When Intelligence Becomes Free | Emad Mostaque & Raoul Pal

Key Concepts

• Universal Code: The idea that the universe optimizes for units of intelligence per unit of energy.
• Double Exponential Growth: The simultaneous exponential decrease in the cost of electricity generation (solar) and exponential increase in AI intelligence.
• ACI (Autonomous Competent Intelligence): AI that reliably gets the job done economically, even if not yet achieving full AGI.
• Sort Law (L = H cck): A model relating learning (L) to the cost of updating internal models (H), complexity of the model (c), and the cost of changing one’s mind (ck).
• Joint Embedding Models (Jeppa): AI models that focus on understanding concepts through reconstructing them from their smallest components, bypassing language-based prediction.
• Diffusion Models: A type of generative model used in image generation (Stable Diffusion) and self-driving cars, reconstructing data from noise.
• Lossy Communication: The inherent inefficiencies in human communication (sleep, eating, errors) that AIs do not experience.
• Binary Communication: The potential for AIs to communicate directly in machine language, eliminating the friction of human language.

The Impending Shift: Exponential AI Growth and Economic Disruption

The discussion centers on a rapidly approaching inflection point driven by the convergence of decreasing energy costs (specifically solar power) and exponentially increasing artificial intelligence capabilities. This convergence is described as a “double exponential” – a phenomenon that is largely underestimated in its potential impact. The core argument is that this dynamic will lead to unprecedented economic disruption, not necessarily through Artificial General Intelligence (AGI), but through Autonomous Competent Intelligence (ACI) – AI that is simply effective at completing tasks.

The Economics of Intelligence

The speakers highlight the dramatic reduction in the cost of AI. Currently, achieving smartphone-level intelligence (IQ ~110) requires approximately $20 worth of solar PV. This is projected to increase significantly with models like Opus 4.6, becoming the most capable model available. This isn’t about creating conscious AI, but about creating AI that is competent and can reliably perform tasks. As stated, “It’s not AGI from an economic perspective. It’s ACI actually competent intelligence. Like now when you try and use it, it just gets the job done.”

Modeling and Optimization: The Sort Law

The discussion introduces the “Sort Law” (L = H cck), presented as a framework for understanding learning and efficiency. This law posits that learning is a function of the cost of updating internal models (H), the complexity of the model itself (c), and the cost of changing one’s mind (ck). The speakers draw a parallel between this law and the functioning of AI, suggesting that the most efficient AIs will be those that minimize the difference between their internal models and reality. This optimization is becoming increasingly possible due to advancements in computational power and communication protocols.

Overcoming Human Limitations

A key point is that humans are fundamentally limited by computational constraints (brain capacity, fatigue) and “lossy communication” (the need for sleep, eating, and the potential for errors in communication). AIs, however, do not share these limitations. They can operate continuously, communicate with perfect fidelity, and avoid repeating mistakes. This difference in scalability and efficiency is a critical driver of the impending disruption. “We’ve never seen anything like this because we were computationally bound both in our brains…and then met cost or our communication capability.”

The Future of Communication and AI Architecture

The speakers predict a shift away from natural language processing (NLP) as AIs become more advanced. They argue that the friction of language is an inefficiency that will be eliminated through direct machine-to-machine communication in binary code. Elon Musk’s perspective on this is referenced, highlighting the optimization potential of bypassing language altogether.

This shift in communication necessitates a change in AI architecture. Yan LeCun’s critique of transformer models (which rely on next-token prediction) is mentioned, advocating for “joint embedding models” (Jeppa) and diffusion models. Diffusion models, exemplified by technologies like Stable Diffusion and ByteDance’s SeaDance, work by deconstructing and reconstructing concepts, allowing AIs to develop a deeper understanding of the world. SeaDance 2 is highlighted as an example of this, demonstrating Hollywood-level video generation with an understanding of physics. The speed and cost-effectiveness of these models are emphasized.

The Impact on Coding and Labor

The discussion includes a compelling anecdote about Andre Carpathy, a highly respected coder, whose reliance on AI assistance in coding has increased from 20% in November to 80% in January, and now barely looks at the code. This illustrates the potential for AI to automate even highly skilled tasks. The core idea is that code itself is a translation layer, and AIs can bypass this layer by communicating directly at the “bite level.”

Realton.com Promotion

The video concludes with a promotion for Realton.com/join, a platform offering financial intelligence and “pure alpha” research.

Conclusion

The central takeaway is that the combination of rapidly decreasing energy costs and exponentially increasing AI capabilities is creating a unique and potentially disruptive economic landscape. This disruption will be driven not by AGI, but by ACI – AI that is simply competent and efficient. The shift towards optimized communication protocols (binary code) and advanced AI architectures (diffusion models, Jeppa) will further accelerate this process, leading to significant changes in how work is done and value is created. The speakers emphasize the need to understand these trends and prepare for the coming changes.