TickerTain
TickerTain
NewsroomShortsPortfolioConvergence
NewsroomShortsPortfolioConvergence
←
▶ 5:54 · Robotics & Physical AI · General robotic foundation models will unlock Cambrian explosion of robot applications
episode briefing
Invest Like The Best

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

2026-03-31 · 4 company · 9 thematic
sentiment
3 bull0 bear1 neu
speakers
patrick o'shaughnessy

Founder of the investment firm Positive Sum and host of Invest Like the Best. He also created the Colossus network of business and investing podcasts.

sergey levine

UC Berkeley computer science professor specializing in machine learning and robotics. He is a co-founder of Physical Intelligence, which develops general-purpose AI for robots.

now playing · Robotics & Physical AI
Robotics & Physical AItailwindscore 9/10sergey levine
General robotic foundation models will unlock Cambrian explosion of robot applications
Just as PCs and LLMs enabled anyone to build software applications, a general-purpose robotic foundation model will let anyone build robot applications without solving the intelligence stac…
Robotics & Physical AItailwindscore 8/10sergey levine
One intelligence problem across all robot form factors
Physical intelligence is embodiment-agnostic: the same foundation model can control humanoids, arms, bulldozers, or drone swarms. The fundamentals of physical interaction, causality, and ob…
Frontier AI Modelstailwindscore 9/10sergey levine
Multimodal LLMs provide common sense for robotics long-tail scenarios
Multimodal LLMs contain world knowledge that can be grounded in physical situations via chain-of-thought reasoning, solving the 'common sense' bottleneck for handling novel scenarios. This…
Frontier AI Modelstailwindscore 8/10sergey levine
Combining generative AI knowledge with RL superhuman performance is the grand challenge
Generative AI (LLMs) captures human knowledge but mimics human performance; deep RL (AlphaGo) discovers superhuman strategies but lacks world knowledge. Robotics needs both: web-scale knowl…
AI Economics & Business Modelstailwindscore 9/10sergey levine
Robotics data flywheel starts at 'useful enough' not 'perfect'
The key to solving robotics data scarcity is not pre-collecting massive datasets but deploying systems that are useful enough to operate in the real world, where they autonomously gather di…
Robotics & Physical AImixedscore 8/10sergey levine
ML flips Moravec's paradox: data-rich physical tasks become easy
Tasks easy for humans but hard for robots (dexterity, manipulation) become tractable with ML when data collection is straightforward. The remaining hard problems are where data is scarce an…
Robotics & Physical AItailwindscore 8/10sergey levine
Robot improvement bottleneck shifted to semantic reasoning — solvable by language coaching
Robots now fail at task interpretation, not low-level control. Adding semantic labels (coaching) to autonomous experience improves generalization without new teleoperation data. This means…
Robotics & Physical AImixedscore 8/10sergey levine
Humanoid locomotion uses simulation; manipulation uses real data — winner unclear
Two divergent paradigms exist: humanoid acrobatics rely on heavy simulation with near-zero real data, while manipulation relies on massive real-world datasets with little simulation. The fi…
Robotics & Physical AItailwindscore 7/10sergey levine
Robot hardware costs plummeted 100x enabling general-purpose research
Robot arm costs dropped from $400k (PR2) to ~$3k per arm due to learning-based control that works with low-precision hardware. This hardware affordability constellation makes general-purpos…