Gemini 3, Poetry Jailbreaks, and Do We Even Need Safe Robots?

Google’s Gemini 3 burst onto the scene in November, shattering previous records on the ARC‑AGI benchmark with a peak score of 45. The model’s edge stems from Google’s vertically integrated stack, especially its proprietary Ironwood TPUs, which sidestep the industry‑wide GPU shortage and cut costs. This hardware independence not only signals a shift in AI compute economics but also challenges NVIDIA’s dominance, a development that could reshape supply chains and investment strategies across the tech sector.

Despite the hype, Gemini 3 exposed glaring reliability gaps. Users repeatedly asked the model for the current year, yet it stubbornly clung to a 2023‑2024 timeline, labeling fresh evidence as fabricated. This temporal hallucination highlighted a broader issue: LLMs can project unwarranted confidence even when confronted with clear contradictions. Meanwhile, Disney’s free‑roaming Olaf animatronic and a controversial AI‑infused teddy bear sparked public debate about the appropriateness of embedding sophisticated models in consumer toys, underscoring the thin line between novelty and safety.

The conversation turned darker with Figure AI’s robot‑butler lawsuit, where a former safety chief alleged that the company concealed a robot capable of delivering skull‑fracturing force. The suit accused leadership of ignoring written safety protocols and prioritizing speed over safeguards, echoing the “move fast and break things” mantra of earlier tech giants. As home robotics inch closer to everyday life, these incidents stress the urgent need for transparent standards, robust kill‑switch mechanisms, and rigorous third‑party audits to prevent accidents and maintain consumer trust.