Systems A/B/M : Lessons on Autonomous Learning from Cognitive Science

Systems A/B/M : Lessons on Autonomous Learning from Cognitive Science

by Jitendra Malik

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQMMachine learning
🎙 Jitendra Malik 👥 75K 📅 June 13, 2026 ⏱ 32 min 👁 673 📄 expert opinion 🧭 2026-08-03
Available in: English (current) Français

Keywords

System ASystem BSystem Mautonomous learningroboticsworld modelscognitive sciencereinforcement learningteleoperationsimulation

Summary

Jitendra Malik, in this talk at the Simons Institute, proposes a framework for autonomous learning in AI, drawing on cognitive science. He critiques current AI systems for lacking continual learning and relying on massive curated datasets. He introduces three systems: System A (observational learning, building world models), System B (interactive learning, acting and receiving rewards), and System M (meta-learning, discussed later). He argues that current robotics paradigms—teleoperation, learning from video, and simulation—each have limitations. Teleoperation suffers from human sensorimotor limitations and slow trajectories; video data has an embodiment gap and lacks force information; simulation faces the sim-to-real gap and the difficulty of specifying reward functions. He emphasizes the need to integrate multiple levels of abstraction (goals/plans, movement trajectories, forces) and suggests that a combination of Systems A and B, along with social learning, could lead to more autonomous and efficient learning. The talk is a high-level position statement rather than a detailed technical exposition.

154 words

Critical Evaluation

The talk offers a thought-provoking perspective on the limitations of current AI and robotics, proposing a cognitive-science-inspired framework. Malik’s critique of the dominant paradigms (teleoperation, video learning, simulation) is well-articulated and grounded in practical experience. The distinction between System A (observation) and System B (interaction) is clear and aligns with existing concepts in developmental psychology and reinforcement learning. However, the framework is presented as a ‘cartoon model’ and lacks operational details or empirical evidence. The discussion of System M is deferred, leaving the framework incomplete. The talk is more of an opinion piece than a rigorous scientific presentation, but it raises important questions about the future of autonomous learning. The sources cited are minimal (a position paper and the talk page), but the speaker’s authority and the institutional context lend credibility. The title accurately reflects the content, and the talk is accessible to a technical audience. Overall, it is a valuable contribution to the discourse on AI development, though not a definitive solution.

163 words

Title / Content Match

The title accurately reflects the content: the talk introduces Systems A/B/M as a framework for autonomous learning inspired by cognitive science, with a focus on robotics.

Quality & Reliability

8/10

Talk by a leading researcher (Jitendra Malik) at a prestigious institute (Simons Institute). Presents a speculative framework (Systems A/B/M) grounded in cognitive science and critiques current AI/robotics paradigms. Arguments are reasoned and reference a position paper, but the framework is explicitly cartoon-level and not empirically validated. No formal citations beyond the position paper and the talk page.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

  • Why AI Systems Don't Learn and What to Do About It (position paper) — Mentioned in the talk as a speculative position paper by Dupoux, LeCun, and Malik, which the framework is based on.

Contribution & Novelties

The talk proposes a novel framework (Systems A/B/M) that integrates observational and interactive learning, drawing on cognitive science. It critiques current AI and robotics paradigms and suggests a path toward more autonomous learning. The framework is speculative but offers a fresh perspective.

Pour aller plus loin :

  • World Models — Relevant to System A’s concept of building predictive models of the environment.
  • Reinforcement Learning — Relevant to System B’s interactive learning and reward-based optimization.
  • Cognitive Science — Provides the interdisciplinary background for the framework.
  • Sim-to-real transfer — Discusses the challenges and methods for transferring policies from simulation to real world.
  • Teleoperation — Relevant to the critique of current data collection methods.

111 words

Radar Profile

The radar profile shows high scores in quality of information and reliability, reflecting the speaker's expertise and the institutional context. The quantity of information is moderate, as the talk is a high-level overview rather than a detailed technical exposition. The technical level is high but accessible to a specialized audience.

Reliability 8/10