The Fly in the Machine - Sam Glassenberg

The Fly in the Machine

A few weeks ago one of the most groundbreaking papers in computational intelligence of the 21st century was published. It’s comparable in impact to that of the attention mechanism – the paper that birthed the LLM explosion.

Why? This paper reveals how we, as a species, will ultimately defeat death.

The Story:

One day in March, 2024 a fruit fly woke up from a nap. It found itself in an eerily silent, dark room.

It felt some dirt on its antenna, which it promptly started grooming. Then it tasted something sweet and moved its mouth (more precisely, its proboscis) toward the sugar.

Only the room doesn’t actually exist. Or the sugar. Or the dirt.

The fly was simulated entirely on a computer.

Simulated down to each one of its 50 million synapses. Each synapse which had just been painstakingly mapped in one of the greatest scientific accomplishments of the 21st century: The complete mapping of the fruit fly connectome.

Fruit Fly Connectome
Fruit Fly Connectome

They Did It. They Uploaded a Fruit Fly.

Sure, it’s only 125,000 neurons. But it has most of the basic building blocks. The fruit fly can see, hear, and taste. It responds to caffeine and other neurochemicals.

Could this be the “Chicago Pile” moment for neural scanning and simulation?

Scale this up to 100B neurons and 100T connections and you can’t do this to a human. At which point, we’ve done it – we live forever.

Doing some back of the napkin math with ChatGPT, we aren’t very far away from the compute power needed to do this.

  • ANNs require approximately 2 FLOPs per simulated neural connection.
  • Simplified spiking models require ~10 FLOPs.
  • Hodgkin-Huxley Model simulates ion channel dynamics and membrane potentials using differential equations. Each time step requires solving several differential equations for ion currents, gating variables, and membrane potential. My gut tells me this is overkill but even that is around 1,000 FLOPS per neuron.

That puts us at human-level ANNs around 2030, and human-level simulations with HHM accuracy around 2041. I’m sure we’ll find plenty of optimizations to pull this in, as huge parts of the human brain follow highly repeatable patterns that can be abstracted.

Stay healthy, folks – if you can make it just a little longer you may just get to live forever.

Leave a Reply

Your email address will not be published. Required fields are marked *