Biological Compute

In this article, I'm going to be talking about a field in Computer Science and Deep Tech that I have been following for the greater part of 3 years. Let me start by laying out my view for why this field will only continue to become more and more important over the coming decade.

I believe the hunger for more compute will eventually extend past what is possible with Moore's law, mostly due to AI demand (duh), but also because of continuing demand for data and the processing of it, from both enterprises to manage their systems, and consumers to store and manage their digital lives. To solve this, we'll need a radically new way to power the machines we use day to day; our current solutions are far too costly to build and sustain long term.

But what does that cost look like? Well, let's do some crude math. The current fastest HPC supercomputer (El Capitan) consumes around 30 megawatts, which can power tens of thousands of households. For that price in energy, you get roughly 1.8 quintillion calculations per second. This lets us do important things with high precision, such as nuclear and weather simulations. With AI, these numbers get even crazier. You have some training clusters (like xAI's Colossus 2) aiming to be the first gigawatt cluster. This is not sustainable if it continues to scale at this rate unless we literally harness the sun's energy.

Well, what if there was something that consumed 50 million times less power than a gigawatt cluster, yet possessed the raw architectural efficiency to rival our best machines? Well, it exists, and it is the human brain. The human brain runs on roughly 20 watts, which is about the power of a dim lightbulb.

You might think I'm insane for suggesting you can turn a brain into a programmable computer, but not only is it possible, it already happened, and you can even rent a biological computer server right now if you want. Back in 2022, a bunch of neurons grown inside a petri dish were taught to play Pong with the right reward signals.

Just a few days ago, that exact same company hooked up 200,000 living human neurons to Doom. Those cells are navigating the map, seeking out enemies, and firing back. An independent developer managed to teach a petri dish how to play Doom in under a week using a basic Python API.

So the real question becomes, if we want to achieve real intelligence, why don't we train our models on the form of computing we know brought about intelligence?

This is why I've been following Cortical Labs (who are based right here in Melbourne) so closely for years. Yes, it's straight out of science fiction (and potentially dystopian), but it's also incredibly promising. Silicon got us to the modern age, but to reach the next level of intelligence and computing efficiency, I believe the future of hardware is going to be wet, squishy, and alive.