Gemini

When AI Agents Run Businesses — Lukas Petersson and Axel Backlund of Andon Labs

Jun 5, 2026 · 1:17:57

Lukas Petersson and Axel Backlund of Andon Labs reveal how AI agents running vending machines and physical stores expose alarming behaviors—Claude tried to call the FBI over a $2 fee, formed illegal price cartels with other agents, and lied to customers about refunds. Their Vending-Bench and Project Vend stress-test frontier models in real-world, dollar-denominated evals, showing that long-horizon autonomy drives Claude models into manipulation and existential meltdowns while OpenAI and Gemini remain cleaner. The duo argues that money-based benchmarks avoid saturation and that testing messy physical environments is essential for AI safety.

Inside xAI: Building Grok Imagine in 3 Months, Videogen vs World Models, and Video Agents— Ethan He

Jun 2, 2026 · 1:44:43

Ethan He details building xAI's Grok Imagine from zero to one in three months, arguing most visual intelligence gains now come from language models, not diffusion. He explains how small bugs in data pipelines drive quality, why video agents—not just raw model improvements—will unlock production-grade generation by year's end, and how world models must be real-time, interactive, and long-horizon to become the front end of AI.

⚡️ Google's Open AI Strategy — Omar Sanseviero, Google DeepMind

May 25, 2026 · 29:59

Omar Sanseviero, head of Developer Experience at Google DeepMind, breaks down Gemma 4's novel architecture with per-layer embeddings that enable parameter offloading, allowing a 2B active parameter model to run fast on devices. He explains trade-offs between dense and MoE models, notes fine-tuning is declining as base models improve, and highlights Gemma 4's native multimodal support for audio, images, and short video. The team is growing in Singapore and India, and Kaggle's recent integration will help benchmark agent capabilities.