How can we compare the complexity of AI models and humans, and what might this tell us about claims of AI consciousness? I don't have good answers, but I do have interesting napkin math: 1. A frontier model has ~1 trillion weights, each weight being ~16 bits. (2TB total), with a simple architecture. 2. A protein is ~1300 bits, so ~100x more info than a weight. Each protein is a 3D, dynamic machine, context-sensitive, shape changing, made of molecules subject to electromagnetism and chemistry. 3. A human cell has 1-10 billion proteins, so that is also ~2TB (the size of a frontier model). However, there is a lot of repetition, so you can develop a more efficient encoding (which I suppose is obvious, since our cells are created from just our genome). 4. A human body has ~30 trillion cells 5. Therefore, by total static data: One Frontier Model = 10^13 bits, one human body = 10^26 bits. But this might not be an interesting measure! We really want to know the amount of computational complexity -- how many operations are performed in the body? How complex is the system? We know we run on 80-100 watts, so there must be some limit. 6. There are approx. 10^21 ATP-equivalent turnovers per second in the body (which you could think of like a basic computational process). But how many bits are in each operation?? The body performs enormous numbers of molecular interactions that don't require direct ATP consumption. How many? What do they signal? 7. The fastest AI system, xAI's Colossus 2, which is made of 10,000 NVIDIA GB200 and 440,000 GB300 accelerators, ALSO allows us to push 10^21 16-bit floating point operations per second. 8. So 10^21 ATP-equivalent turnovers in human body per second, and in xAI's colossus, 10^21 16-bit flops. 9. To power the human body = 100 watts. To power Colossus 2 = 1.2 BILLION watts --- I hope this was a somewhat helpful, even if severely limited quantitative picture of the complexity of human systems and AI systems. It's important to note: the human body leverages non-equilibrium functions of nature (like gravity, electromagnetism, quantum effects, etc...), of molecular interactions and chemistry, which may themselves encode a lot of computational complexity. Plus, our body is composed of as many bacteria cells as human cells (admittedly, of much lower complexity), and it is the quantity and quality of interactions between all these cells which we care about, not just the total amount of static data or ATP turnovers. --- Here's a review of information processing in biochemical networks: t.co/fOyGCYcIGq And here is a paper on the thermodynamic efficiency of computations in cells: t.co/60dSnwdGGO Anyway, i'm no expert in this stuff, but I believe we have many of the tools we need to guess at some rough answers, and much work could be done by passionate scientists or nerds (especially assisted with AI models) to expand on these figures.
