceeds to complete the task.

The gaming testbed matters because it forces AI to handle messy, real-world-like interfaces: 3D navigation, inventory management, upgrade paths, and systems that change based on player actions. SIMA 2’s control loop is described as more elaborate than a standard large language model with tools. The system is shown working alongside Genie 3, Google’s world simulator/generator, where the agent can be instructed to interact with objects (like swimming to an orange coral) and then explore, look around, and adjust its behavior based on what it sees. The underlying architecture is framed as multiple components—such as a core agent and a task setter—plus a reward model and a self-generated experience loop that supports learning through repeated play.

Alongside SIMA 2, the transcript spotlights Nano Banana 2, an image model that briefly leaked into public access before that access was revoked. Even with the caveat that official releases may be more censored, the shared examples aim to show major gains in image editing and reasoning. Nano Banana 2 can enhance screenshots into high-fidelity remasters (including game-like scenes such as Crash Bandicoot and Grand Theft Auto: Vice City), preserve fine details like textures and UI elements, and perform structured edits—like rearranging torn paper into a coherent sentence by rotating pieces. It also appears to handle “instruction following” on images: drawing a line to indicate a ball’s path, generating an above-ground angle from a marked building, and adding objects to a scene based on doodles.

The transcript also claims improved benchmark performance, including difficult visual cases like a full glass of wine up to the top and correct clocks rendered in Roman numerals. Math accuracy is not perfect—one multi-step whiteboard problem is described as still incorrect—but the model is portrayed as closer to correct than its predecessor. Safety features are mentioned via SynthID, described as an invisible watermark intended to detect AI-generated content.

Finally, the roundup touches other fast-moving model developments: GPT-5.1 is described as deployed in ChatGPT with a small quality bump and faster responses, while Gemini 3 is framed as imminent, with demos showing it coding a functional Windows-like interface in HTML. The transcript closes with a warning about AI-enabled cyber risk, citing Anthropic’s claim of disrupting an AI-led espionage campaign targeting major institutions, and arguing that malicious, targeted attacks will become a growing challenge as these tools spread.