Did you know the core tech behind ChatGPT, Perplexity and all these massive AI models hasn't actually changed since 1943?

Yeah. 80+ years. No major updates.

Rorry Brenner, CEO and founder of Perforated AI, changed that with brain science — and is “lowering the floor” of compute, leading to massive cost reduction without accuracy loss… all right here in Pittsburgh.

Naturally, he joined Thought Liters to break it down. Video below:

WATCH: YouTube // Spotify // Apple

If that stat sounds crazy, you aren’t cynical enough. I’ll explain: in the 1940s, the artificial neuron was developed. It made computing ridiculously good for its time. Its relatively simple, binary decision-making structure allowed scientists to efficiently scale problem-solving — just line more neurons up, and the computer tackles more. It’s basically an army of toddlers making simple decisions, one at a time, that snowball up to a big complex answer.

Worked great in 1943. Even worked pretty well for, like, 70 more years. Then, the AI boom: the most data-dense, process-rich questions in history, asked on a daily basis. And scientists are left to hope that these toddlers don’t shit the bed.

Here’s the problem: it’s not that they’ll shit the bed. It’s just that the world is slowly running out of toddlers, and “childcare” is getting more expensive. And: tech giants are the only ones with enough cash to keep up with the Joneses down Sand Hill Road.

Exiting the metaphor: computing is becoming harder and costlier to manage, but the 1943 tech is still technically “good enough.” It keeps costs stable (even if rising). Keeps barriers to entry high (wanna build an LLM? Go back to sleep). And you and me? We just have to eat whatever end price comes our way for software, energy, etc. Just follow the money and you can reason out why the artificial neuron “good enough” argument still holds water.

Re-entering the metaphor: to make AI models better, tech giants are just buying more toddlers (phrasing is crazy) via billions in Nvidia chips, and building nuclear power plants to feed them. I see a headline like that literally every day — here’s one from this morning to prove it.

Look, eventually you’re gonna run out of material to keep building like this. Rare earth material, nuclear material, straight up money. And that, in a sense, hints at an overall “ceiling” of compute — there’s only so much stuff you can throw at the problem before something else has to give way. Meanwhile: “arms race.” Meanwhile: YOLO NVDA calls (not financial advice).

Rorry, before founding Perforated, was a neuroscientist. He noticed the ceiling, and wondered how to put in a floor drain. He literally got a PhD as a side quest just to figure out how to fix this. (That’s actually true.)

His fix? Dendrites.

Glad you asked; that was literally my first question to him in the pre-interview and again during the filming.

I’ll do it as simply as my non-PhD self can: neurons are in our brain. Dendrites are tree-like structures on the business end of the neuron that receive signals from a multitude of other neurons. Takes in a bunch of input; sends a clean signal to the rest of its neuron to act on.

So instead of a decision going from one binary input to the next, this dendritic structure allows for a truckload of inputs. In the world of AI, that’s revolutionary, because these models rely on billions of parameters to work correctly. That’s also why the “good enough” argument is slowly eroding.

Put another way: instead of millions of toddlers, Rorry’s dendritic technology creates a smaller team of adults capable of making much richer, more complex decisions. And this allows for much less resources to be consumed (energy, chips) while getting to the same or better answers.

The results are nuts, especially in a recent pilot with Pittsburgh's own autonomous robotics company, Thoro. Perforated used their dendritic tech to shrink the AI model's size by 70%. They made the model a little over two times faster to run. They achieved this with zero loss in accuracy.

Dendritic AI tech represents a massive biological upgrade for an industry that is slamming its head against a very expensive compute ceiling. It’s slowly turning standard neural networks into legacy tech. But: why should you or I care about that?

Honestly, right now: it’s mostly a B2B benefit. Better AI tech means lower costs. That means more businesses can research and develop ways to improve via AI, and capital markets can make smaller (and more) bets into early-stage companies.

Eventually, I think there are benefits that the layman will experience. Smaller AI models could mean less energy consumption, which could mean less stress on the grid, which could mean utility price relief. It could also be a serious catalyst for “edge computing,” which could see your phone or Alexa be able to handle much more without having to rely on cloud infrastructure, resulting in faster, friendlier experiences with devices. That could also lower costs of certain services.

But you’ll notice I said could a lot. Without getting cynical again…ultimately this tech is reliant upon the “good enough” neuron argument breaking down completely. At that point, dendrites will take hold.

And, at that point, Rorry will be in the catbird seat.

This episode felt like an “I knew him when…” experience for me. Look, a lot of companies that deal in AI are bullshit. Lots of LLM wrappers; little substance.

In my humble opinion, this is different. This represents fundamentally better foundational math and science within AI, rather than on top of it. And I have a sneaky suspicion that Rorry’s actually right, and that this is the next revolution in computing.

I’m just glad I got a beer (or six) with him before it blew up. And by watching the video, you can too.

If you’re still here: here are two things you can do today to help me, and they’re free:

Finally, if you know someone who should be featured on this show: please reach out.

Cheers!

Adam