Jitendra Malik

AI researcher specializing in computer vision and robotics with a focus on sensorimotor learning.

Mentions

Appeared in

Discussed in

Key positions and views

Sensorimotor Skills Precede Language: True, robust AI must first master physical interaction, sensory perception, and motor control before developing advanced language, mirroring the developmental path of humans.

Quadrupedal Locomotion is Nearly Solved: Through techniques like Rapid Motor Adaptation (RMA), the challenge of enabling four-legged robots to walk and adapt to varied, uneven terrain is largely resolved.

Bipedal Locomotion Remains a Grand Challenge: Unlike its four-legged counterpart, two-legged locomotion is significantly harder due to the complexities of balance and is still an unsolved problem in robotics.

LLMs are Powerful but Ungrounded: While models like GPT-4 demonstrate impressive text-processing capabilities, they lack the grounding in physical, sensory-motor experience required for true, non-brittle intelligence.

AI Research is Bifurcating: The field is splitting into corporate-led 'big science' for training massive models and university-led 'small science' for conceptual exploration, making it impossible for academia to compete on model scale.

Podcast consensus on Malik

Points of consensus

▶Jitendra Malik consistently argues that true artificial intelligence must be built on a foundation of sensorimotor learning, where physical interaction precedes language capabilities, mirroring human evolution. This view is central to his entire research philosophy.May 2026

▶He and his collaborators successfully developed Rapid Motor Adaptation (RMA), a technique that enables legged robots to adapt to diverse and challenging terrains in near real-time by using the discrepancy between commanded and sensed actions.May 2026

▶Malik provides a clear distinction between solved and unsolved problems in AI, stating that quadrupedal locomotion is 'pretty close to solved,' while bipedal locomotion and human-level 3D reconstruction from a single image remain significant open challenges.May 2026

▶He believes that while Large Language Models are powerful, their learning process is analogous to reading books and they lack the 'grounded' experience from physical interaction, which is necessary to make them more robust and less brittle.May 2026

Points of debate

▶Malik highlights a growing divide in AI research between 'big science' (resource-intensive projects like training GPT-4, only feasible at large companies) and 'small science' (novel exploration suitable for university labs), creating a tension in the ecosystem of innovation.May 2026

▶He presents a contrast between the capabilities of current AI, noting that while a system like GPT-4 can score in the 90th percentile on a law exam, his own advanced robots could not climb stairs without the addition of vision, underscoring the different axes of AI progress.May 2026

▶There is a tension in his view of LLMs: he analogizes their learning to how humans read books, a powerful form of knowledge acquisition, yet simultaneously argues this method is fundamentally incomplete without sensory-motor grounding.May 2026

▶He contrasts the near-solved problem of four-legged (quadrupedal) locomotion with the much harder, unsolved problem of two-legged (bipedal) locomotion, highlighting a key frontier in robotics research.May 2026

Key themes

▶Embodiment First: A Developmental Approach to AIMay 2026

Malik's core thesis is that AI development should mirror human evolution, where physical intelligence (e.g., dexterity from an opposable thumb) preceded complex language. He bets his career on the idea that grounding AI in sensory-motor competencies like seeing, moving, and manipulating objects is the necessary foundation for creating robust, less brittle intelligence.

This 'embodiment-first' strategy represents a contrarian bet against the dominant language-first paradigm, suggesting that long-term AGI investors should consider robotics and sensorimotor research as critical, potentially undervalued, enabling technologies.

▶Rapid Motor Adaptation (RMA) as a Robotics BreakthroughMay 2026

Malik details the development of RMA, a technique enabling a robot to adapt its movements to changing terrain in under a second. The system's core principle is to use the discrepancy between a commanded action and the actual sensed motion to infer terrain properties and adjust its policy, allowing a single neural network to navigate everything from rocky riverbeds to muddy piles.

RMA's success in creating a single, adaptable policy for diverse terrains is a significant step toward deploying autonomous robots in unpredictable, real-world environments, moving beyond the structured confines of factories.

▶Mapping the Frontier: Solved vs. Unsolved Problems in AIMay 2026

Malik provides a clear assessment of the current state of AI capabilities, declaring quadrupedal locomotion 'pretty close to solved.' Conversely, he identifies bipedal locomotion and recovering a human-level 3D model from a single image as major unsolved challenges, illustrating the uneven progress across different domains of AI research.

Malik's delineation of solved and unsolved problems offers a valuable roadmap for R&D, signaling to analysts that bipedalism and advanced single-image 3D vision are high-impact areas ripe for future breakthroughs and investment.

▶The Bifurcation of AI Research: 'Big Science' vs. 'Small Science'May 2026

Malik characterizes the AI research landscape as splitting into two camps: the 'small science' of individual-led exploration common in universities, and the 'big science' of massive, resource-heavy projects like training GPT-4 or Meta's Segment Anything, which are now the exclusive domain of large corporations. He explicitly states that training a GPT-scale model is no longer feasible in an academic lab.

This trend suggests a consolidation of power in foundational model development within a few corporate entities, potentially creating a bottleneck for academic innovation and shifting the role of universities toward theoretical work and niche applications.

Source episodes

Sentiment over time

Not enough data for timeline

Changes over time

Foundational Belief

Malik grounds his research philosophy in evolutionary biology, citing fossil data showing that the human hand's evolution preceded the brain's, forming the basis for his 'embodiment-first' approach to AI.

Core Research Strategy

He bets his scientific career on the approach of developing physical, sensorimotor intelligence in AI before focusing on language, believing this is the path to fully realized AI.

Key Technical Development

Malik and his collaborators develop Rapid Motor Adaptation (RMA), a technique for robot locomotion that uses the gap between commanded and sensed motion to adapt to new terrain in less than a second.

Empirical Validation

The RMA policy is proven effective, allowing a robot dog to traverse diverse terrains. During this work, his group discovers that vision is a necessary component for more complex tasks like climbing stairs, which the blind version could not handle.

State of the Field Assessment

He assesses the current state of robotics and computer vision, declaring quadrupedal locomotion nearly solved but identifying bipedal locomotion and single-image 3D reconstruction as major remaining challenges.

Broader AI Commentary

Malik comments on the broader AI landscape, acknowledging the power of LLMs like GPT-4 while critiquing their lack of grounded experience and noting the structural split in AI research between corporate 'big science' and academic 'small science'.

Suggested prompts

How does Jitendra Malik's 'embodiment-first' philosophy challenge the current trajectory and scaling laws of large language model development? &nearr;What are the specific technical hurdles that make bipedal locomotion significantly harder than quadrupedal locomotion according to Malik's framework? &nearr;If Malik's theory that grounded experience is necessary for robust AI is correct, what are the investment implications for robotics and hardware companies versus pure software AI firms? &nearr;How does the 'big science' vs. 'small science' dynamic described by Malik impact the potential for disruptive innovation from academic institutions in the age of foundation models? &nearr;

Key concepts

Sensorimotor Learning 1 ep Rapid Motor Adaptation (RMA) 1 ep Quadrupedal Locomotion 1 ep Bipedal Locomotion 1 ep Computer Vision 1 ep Grounded Learning 1 ep 3D Reconstruction 1 ep Neural Radiance Fields (NERF) 1 ep Large Language Models (LLMs) 1 ep Big Science vs. Small Science 1 ep

Notable quotes

“I think quadrupedal locomotion is pretty close to salt, I would say.”

Jitendra Malik · S3 E20 Jitendra Malik on Building AI from the ground up: Sensorimotor learning before language

“And this discrepancy is the signal that we can use to adapt.”

Jitendra Malik · S3 E20 Jitendra Malik on Building AI from the ground up: Sensorimotor learning before language

“Bipedal locomotion is harder. So when you have two feet versus four feet, four feet are easier because balance is easier... it doesn't seem like that problem is solved.”

Jitendra Malik · S3 E20 Jitendra Malik on Building AI from the ground up: Sensorimotor learning before language

“Climbing upstairs, it couldn't do. So for that, we found it necessary to introduce the use of vision.”

Jitendra Malik · S3 E20 Jitendra Malik on Building AI from the ground up: Sensorimotor learning before language

Report last updated: May 5, 2026

Get started free

Back to Entities Intelligence Report