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

Key takeaways

• Research now often means "engineering with unknowns": Omar Sanseviero from Google DeepMind notes that many researchers spend their days on "ablations," which means moving pieces around to see what works. He believes this is "much more engineering rather than for like research," unless you are designing new fundamental architectures. Read more →
• Gemma 4 brings multimodal AI to your pocket: Google DeepMind's Omar Sanseviero confirmed that Gemma 4's smaller models can now process audio, images, and short videos (30-60 seconds) right on a device, a leap for edge computing previously reserved for larger, cloud-based AI. Read more →
• Google's Gemma 4, an on-device model, already matches the state-of-the-art capabilities from 1 to 1.5 years ago for local functions like agentic tasks and conversational AI. Read more →
• Google DeepMind’s Gemma 4 model uses a novel E2B architecture that loads only 2 billion “effective” parameters onto the GPU for inference, even though the model itself contains nearly 5 billion parameters. Read more →
• General fine-tuning for broad conversational changes in large language models is rapidly becoming unnecessary, with models like Google DeepMind's Gemma 4 performing exceptionally well out-of-the-box. Read more →
• Google DeepMind architected Gemma with two distinct models: a 31B dense model for maximum "raw intelligence," and a 27B Mixture-of-Experts (MOE) model optimized for "extremely fast inference" on consumer GPUs. Read more →