Karl Friston and Gary Marcus on AI's Future, LLMs, and Active Inference

Critiques of Current AI & LLMs

• ⚠️ Gary Marcus criticizes Large Language Models (LLMs) for producing "authoritative B.S.," being unreliable, prone to hallucinations, and exhibiting overconfidence without genuine understanding.
• 💡 Marcus highlights an "intellectual monoculture" in AI, where scaling LLMs is pursued despite fundamental limitations like boneheaded reasoning and energy inefficiency, arguing it's not a sustainable path.
• 🧠 Karl Friston views deep learning as a "distraction" from proper AI, noting that current generative AI focuses on content generation rather than developing generative models for understanding the causes of sensory inputs.

The Role of Uncertainty and Agency

• 🔑 Friston explains that Yann LeCun's approach misses a crucial element by setting "temperature to zero," effectively removing uncertainty from neural networks, which prevents principled uncertainty quantification and information seeking.
• 🎯 Without the capacity to encode uncertainty, systems cannot evaluate model quality, know what questions to ask, or exhibit true agency and curiosity.
• 🚀 Agency in AI requires generative models that predict the consequences of actions and drive information-seeking behavior to resolve uncertainty, a capability current LLMs lack.

Active Inference and Generative Models

• ✅ Friston emphasizes that minimizing free energy involves maximizing entropy, aligning with Ockham's principle and complexity minimization, which keeps options open and avoids overly precise explanations.
• 💡 Current machine learning models, unable to encode uncertainty, cannot measure their own complexity, leading to overfitting and reliance on engineering workarounds like dropout.
• 🌱 Active Inference offers a first-principles approach to building adaptive, autonomous systems by optimizing learning through epistemic foraging, efficiently seeking information to resolve uncertainty and minimize energy waste.

Towards a Principled AI Future

• 🔮 Both Friston and Marcus advocate for AI systems grounded in explicit world models that can be interrogated and provide a causal understanding of the environment.
• 🤝 Marcus suggests that future AI needs to integrate both fast, automatic neural networks and slow, deliberative symbolic AI, moving beyond blank-slate learning to incorporate innate priors for more robust intelligence.
• 🌍 Friston envisions a future where AI helps resolve global uncertainties and problems, fostering mutual understanding and communication, while celebrating the diversity of intelligences as essential for sustainable ecosystems.

World Models and Neuro-Symbolic Approaches

• 🧩 Marcus argues that LLMs lack true world models because they cannot consistently apply rules or demonstrate coherent understanding, even for simple domains like chess, despite vast training data.
• 🛠️ The current trend of "bolting on" symbolic systems (like Python interpreters) to LLMs is a "Frankenstein" approach that, while showing some improvement, highlights the need for a principled neuro-symbolic integration rather than superficial additions.
• 🔬 Friston distinguishes between implicit world models (content-to-content mapping) in current generative AI and true understanding world models (cause-to-content mapping), stressing the latter as fundamental for genuine intelligence and agency.