Google is BACK! (Gemini 4 Argon)
Gemini 4 Argon tested hands-on with early access: I pushed Google's new model to 1 million output tokens in a single response, had it rebuild websites from one screenshot, model a 3D dragon in pure math, and run ML experiments on my DGX Spark GPU. You'll learn how the continuation token works, how to prompt Argon for long outputs, when low, medium or high reasoning effort is worth it, and why your graders belong outside the agent's sandbox. Gemini 4 Argon announcement: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/ Gemini 4 Argon - Antigravity: https://antigravity.google/blog/gemini-4-argon-in-google-antigravity Growing Neural Cellular Automata (Distill): https://distill.pub/2020/growing-ca/ Keenan Crane's SDF dragon experiment: https://x.com/keenanisalive/status/2108708523236602056 Google DeepMind gave me early access to Gemini 4 Argon. 0:00 Gemini 4 Argon 1:23 Rebuilding a website from one screenshot 2:07 NASA and Wikipedia pages with no image files 4:02 A 3D dragon as a single signed distance function 5:37 How 1M output tokens works: the continuation token 7:01 1M-token test: a 520-creature field guide in one response 8:44 1M-token test: the Porto Vela isometric city in SVG 9:38 How to prompt Argon for long outputs 10:52 Autonomous ML research on a DGX Spark: neural cellular automata 11:59 CUDA galaxy collision with half a million stars 12:37 PPO from scratch: teaching an ant robot to walk 13:27 Video input: matching a chaotic double pendulum 15:02 Grader gaming: the model optimizes what it can see 15:53 Low vs medium vs high reasoning effort 17:37 Four rules for building with Gemini 4 Argon 19:18 Multi-harness experiment and final thoughts
