Subvolts take Understand Gemini's two-step AI training and new agentic features from a Google DeepMind leader. Dec 12, 2024 - Developers
- Learn about the training methodologies and architectural innovations behind advanced AI models like Gemini.
- Business
- Understand the progression of AI capabilities and how general-purpose agents might impact future applications.
This episode features a key leader discussing the technical evolution and new capabilities of a major AI model.
Gemini 2.0 and the Evolution of Agentic AI with Oriol Vinyals
Audio streams from Hannah Fry. Your place is remembered on this device.
From Subvolts
Summary and analysis by Subvolts editorial, AI-assisted.
Google DeepMind's podcast features Oriol Vinyals, Gemini co-lead, discussing the evolution of AI agents from single-task to general-purpose models like Gemini. Vinyals explains the two-stage training process: pre-training via imitation learning and post-training via reinforcement learning. The episode also covers scaling complexities, architectural innovation, and new agentic capabilities released by Google DeepMind.
This episode explains how AI like Gemini learns, first by copying and then by practicing to get better. It covers how these AI systems are becoming smarter and can do more things, which could lead to helpful new tools for everyone.
This episode offers insight into how advanced AI models like Gemini are built, moving from simple imitation to complex learning. It's for anyone interested in the technical underpinnings of modern AI and how labs are pushing towards more capable, general AI systems. The discussion on two-stage training and architectural innovation directly relates to current efforts across frontier labs to create more powerful and versatile AI.
- HostHannah Fry
- GuestOriol VinyalsVP of Drastic Research and Gemini co-lead, Google DeepMind
- Episode Title: Gemini 2.0 and the Evolution of Agentic AI with Oriol Vinyals
- Show: Google DeepMind: The Podcast
- Length: 49 minutes
- Guest: Oriol Vinyals, VP of Drastic Research and Gemini co-lead
- Training Process: Two-step: pre-training (imitation learning) and post-training (reinforcement learning)
In this episode, Hannah is joined by Oriol Vinyals, VP of Drastic Research and Gemini co-lead. They discuss the evolution of agents from single-task models to more general-purpose models capable of broader applications, like Gemini. Vinyals guides Hannah through the two-step process behind multi modal models: pre-training (imitation learning) and post-training (reinforcement learning). They discuss the complexities of scaling and the importance of innovation in architecture and training processes. They close on a quick whirlwind tour of some of the new agentic capabilities recently released by Google DeepMind.
Note: To see all of the full length demos, including unedited versions, and other videos related to Gemini 2.0 head to YouTube.
Future reading/watching:
- Gemini 2.0
- Decoding Google Gemini with Jeff Dean
- Gaming, Goats & General Intelligence with Frederic Besse
Thanks to everyone who made this possible, including but not limited to:
Presenter: Professor Hannah Fry
Series Producer: Dan Hardoon
Editor: Rami Tzabar, TellTale Studios
Commissioner & Producer: Emma Yousif
Music composition: Eleni Shaw
Camera Director and Video Editor: Bernardo Resende
Audio Engineer: Perry Rogantin
Video Studio Production: Nicholas Duke
Video Editor: Bilal Merhi
Video Production Design: James Barton
Visual Identity and Design: Eleanor Tomlinson
Commissioned by Google DeepMind
—
Subscribe to our YouTube channel
Find us on X
Follow us on Instagram
Add us on Linkedin
Please leave us a review on Spotify or Apple Podcasts if you enjoyed this episode. We always want to hear from our audience whether that's in the form of feedback, new idea or a guest recommendation!
Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
- Agentic AI
- Oriol Vinyals
- AI that can act and make decisions on its own to achieve goals.
- Multimodal models
- Oriol Vinyals
- AI that can understand and process information from different types of data, like text, images, and sound.
- Imitation learning
- Oriol Vinyals
- Teaching AI by showing it examples of how to perform a task.
- Reinforcement learning
- Oriol Vinyals
- Teaching AI by letting it try things and rewarding it for good outcomes.
44 minCan AI help us better predict the weather?
WithPeter Battaglia
45 minThe mathematics of AI uncertainty
WithZoubin Ghahramani
53 minSame guests, other shows
42 minGemini co-leads on project origins and what's next
WithJeff Dean, Koray Kavukcuoglu, Noam Shazeer, Oriol Vinyals
1 h27 minLife is like a game
WithOriol Vinyals, Max Jaderberg, Raia Hadsell, David Silver
35 minBehind the Scenes of Gemini 2.0
WithTulsee Doshi
43 minSmaller, Faster, Cheaper & The Story of Flash 8B
WithEmanuel Taropa
Questions people ask
- What is the two-step process for training multimodal models?
- The process involves pre-training using imitation learning and post-training using reinforcement learning.
- Where can I see full-length demos of Gemini 2.0?
- Full-length and unedited demos of Gemini 2.0 are available on YouTube.
Page generated Oct 19, 2026. Summaries are Subvolts' own; the audio, artwork and notes belong to Hannah Fry.