Invest Like the Best• Mar 31, 2026• 1:15:15Interview

World's Top Researcher on AI, LLMs, and Robot Intelligence

From Invest Like the Best

Sergey Levine•Co-founder and Researcher, Physical Intelligence

Executive Summary

Physical Intelligence is developing general-purpose, body-agnostic foundation models for robotics, aiming to create a universal 'brain' that can power any robot for any task.
The company's core thesis is that a generalist approach will be more effective long-term than building specialized, single-task robots, mirroring the success of large language models over narrow NLP solutions.
A primary challenge in robotics is the lack of an 'internet-sized' dataset.
The key strategy is to deploy useful robots into the real world to collect diverse, real-world data at scale.
The ultimate goal is to create a platform that dramatically lowers the barrier to entry for robotics innovation, enabling a 'Cambrian explosion' of new robotic forms and applications.

9 quotes

Concerns Raised

The lack of a large, pre-existing dataset for training robotic models remains a fundamental bottleneck.
The value of generalization is difficult to demonstrate, as highly-controlled, single-task demos often appear more impressive to the public.
There is an unresolved dichotomy between simulation-heavy approaches (common for locomotion) and real-world data-heavy approaches (for manipulation).

Opportunities Identified

Creating a general-purpose foundation model could unlock a massive wave of innovation in robotic hardware and applications.
Leveraging multimodal language models can imbue robots with common sense and high-level reasoning capabilities.
The dramatic decrease in hardware costs makes widespread experimentation and data collection more feasible.
Robots have the potential to surpass human capabilities in speed, precision, and efficiency for many physical tasks.

Key Themes

Generalist vs. Specialist Robotics

The conversation centers on the thesis that developing a single, general-purpose robotic foundation model is ultimately more viable and powerful than creating specialized models for narrow tasks like washing dishes. This approach leverages data from diverse sources to build a foundational understanding of physical interaction, which can then be applied to new tasks and robots more efficiently.

This strategic bet mirrors the evolution of AI, where general language models like GPT surpassed narrow, task-specific models. If successful, it could define the dominant paradigm for the next generation of robotics.

The Data Scarcity Problem

Unlike LLMs which could be trained on the vast text of the internet, robotics lacks a pre-existing, large-scale dataset of physical interaction. The discussion highlights the critical need for effective data acquisition strategies, such as deploying robots in the real world, and the ongoing debate between using real-world data versus simulation.

Data is the key bottleneck for progress in robotic learning. The winning strategies for data collection and utilization will likely determine which companies lead the field.

Foundation Models for the Physical World

The core idea is to apply the foundation model paradigm from AI to the physical world. This involves creating a base intelligence that understands physics, cause-and-effect, and manipulation, which can then be prompted or fine-tuned for specific applications, regardless of the robot's physical form (embodiment).

This approach abstracts the 'intelligence' layer from the 'hardware' layer, potentially commoditizing robotic intelligence and allowing for rapid innovation in physical robot design and application.

Democratizing Robotics Innovation

A successful general-purpose model would act as a platform, significantly lowering the barrier to entry for creating new robotic applications. Entrepreneurs and engineers could build novel hardware and software solutions on top of this foundational intelligence, without needing to solve the core AI problem themselves.

This could trigger a 'Cambrian explosion' of creativity in robotics, similar to how the personal computer and the internet unleashed waves of software innovation, leading to unforeseen applications and industries.

Embodiment and Form Factor

The discussion explores the idea that intelligence should be 'body-agnostic.' While humanoid robots capture the public imagination, the future will likely consist of a wide variety of robot forms optimized for different tasks. A general foundation model would be able to control this diverse ecosystem of hardware.

Focusing on a general intelligence model rather than a specific form factor allows for greater flexibility and prevents being locked into a potentially suboptimal hardware design, fostering broader innovation.

Get started free

Topics

Robotics Foundation Models Physical Intelligence Artificial Intelligence Machine Learning Generalization Embodied AI Data Acquisition Simulation vs. Real-World Data Humanoid Robots Robot Form Factor Multimodal Models Reinforcement Learning Common Sense Reasoning Sergey Levine

Processed Apr 3, 2026 yt-dlp + mlx-whisper + Gemini