AI Agents: When Research Becomes Engineering
Omar Sanseviero of Google DeepMind reveals how AI agents are turning research into automated engineering, freeing human experts to find entirely new discoveries.
⚡️ Google's Open AI Strategy — Omar Sanseviero, Google DeepMind
With swyx, Alessio Fanelli, Omar Sanseviero · Sunday, May 24, 2026
Omar Sanseviero from Google DeepMind discusses the release and capabilities of Gemma 4, including its novel E2B architecture for on-device inference, its multimodal and multilingual advancements, and its positioning relative to larger models like Gemini. The conversation also explores broader trends in AI development, such as the evolving role of fine-tuning, the challenges of MOE models, and the increasing convergence of research and engineering practices within Google and the wider AI community.
Gemma 4: On-Device AI That Speaks 140 Languages & Sees Video
Omar Sanseviero unveils Gemma 4's multimodal AI: 140 languages, on-device audio/video (30-60s), object detection. New product vectors just opened up.
Gemma Won't Eat Gemini: Google's On-Device AI Strategy
Omar Sanseviero of Google DeepMind clarifies Gemma 4's on-device role vs. Gemini's cloud dominance. Local AI shines for agents; big models own knowledge. Plan your AI strategy.
Gemma 4's E2B: Why Your On-Device AI Just Got Faster
Google DeepMind's Omar Sanseviero reveals Gemma 4's E2B architecture. It loads 2B 'effective' parameters on device for blazing fast AI on phones & PIs.
Google DeepMind: Your LLM Fine-Tuning Strategy Is Obsolete
Omar Sanseviero from Google DeepMind explains why general LLM fine-tuning is dead. Learn how prompting and agentic tools reshape customization for founders.
MOE Models: Fast Inference, Brutal Fine-Tuning Reality
Omar Sanseviero reveals Gemma's MOE 27B shines in speed but struggles with fine-tuning. Founders: choose dense for custom LLMs, MOE for raw inference.