Google AI: Release Notes – Détails, épisodes et analyse

Tulsee Doshi, Head of Product for Gemini Models joins host Logan Kilpatrick for an in-depth discussion on the latest Gemini 2.5 Pro experimental launch. Gemini 2.5 is a well-rounded, multimodal thinking model, designed to tackle increasingly complex problems. From enhanced reasoning to advanced coding, Gemini 2.5 can create impressive web applications and agentic code applications. Learn about the process of building Gemini 2.5 Pro experimental, the improvements made across the stack, and what’s next for Gemini 2.5.

Chapters:

0:00 - Introduction
1:05 - Gemini 2.5 launch overview
3:19 - Academic evals vs. vibe checks
6:19 - The jump to 2.5
7:51 - Coordinating cross-stack improvements
11:48 - Role of pre/post-training vs. test-time compute
13:21 - Shipping Gemini 2.5
15:29 - Embedded safety process
17:28 - Multimodal reasoning with Gemini 2.5
18:55 - Benchmark deep dive
22:07 - What’s next for Gemini
24:49 - Dynamic thinking in Gemini 2.5
25:37 - The team effort behind the launch

Resources:

• Gemini → https://goo.gle/41Yf72b
• Gemini 2.5 blog post → https://goo.gle/441SHiV
• Example of Gemini’s 2.5 Pro’s game design skills →  https://goo.gle/43vxkq1
• Demo: Gemini 2.5 Pro Experimental in Google AI Studio → https://goo.gle/4c5RbhE

Dave Citron, Senior Director Product Management, joins host Logan Kilpatrick for an in-depth discussion on the latest Gemini updates and demos. Learn more about Canvas for collaborative content creation, enhanced Deep Research with Thinking Models and Audio Overview and a new personalization feature.

0:00 - Introduction
0:59 - Recent Gemini app launches
2:00 - Introducing Canvas
5:12 - Canvas in action
8:46 - More Canvas examples
12:02 - Enhanced capabilities with Thinking Models
15:12 - Deep Research in action
20:27 - The future of agentic experiences
22:12 Deep Research and Audio Overviews
24:11 - Personalization in Gemini app
27:50 - Personalization in action
29:58 - How personalization works: user data and privacy
32:30 -The future of personalization

Jack Rae, Principal Scientist at Google DeepMind, joins host Logan Kilpatrick for an in-depth discussion on the development of Google’s thinking models. Learn more about practical applications of thinking models, the impact of increased 'thinking time' on model performance and the key role of long context.

01:14 - Defining Thinking Models
03:40 - Use Cases for Thinking Models
07:52 - Thinking Time Improves Answers
09:57 - Rapid Thinking Progress
20:11 - Long Context Is Key
27:41 - Tools for Thinking Models
29:44 - Incorporating Developer Feedback
35:11 - The Strawberry Counting Problem
39:15 - Thinking Model Development Timeline
42:30 - Towards a GA Thinking Model
49:24 - Thinking Models Powering AI Agents
54:14 - The Future of AI Model Evals

Explore the synergy between long context models and Retrieval Augmented Generation (RAG) in this episode of Release Notes. Join Google DeepMind's Nikolay Savinov as he discusses the importance of large context windows, how they enable Al agents, and what's next in the field.

Chapters:
0:52 Introduction & defining tokens
5:27 Context window importance
9:53 RAG vs. Long Context
14:19 Scaling beyond 2 million tokens
18:41 Long context improvements since 1.5 Pro release
23:26 Difficulty of attending to the whole context
28:37 Evaluating long context: beyond needle-in-a-haystack
33:41 Integrating long context research
34:57 Reasoning and long outputs
40:54 Tips for using long context
48:51 The future of long context: near-perfect recall and cost reduction
54:42 The role of infrastructure
56:15 Long-context and agents

Learn more

• AI Studio: https://aistudio.google.com/
• Gemini Canvas: https://gemini.google.com/canvas
• Mariner: https://labs.google.com/mariner/
• Gemini Ultra: https://one.google.com/about/google-a...
• Jules: https://jules.google/
• Gemini Diffusion: https://deepmind.google/models/gemini...
• Flow: https://labs.google/flow/about
• Notebook LM: https://notebooklm.google.com/
• Stitch: https://stitch.withgoogle.com/

Chapters

• 0:59 - I/O Day 1 Recap
• 02:48 - Envisioning I/O 2030
• 08:11 - AI for Scientific Breakthroughs
• 09:20 - Veo 3 & Flow
• 7:35 - Gemini Live & the Future of Proactive Assistants
• 20:30 - Gemini in Chrome & Future Apps
• 22:28 - New Gemini Models: DeepThink, Diffusion & 2.5 Flash/Pro Updates
• 27:19 - Developer Momentum & Feedback Loop
• 31:50 - New Developer Products: Jules, Stitch & CodeGen in AI Studio
• 37:44 - Evolving Product Development Process with AI
• 39:23 - Closing

Connie Fan, Product Lead for Gemini's coding capabilities, and Danny Tarlow, Research Lead for Gemini's coding capabilities, join host Logan Kilpatrick for an in-depth discussion on how the team built one of the world's leading AI coding models. Learn more about the early goals that shaped Gemini's approach to code, the rise of 'vibe coding' and its impact on development, strategies for tackling large codebases with long context and agents, and the future of programming languages in the age of AI.

Watch on YouTube: ⁠https://www.youtube.com/watch?v=jwbG_m-X-gE⁠

Chapters:

0:00 - Intro
1:10 - Defining Early Coding Goals
6:23 - Ingredients of a Great Coding Model
9:28 - Adapting to Developer Workflows
11:40 - The Rise of Vibe Coding
14:43 - Code as a Reasoning Tool
17:20 - Code as a Universal Solver
20:47 - Evaluating Coding Models
24:30 - Leveraging Internal Googler Feedback
26:52 - Winning Over AI Skeptics
28:04 - Performance Across Programming Languages
33:05 - The Future of Programming Languages
36:16 - Strategies for Large Codebases
41:06 - Hill Climbing New Benchmarks
42:46 - Short-Term Improvements
44:42 - Model Style and Taste
47:43 - 2.5 Pro’s Breakthrough
51:06 - Early AI Coding Experiences
56:19 - Specialist vs. Generalist Models

⁠

⁠

A conversation with Sergey Brin, co-founder of Google and computer scientist working on Gemini, in reaction to a year of progress with Gemini.

Watch on YouTube: https://www.youtube.com/watch?v=o7U4DV9Fkc0

Chapters

0:20 - Initial reactions to I/O
2:00 - Focus on Gemini’s core text model
4:29 - Native audio in Gemini and Veo 3
8:34 - Insights from model training runs
10:07 - Surprises in current AI developments vs. past expectations
14:20 - Evolution of model training
16:40 - The future of reasoning and Deep Think
20:19 - Google’s startup culture and accelerating AI innovation
24:51 - Closing

Ani Baddepudi, Gemini Model Behavior Product Lead, joins host Logan Kilpatrick for a deep dive into Gemini's multimodal capabilities. Their conversation explores why Gemini was built as a natively multimodal model from day one, the future of proactive AI assistants, and how we are moving towards a world where "everything is vision." Learn about the differences between video and image understanding and token representations, higher FPS video sampling, and more.

Chapters:

0:00 - Intro
1:12 - Why Gemini is natively multimodal
2:23 - The technology behind multimodal models
5:15 - Video understanding with Gemini 2.5
9:25 - Deciding what to build next
13:23 - Building new product experiences with multimodal AI
17:15 - The vision for proactive assistants
24:13 - Improving video usability with variable FPS and frame tokenization
27:35 - What’s next for Gemini’s multimodal development
31:47 - Deep dive on Gemini’s document understanding capabilities
37:56 - The teamwork and collaboration behind Gemini
40:56 - What’s next with model behavior

Watch on YouTube: https://www.youtube.com/watch?v=K4vXvaRV0dw