Greylock · Video summary
Building AI for the Physical World with Generalist
Pete Florence, co-founder and CEO of Generalist, discusses the company's journey in building general intelligence for the physical world, highlighting their latest model, Gen 1.5, and its emergent capabilities in robotics.
Watch on YouTube ↗Key takeaways
- 🍌A robot using a banana as a brush demonstrates emergent physical generalization, adapting to novel tools for a known task.00:25
- 📈The rapid progress in robotics AI was not guaranteed; it required significant advancements in training recipes and scaling efficiencies.01:08
- 🌍Generalist's long-term goal is to build general intelligence for the physical world, automating undesirable tasks and accelerating scientific discovery.04:17
- 🧠The emergence of one-shot and few-shot learning, coupled with generalization, is a key milestone in robotics, similar to the impact of GPT-3 in language models.06:50
- 📏Gen Zero established strong scaling laws in robotics, providing a predictable ability to improve models with more compute, data, and model size.08:38
- 🏆Gen One demonstrated mastery across several tasks from a general model, achieving high success rates and improving speed and improvisational intelligence.10:22
Chapters
- 0:00Introduction to Generalist's AI for the Physical World
- 0:42The Inception Story and Technical Risks
- 1:48Defining Generalist's Long-Term Goal
- 5:01Parallels Between Language and Robotics Models
- 8:14The Evolution of Generalist's Models: Gen Zero, Gen One, and Gen 1.5
- 18:10The Concept of Cross-Embodiment
- 23:10Research Philosophy and Working with Customers
- 28:37The Future of Generalist and Robotics AI
The digest
Generalist is on a mission to build general intelligence for the physical world, aiming to revolutionize how robots interact with their environment. Pete Florence, co-founder and CEO, sheds light on the company's ambitious goals, the technical challenges they've overcome, and the exciting advancements in their robotics models.
The Unforeseen Progress in Robotics
Just a few years ago, the idea of robust, general-purpose AI in robotics seemed distant. Florence recalls his graduate school days, where models were trained on mere hours of data and had tens of thousands of parameters. The progress since then has been exponential, moving from thousands of hours of data at Google to the current scale where models are the result of tens of thousands of training runs and experiments. This rapid advancement wasn't a given; it required figuring out the right training recipes and scaling efficiencies that fundamentally shifted the capabilities of these models.
Keep reading in The Digest
The full 2-minute article, 3 more sections, and every new episode from Greylock and the creators you follow.
- The Core Drivers of Progress
- Generalist's Long-Term Vision
- Parallels with Language Models
