Spatial World Models for AGI: Approaches, Funding, and Provenance

The Shift Towards Spatial World Models for AGI

• 🚀 The field of AI is on the cusp of evolution, moving from current transformer-based models towards Artificial General Intelligence (AGI).
• 💡 Spatial world models are emerging as the dominant theme, recognized as crucial for achieving true AGI.
• ⚠️ Current Large Language Models (LLMs) are acknowledged as insufficient for reaching AGI, necessitating new foundational approaches.

Diverse Approaches to Spatial World Models

• 🤖 Google's Genie 3 represents a pure autoregressive, transformer-based approach, excelling in engineering but facing challenges with long-term object persistence.
• 🌍 Fei Fei Li's World Labs Marble offers generative 3D spatial modeling, specifically for static objects, characterized by persistent object memory and dynamic camera views.
• 🧠 Yann LeCun's AMI Labs is expected to deliver products based on his VL-JEPA (Vision Language Joint Embedding Predictive Architecture) framework.
• 🌌 Verses.ai, featuring Karl Friston, utilizes generative active inference for dynamic objects, rooted in variational inference.
• 💡 A new entrant, Him Ditt's General Intuition, further solidifies the trend towards spatial world modeling startups.

Significant Investment in Spatial AI

• 💰 Major tech giants like Google DeepMind and Meta FAIR are investing billions annually into AI research, with a significant portion directed towards spatial world models.
• 📈 Four key startups (AMI Labs, World Labs, Verses.ai, General Intuition) have collectively secured over $1 billion in initial investment rounds, signaling a major industry shift.
• 💸 Verses.ai is noted as significantly underfunded compared to its peers, despite its foundational active inference approach, highlighting a potential disparity in market valuation versus conceptual utility.

Provenance and the Genius Behind the Models

• 🌟 The development of these spatial world models stems from decades of work by creative geniuses, not sudden new formulations.
• 🖼️ Fei Fei Li's long-term fascination with imagery led to the creation of ImageNet, a pivotal dataset for computer vision, and now her focus on spatial intelligence.
• 🌐 Yann LeCun is recognized as the original inventor of Convolutional Neural Networks (CNNs), evolving his work into JEPA and VL-JEPA for his new venture.
• 🧠 Karl Friston developed Active Inference, a dynamic form of variational inference based on the free energy principle, through over a decade of dedicated research.

Future Implications for AGI Development

• ✅ Achieving AGI will require robust solutions for knowledge representation (including spatial, temporal, and abstract concepts), reasoning, and values.
• 🔍 A critical step for evaluating future AGI systems is a deep understanding of generative versus non-generative AI and effective object representation.
• 📚 Individuals and organizations are encouraged to engage in self-study or structured courses to grasp these fundamental concepts for informed decision-making in the evolving AI landscape.