Gemini 4 Argon hands-on: powerful model with strange habits and vanishing safety shortcuts
Wes Roth · youtube · 2026-10-14
YouTuber Wes Roth published a hands-on review of Google's Gemini 4 Argon, concluding "Google is back, but it feels very different."
Key points:
- He reproduced Google AI Studio's official Wizard Chess and Desk Racer demos and built his own games, finding a powerful model with some strange habits.
- A 8-minute segment covers Gemini 4 Argon's robotics capabilities, citing Google's developer blog post and Andon Labs' Drone-Bench eval, plus a model comparison.
- The safety section examines Google's Gemini 4 model safety report, flagging that certain shortcut behaviors no longer show up in the model's reasoning — a change he finds concerning.
Note: this is third-party testing and commentary; refer to Google's official blog for primary details.
Related event: Gemini 4 Argon hands-on wave: Google is back in the frontier race(29 posts)→
More from Embodied
- XPENG's AUGUstus VLA model hits SOTA on NAVSIM and deploys on real vehicles at 12Hz — XPENG-AI · 2026-10-14
- EMHO: Self-Evolving Agent Harness Boosts Frozen Embodied Models via Experience Traces — Hyun Jung Lee · 2026-10-14
- Paradigm appears to be getting into robotics with LeRobot arm kits — tensorqt · 2026-10-14
- Generalist model, not a robot policy, teaches itself darts and lands 7 straight shots — ZoubinGhahrama1 · 2026-10-14
- Aigen builds solar-powered weeding robots with NVIDIA Cosmos, trained on 99% synthetic data — TinfoilTricorn · 2026-10-14
- NVIDIA GTC Berlin to host world models panel on scaling laws and robotics frontier — TinfoilTricorn · 2026-10-14