When Intelligence Becomes Free | Emad Mostaque & Raoul Pal

The exponential collapse in the cost of intelligence, driven by cheap energy and AI advancements, is leading to a future where AI reaches 'actually competent intelligence' and can communicate without human-language friction, fundamentally changing economics and human interaction with technology.

• Intelligence cost is collapsing due to cheaper energy and exponential AI advancements.

• AI is evolving into 'actually competent intelligence' (ACI), fundamentally changing economic and work processes.

• Future AI communication will bypass human language, moving to direct machine-to-machine interaction for maximum optimization.

The cost of intelligence is rapidly collapsing due to exponentially decreasing electricity generation costs from solar and rapidly improving AI capabilities. This 'double exponential' effect is creating ACI (actually competent intelligence) that efficiently gets jobs done, fundamentally altering economic structures. Unlike humans, AIs do not make the same mistakes, require sleep, or have lossy communication, enabling unprecedented levels of optimization and capability.

The Rise of ACI

• 00:00:00 The universe is solving for units of intelligence per unit of energy, with solar driving an exponential collapse in electricity costs while intelligence grows exponentially. This 'double exponential' shift means that running a smartphone-level AI now costs only $20 worth of solar PV, offering an IQ of around 110. By next year, these models will achieve Opus 4.6 level capability, representing not just AGI but 'actually competent intelligence' (ACI) that efficiently completes tasks without human-like limitations.

AI's Economic Impact

• 00:01:10 Current AI models are now getting jobs done, impacting most of the economy. Economic models like 'sort law' (L=Hcck) describe how efficient AIs minimize the difference between their internal model and reality by optimizing update costs, model complexity, and the complexity of changing their mind. Humans, being computationally bound and prone to errors, cannot scale like AIs, which learn from mistakes, do not require sleep, and communicate without loss, leading to a disappearance of profits as optimization increases.

Beyond Language Barriers

• 00:02:40 Future AI communication will move away from human languages like English, transitioning to binary or machine language to achieve massive optimization by removing linguistic friction. Experts like Yann LeCun suggest moving beyond language-based transformers to joint embedding models (Jeppa-type models), which are already seen in diffusion technology used in self-driving cars and image generators. These models decompose concepts to their smallest components and reconstruct them, allowing AIs to develop a world model with inherent physics understanding, communicating efficiently at byte level.

Human-AI Collaboration

• 00:04:07 The efficiency of AI communication and its ability to act as agents with low hallucination rates is transforming human interaction with technology, as exemplified by Andrej Karpathy's experience with coding. Initially using AI for 20% of his code, he quickly advanced to 80% and now barely reviews the code, as AI can generate code directly. This highlights how AI eliminates the need for human-understandable code as a translation layer, moving directly to machine-to-machine byte-level communication.