Google Gemini: Open Source and Deep Thinking Models - Sam Witteveen

What if your AI model could think before it answered? Learn how deep reasoning is transforming AI from a simple tool into a true coding partner.

#1about 3 minutes

Navigating the current AI hype and research secrecy

The term "AI" is often used for marketing, while the competitive landscape has made research labs more secretive about their work.

#2about 3 minutes

Understanding Google's open weights Gemma models

Gemma models are "open weights," not fully open source, meaning developers can use the model weights for commercial projects but don't have the training data or code.

#3about 2 minutes

The training process of large language models

LLMs are trained in stages, starting with pre-training to predict the next token, followed by post-training and reinforcement learning to align with human instructions.

#4about 2 minutes

Comparing proprietary Gemini and open Gemma models

Gemini models are Google's proprietary, cloud-based offerings, while Gemma models are open-weight versions that can be run on-premise or on-device.

#5about 3 minutes

How AI models are becoming smaller and more efficient

Models are made smaller and more powerful through training on vast amounts of data and using techniques like distillation to transfer knowledge from larger models.

#6about 7 minutes

The rise of multilingual and multimodal AI

Modern models like Gemma 3 are trained on over 140 languages, and multimodal models like Gemini can process text, audio, video, and images simultaneously.

#7about 4 minutes

Improving text in AI-generated images and videos

Recent models are finally getting better at rendering accurate text in images, while video generation models can create complex, realistic scenes from prompts.

#8about 3 minutes

AI tools for advanced content creation and editing

AI-powered tools can now edit video by manipulating the transcript and even insert new words in a cloned voice, simplifying the post-production process.

#9about 4 minutes

Using large language models as a learning tool

LLMs serve as powerful educational tools by explaining complex concepts at different levels, acting as a personal tutor for learning new technical skills.

#10about 6 minutes

How deep reasoning models 'think' before answering

Reasoning models improve accuracy by generating an internal monologue to break down a problem and explore possibilities before providing a final answer.

#11about 10 minutes

The evolution of AI-powered coding assistants

AI coding tools are evolving into agents that can write code, run tests, and fix errors, shifting the developer's role toward architectural and conversational guidance.

#12about 2 minutes

How to get started with Google's Gemma models

Developers can start experimenting with Gemma models through Google's AI Studio, Hugging Face, or by running them locally with tools like Ollama and LM Studio.