Jun 24, 2025

Google Gemma and Open Source AI Models - Clement Farabet

What if your local AI could offload complex reasoning to a powerful cloud model? Learn Google's hybrid strategy for combining on-device Gemma with cloud-based Gemini.

#1about 6 minutes

Tracing a 20-year journey through AI's evolution

Clement Farabet recounts his career from early neural net experiments and building custom hardware to the inflection point of AlexNet and joining Google DeepMind.

#2about 2 minutes

Understanding the roles of Gemini and Gemma models

Gemini is a powerful, cloud-based API model for complex reasoning, while Gemma is an open model designed for efficient, customizable on-device deployment.

#3about 4 minutes

How Gemma models evolve and support customization

Gemma's capabilities are often distilled from more advanced Gemini models, and the community creates specialized variants like Med-Gemma for specific domains.

#4about 3 minutes

Understanding the open and permissive license of Gemma

Gemma's license allows for broad commercial use and modification, providing developers with the model weights and inference code without a subscription model.

#5about 6 minutes

Building web apps and live experiences with AI Studio

AI Studio enables developers to generate entire web applications from prompts, which Gemini 1.5 Pro can then debug and deploy to Cloud Run.

#6about 7 minutes

Envisioning the future of AI-driven contextual operating systems

On-device models like Gemma are paving the way for contextual operating systems where traditional apps and file systems are replaced by natural language interactions.

#7about 10 minutes

Defining AI agents and navigating the path to AGI

True AI agents possess autonomous action capabilities, which are best developed within constrained environments to solve complex tasks safely and avoid open-ended risks.

#8about 6 minutes

How developers can get started and contribute to AI

Developers should start with open models on platforms like Hugging Face and focus on creating novel user interfaces or fine-tuning models for specific domains.

#9about 2 minutes