Object-Centered AI Models Grounded in Physics for True Understanding
TL;DR
- AI development has prioritized scaling large models over developing architectures that mimic biological intelligence, potentially leading to systems that excel at pattern matching but lack true understanding or reasoning capabilities.
- The current AI paradigm, heavily reliant on transformers and large-scale data, may be fundamentally flawed by grounding models in language rather than physics, hindering the development of AI that truly understands the world.
- A shift towards object-centered, modular AI models, akin to video game engines, could enable more efficient learning, better generalization, and the capacity for genuine creativity and systems engineering.
- True AI advancement requires models grounded in the physical world, enabling them to understand object interactions and causal relationships, which is crucial for tasks beyond simple prediction or mimicry.
- Current AI systems lack the ability to "know what they don't know," hindering their ability to adapt to novel situations; a key improvement would be systems that can identify uncertainty and seek external knowledge.
- The reliance on reward functions in AI is problematic due to inherent ambiguity and the potential for unintended consequences, suggesting a need for alternative alignment strategies that focus on understanding beliefs and values.
- Future AI development should focus on building cognitively inspired models that incorporate principles of biological intelligence, enabling them to learn and reason in ways analogous to human cognition.
Deep Dive
The future of artificial intelligence may lie not in scaling up current large language models, but in developing more sophisticated, modular AI systems inspired by the human brain's ability to reason, learn, and adapt. Current AI paradigms, largely focused on prediction and pattern matching, are reaching their limitations. A more promising direction involves creating AI that understands the world through objects and their interactions, mirroring human cognition and enabling true problem-solving and innovation.
The dominant approach in AI development, exemplified by large language models (LLMs) and transformer architectures, has focused on scaling up computational power and data. This has led to impressive capabilities in pattern recognition and prediction, but it fundamentally misunderstands how biological intelligence operates. Dr. Jeff Beck argues that the brain functions less like a prediction engine and more like a scientist, constantly forming and testing hypotheses about a world governed by objects and forces. This "Bayesian brain" approach, where the brain performs optimal inference and quantifies uncertainty, is a more robust foundation for genuine intelligence. The development of automatic differentiation (autograd) was crucial in enabling the current AI boom by transforming complex mathematical problems into engineering challenges, allowing for rapid iteration and scaling. However, this focus on function approximation has overshadowed the need for AI models that possess structured, causal understanding of the world, akin to how humans conceptualize objects and their relationships.
A key implication of this perspective is the need to shift from monolithic models to systems composed of numerous smaller, specialized models. This "lots of little models" approach, analogous to video game engines, allows for greater efficiency, modularity, and adaptability. Such a system would not rely solely on processing raw data like pixels or text but would build internal representations of objects and their interactions. This object-centered approach allows AI to possess a crucial capability: knowing what it doesn't know. When encountering novel situations, like a warehouse robot encountering a cat, the AI can identify its knowledge gap, seek external information (phone a friend), and integrate new knowledge, rather than making potentially catastrophic errors. This contrasts sharply with current AI, which often hallucinates or provides confident but incorrect answers when faced with unfamiliar data. Furthermore, grounding AI models in the physical world, rather than language, is essential. Human language is an unreliable narrator of both our internal thoughts and external reality; self-reporting often contradicts observed behavior. Therefore, AI should be grounded in physics and observable interactions, enabling it to perform complex tasks like engineering and innovation, which require understanding how components fit together and can be rearranged to create novel solutions.
The ultimate goal of this paradigm shift is to create AI that can not only perform tasks but also reason, adapt, and potentially innovate in ways comparable to humans. This requires moving beyond mere pattern matching to building systems that understand causality, uncertainty, and the structure of the world. The challenges ahead include scaling these more complex Bayesian inference methods and developing robust frameworks for managing vast numbers of interacting models. However, the potential payoff is artificial intelligence that is not just a tool for prediction, but a genuine partner in problem-solving and discovery, capable of tackling unprecedented challenges and expanding human capabilities. The development of AI that understands the world through its constituent parts and their interactions, much like humans do, is the path towards more robust, adaptable, and ultimately more intelligent systems.
Action Items
- Design object-centric models: Create modular components representing real-world objects and their interactions, enabling composition and reuse.
- Develop a framework for continuous learning: Implement systems that update models based on new data and interactions, avoiding static deployments.
- Integrate physics-based grounding: Prioritize models grounded in physical principles over language for more robust understanding and prediction.
- Build a library of specialized models: Develop numerous small, focused models for specific objects and phenomena, rather than monolithic systems.
- Implement uncertainty quantification: Ensure models explicitly represent and communicate what they don't know to enable adaptive behavior.
Key Quotes
"Bayesian inference provides us with like a normative approach to empirical inquiry and encapsulates the scientific method writ large right I just believe it's the right way to think about the empirical world... the essence of the bayesian approach is it's about explicit hypothesis testing and explicit models in particular generative models of the world conditioned on those hypotheses."
The speaker, Jeff Beck, posits that Bayesian inference offers a fundamental framework for understanding how we acquire knowledge and conduct scientific inquiry. He emphasizes that this approach involves rigorous hypothesis testing and the construction of generative models, suggesting it's the most accurate way to conceptualize how the world operates.
"The brain doesn't work like a giant prediction engine. It works like a scientist, constantly testing hypotheses about a world made of objects that interact through forces -- not pixels and tokens."
This quote highlights a core argument against the prevailing "prediction machine" model of the brain. Beck proposes that biological intelligence is more akin to scientific investigation, focusing on understanding objects and their interactions through fundamental forces, rather than simply predicting the next element in a sequence.
"The biggest thing was autograd [automatic differentiation]... it turned the development of artificial intelligence from being something that was done by like carefully constructing neural networks and then writing down your learning rules and then going through all that painful process it was tick tick for and they turned it into an engineering problem."
Beck identifies automatic differentiation as a pivotal development in AI. He explains that it transformed AI development from a laborious, manual process into a more accessible engineering discipline, enabling rapid experimentation with diverse architectures and approaches.
"The cat in the warehouse problem... what do we have so now we've we've we've got an ai agent that has been trained to like manage a warehouse... and then one day one day something comes along there's never seen before it's cat... the model has no idea has never seen a cat before... it's like what the hell is this and like it's you know it's screwing with my system... and then it says okay stop right don't run over the cat right let's figure out what's going on."
This anecdote illustrates a key challenge in current AI and the proposed solution. Beck uses the example of a warehouse robot encountering an unfamiliar object (a cat) to demonstrate how AI systems often lack the ability to recognize their own ignorance, whereas the proposed approach would allow the system to identify the unknown, seek external information, and integrate new knowledge.
"If you want an AI that thinks like we do you need to have it grounded in the same domain in which we are grounded and we are grounded in this domain right this is why the embodied bit is such an important thing... we want models that are grounded in the physical world in which we evolved and the reason for this is because that is the world that provides us with these atomic elements of thought."
Beck argues that for AI to achieve human-like intelligence, it must be grounded in the same physical reality that shapes human cognition. He emphasizes that our understanding of the world, built through embodied experience, forms the fundamental building blocks of our thoughts.
"The goal of alignment in an rl setting is to get you is to get it would be to somehow get my reward function or perhaps humanity's collective reward function right into the ai agent this is really really really hard... relying on arbitrarily selected reward functions like seems like a terrible idea."
This quote addresses the significant challenge of aligning AI behavior with human values. Beck expresses skepticism about current methods, particularly relying on arbitrarily defined reward functions in reinforcement learning, suggesting they are prone to unintended consequences and difficult to scale reliably.
Resources
## External Resources
### Books
- **"The Book of Why"** by Judea Pearl and Dana Mackenzie - Mentioned in relation to understanding causality.
### Articles & Papers
- **"A Theory of Human-like Artificial Intelligence"** (Source not specified) - Mentioned as a foundational document for the speaker's research.
- **"DeepMind's AlphaFold"** (Source not specified) - Mentioned as an example of AI success in a specific domain.
- **"Generative Adversarial Networks (GANs)"** (Source not specified) - Mentioned as a type of generative model.
- **"Mamba"** (Source not specified) - Mentioned as an alternative architecture to Transformers that benefits from scaling.
- **"Transformer"** (Source not specified) - Mentioned as a key development in AI, though its importance is debated relative to scaling.
- **"Vision-Language Models"** (Source not specified) - Mentioned as an example of models operating in pixel space.
### People
- **Alexey Guzey** - Mentioned as a collaborator.
- **Alexey Kizenko** - Mentioned as a collaborator.
- **Alexey** - Mentioned as a collaborator.
- **Andrej Karpathy** - Mentioned as a former director of AI at Tesla.
- **Andrew Ng** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Ben Goertzel** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Carl** - Mentioned for his work linking information theory and statistical physics, and evangelizing its application across disciplines.
- **Charles Darwin** - Mentioned in the context of evolution and adaptation.
- **Chris Olah** - Mentioned for his work on visualizing neural networks.
- **Daniel Dennett** - Mentioned in relation to consciousness and intentionality.
- **David Deutsch** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **DeepMind** - Mentioned as a source of research on AI.
- **Demis Hassabis** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Donald Knuth** - Mentioned in relation to algorithms and computer science.
- **Douglas Hofstadter** - Mentioned in relation to consciousness and analogy.
- **Elon Musk** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Geoffrey Hinton** - Mentioned as a pioneer in deep learning.
- **George Box** - Mentioned for the quote "All models are wrong, but some are useful."
- **Judea Pearl** - Mentioned for his work on causality.
- **Julian Togelius** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Karl Friston** - Mentioned as a key figure in active inference and his work on information theory and statistical physics.
- **Kevin** - Mentioned in relation to DreamCoder.
- **Leo Tolstoy** - Mentioned in relation to the concept of emergent phenomena.
- **Marcus** - Mentioned as a collaborator.
- **Mark Zuckerberg** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Max Tegmark** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Max** - Mentioned as a friend and collaborator.
- **Michael Levin** - Mentioned as a collaborator.
- **Nando de Freitas** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Nick Bostrom** - Mentioned in relation to AI safety and existential risk.
- **Nvidia** - Mentioned in relation to hardware for AI scaling.
- **OpenAI** - Mentioned as a leading AI research organization.
- **Paul Churchland** - Mentioned in relation to eliminative materialism.
- **Peter Norvig** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Richard Feynman** - Mentioned in relation to scientific inquiry and understanding.
- **Robert Sapolsky** - Mentioned in relation to biology and behavior.
- **Rodney Brooks** - Mentioned in relation to robotics and embodied AI.
- **Sergey Levine** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Shimon Edelman** - Mentioned in relation to cognitive science.
- **Simon Baron-Cohen** - Mentioned in relation to cognitive science.
- **Stanford** - Mentioned as an institution.
- **Steven Pinker** - Mentioned in relation to cognitive science and language.
- **Tegmark** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Thomas Bayes** - Mentioned in relation to Bayesian inference.
- **Timnit Gebru** - Mentioned as a researcher in AI ethics.
- **Tony Fadell** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Tony Zador** - Mentioned for his work on genetically encoding neural networks.
- **Vicarious** - Mentioned as a company.
- **Yann LeCun** - Mentioned as a proponent of the "AI is just scaling" viewpoint.
- **Zubin Gharami** - Mentioned for his explanation of the Dirichlet Process and Chinese Restaurant Process.
### Organizations & Institutions
- **DeepMind** - Mentioned as a research organization.
- **Google** - Mentioned as a company involved in AI research.
- **MIT** - Mentioned as an institution.
- **Northwestern University** - Mentioned as the institution where the speaker obtained their PhD.
- **OpenAI** - Mentioned as a leading AI research organization.
- **Stanford University** - Mentioned as an institution.
### Tools & Software
- **Autograd** - Mentioned as a key factor enabling AI development by simplifying gradient-based learning.
- **DreamCoder** - Mentioned in relation to program synthesis.
- **GPT-3** - Mentioned as an example of a large language model.
- **LangChain** - Mentioned as a framework for grounding models in a common linguistic space.
- **Microsoft** - Mentioned in relation to AI research.
- **PyTorch** - Mentioned as a deep learning framework.
- **TensorFlow** - Mentioned as a deep learning framework.
### Other Resources
- **Active Inference** - Mentioned as a framework inspired by statistical physics and information theory, with applications across various fields.
- **Bayesian Inference** - Mentioned as a normative approach to empirical inquiry and a model for how the brain works.
- **Bayesian Optimization** - Mentioned as a technique for optimizing Bayesian models.
- **Cellular Automata** - Mentioned as a system with emergent properties and potential for complex behavior from simple rules.
- **Chinese Restaurant Process** - Mentioned as a concept explained by Zubin Gharami.
- **Cognitive Models** - Mentioned as a focus for building AI that thinks like humans.
- **Complex Systems** - Mentioned as a field of study in the speaker's PhD.
- **Convolutional Cellular Automaton** - Mentioned in relation to Alex Mordvintsev's work.
- **Differentiable Programming** - Mentioned as a concept related to Autograd.
- **Dirichlet Process** - Mentioned as a concept explained by Zubin Gharami.
- **Emergence** - Discussed as a concept, with a critique of definitions based on ignorance.
- **Equilibrium Free Energy** - Mentioned in relation to active inference.
- **Free Energy Principle** - Mentioned as a foundational concept in active inference.
- **Gaussian Processes** - Mentioned as a development enabling Bayesian inference.
- **Generative Models** - Mentioned as a type of model used in Bayesian inference.
- **Information Theory** - Mentioned as a field linked to statistical physics and active inference.
- **Intention** - Discussed in the context of AI and human cognition.
- **Language Models** - Mentioned as a common approach for grounding AI models.
- **Lenia** - Mentioned as a system for simulating complex biological-like patterns.
- **Logic Gates** - Mentioned in relation to emergentist AI research.
- **Markov Blankets** - Mentioned as a concept used in active inference.
- **Natural Gradient Methods** - Mentioned as a technique for speeding up gradient inference.
- **Neural Networks** - Mentioned as a core component of AI research.
- **Normalizing Flows** - Mentioned as a deep learning tool for probabilistic reasoning, particularly with images.
- **Pattern Formation** - Mentioned as a field of study in the speaker's PhD.
- **Physics Discovery** - Mentioned as a research area related to understanding effective forces.
- **Prediction** - Discussed as a core function of current AI models.
- **Probabilistic Programming** - Mentioned as a related field.
- **Program Synthesis** - Mentioned as a method for creating programs, potentially aided by AI.
- **Reinforcement Learning (RL)** - Mentioned in relation to reward functions and AI decision-making.
- **Self-Supervised Learning** - Mentioned as a common approach in AI training.
- **Simulation** - Mentioned as a method for testing AI models.
- **Sparse Models** - Mentioned as a contrast to dense models like transformers.
- **Statistical Physics** - Mentioned as a field linked to information theory and active inference.
- **Stochastic Gradient Descent** - Mentioned as a method for training models.
- **Turing Completeness** - Mentioned in relation to cellular automata.