[Meta FAIR Research: Seeing, Understanding, and Collaborating AI]

Study with Donghyung Shin: Meta FAIR Research Report "Seeing, Understanding, and Collaborating AI" (April 18, 2025). Section 1 AI New Leap: Meaning of Meta FAIR Research. 1.1 Journey Toward Advanced Machine Intelligence (AMI): Meta Fundamental AI Research (FAIR) team is conducting research on systems that understand and navigate the complex real world, going beyond AI that simply processes text. FAIR research goal: developing Advanced Machine Intelligence (AMI) that perceives, reasons, and acts comprehensively rather than narrow task-specific AI. Three research pillars: (1) Perception AI -- AI that understands visual and audio information from the physical world; (2) Reasoning AI -- AI that builds models of how the world works and reasons about causality; (3) Collaborative AI -- AI that works alongside humans as a genuine partner rather than a tool. Section 2 Seeing AI (Perception): video understanding (understanding what is happening in video, not just recognizing objects in images); 3D scene reconstruction from 2D images; audio-visual correspondence (understanding what sounds correspond to which visual events); the FAIR perception research is directly applied to Meta products (Instagram content understanding, AR glasses visual processing, video ranking). Section 3 Understanding AI (World Models): JEPA (Joint Embedding Predictive Architecture) -- Meta approach to world models that learns by predicting missing parts of a scene rather than generating pixel-level details; the insight is that good world model representations are abstract (capturing meaning) not concrete (capturing pixels); I-JEPA and V-JEPA demonstrated that self-supervised learning with JEPA objectives produces better representations than generation-based objectives. Section 4 Collaborative AI: multi-agent systems where AI agents collaborate with each other and with humans; the research challenge is enabling agents to develop shared representations and communication protocols; practical applications include AI assistants that can hand off tasks between specialized agents; Section 5 Implications: FAIR research positions Meta as doing fundamental AI research beyond product AI -- the AMI vision is a long-term bet that general AI capability requires advances in perception, reasoning, and collaboration working together rather than in isolation; open-sourcing research (LLaMA, Segment Anything, DINOv2) accelerates the field while building Meta reputation as a genuine research institution rather than just a product company.