Alexander Ksendzovsky, co-founder of The Biological Computing Co., details training AI models using living brain cells

Quite interesting thread on capabilities of real biological neurons (spoiler: they're way more capable than classical artificial neurons in a perceptron) . Nice work @IdoAizenbud and collaborators!

@JeffDean @IdoAizenbud Neurons are just the packaging. The real computers are the helical oscillator networks inside neurons (microtubules).

Watch on YouTube:

https://youtu.be/16NDovAP17U

Listen wherever you get your podcasts: Apple: https://podcasts.apple.com/us/podcast/tbcs-alex-ksendzovsky-the-neurosurgeon-using-brain/id1875709419?i=1000772353795

Spotify: https://open.spotify.com/episode/2i6cFZem4TUwyRoVm1xwEA?si=31fe4f1790fb408a

And you can catch more of my thoughts, including my 3 big takeaways, in our show notes on @UpstartsMediaCo here 👇

https://www.upstartsmedia.com/p/podcast-tbc-alex-ksendzovsky

@JeffDean @IdoAizenbud just a startup working with neurons on our podcast today, The Biological Computing Co

@JeffDean @IdoAizenbud so neurons can potentially be modeled by non-linear neural nets neuronception 🤓

@JeffDean @IdoAizenbud and here i thought i was special for using chatgpt turns out my cat has a better neural net

@JeffDean @IdoAizenbud Very interesting, reminds me of this book I read a while back:

A Thousand Brains: A New Theory of Intelligence

https://share.google/Zjdg2yD9pwk7TaM0G

@JeffDean @IdoAizenbud so biological neurons been doing everything alone this whole time while we sat here scaling compute

@alexrkonrad @BioComputingCo Fascinating, any data on latency versus traditional silicon?

@JeffDean @IdoAizenbud Not surprising — biological neurons are analog computers with 1000+ synaptic inputs each, while even a multi-layer perceptron still relies on a weighted sum and a fixed activation. The efficiency gap between wetware and silicon is still orders of magnitude wide.

Underrated implication: if a single cortical neuron needs a deep net to mimic, a lot of the brain's compute is hiding inside the units, not just in the wiring between them. Though I'd keep that separate from the mind question — how much a unit computes and whether a system adds up to a someone seem like different axes.

@JeffDean @IdoAizenbud 💯

@JeffDean @IdoAizenbud working on robotics, the bottleneck isn't how smart each neuron is. it's whether the whole model fits in memory and runs in <50ms.

the biological neuron thread is cool but i'm not sure it changes what we can actually deploy.

@alexrkonrad @BioComputingCo Another Alex - got to be worth listening to

@JeffDean @IdoAizenbud is the takeaway that one biological neuron is closer to a small multilayer net than to a single unit, or am I oversimplifying the result

@JeffDean @IdoAizenbud Astute take—biological brains outpace simple nets.

@JeffDean @IdoAizenbud Biology’s depth fuels smarter paths forward.

@JeffDean @IdoAizenbud But does it matter? If the systems all converge to the same representations, it may not make any difference for AI