The LLM is the Computer

0:00You know, I had this, um, this really humbling moment yesterday. I spent literally the entire morning conceptualizing this massive multi-tiered project architecture for our team felt like an absolute genius, right? Then we go to lunch to celebrate the bill comes and I completely freeze trying to calculate like a 20% tip on an $80 tab. My brain just totally blue screened. Yeah, that is the classic human paradox right there. Yeah. High-level abstract reasoning is totally fine, but basic arithmetic on the spot just causes a total system failure. Exactly. And I mean, that perfectly mirrors today's large language models. You already know they can solve, uh, Olympiad-level math theories or write complex code, but ask them to do simple step-by-step additionally and they just stumble. So for today's deep dive, we are looking at some incredibly fascinating new research from Percepta. Our mission is to figure out how they manage to literally embed a fully functional virtual computer directly inside an AI's architecture and why that basically changes everything about how models actually think.

1:01It is a massive shift. I mean, to understand why I'm betting a computer inside an LLM is such a breakthrough, we really have to look at the current bottleneck. Right now, when an LLM needs exact math, it pauses, writes a script, and sends it to an external Python interpreter. Right. It's, it's like how humans can't naturally fly. So we build airplanes. LLMs can't compute natively, so they just orchestrate external tools. But what if we actually gave them wings? Exactly. And the thing weighing them down has always been auto regressive decoding. As the model generates an execution trace, looking back at the memory prefix, takes quite a drastic time. It just gets so bogged down. Wait, so it's like a person with severe short-term memory loss trying to do long division. Every time they want to write down a single new digit, they have to reread their entire worksheet from the beginning. Yeah, just to remember where they are, that constant looking back is incredibly slow. The system cannot natively execute the steps efficiently at all. Okay, so before we talk about how perceptive actually solve that memory bottleneck,

2:02a quick note for you listening since I know you're always looking into where agents can make the most impact. If you need help with AI training or automation, integration, or software development, to uncover where agents can really transform your business or personal life, check out embersilk.com for all your AI needs. So back to perceptive, shrinking the memory to 2D solves this speed bottleneck. How does that actually work? Because I assume shrinking attention heads to just two dimensions would completely lobotomize the model. You would think so, right? It seems super counterintuitive, but geometrically, 2D is actually all you need for touring completeness. Wait, really? Just 2D. Yeah, just 2D. Meaning it has the architectural capacity to compute anything a standard computer can, but restricting the attention heads to two dimensions, they transform the mechanism from this slow linear scan into what's called a convex hole query. Oh, wow. So instead of scanning, like a million pass data points one by one, a convex hull is kind of like snapping a mathematical

3:03rubber band around the data. The model instantly finds the outermost boundaries it needs without having to read all the individual dots in the middle. That is a perfect way to visualize it. It mathematically targets the exact data point required, so memory retrieval becomes logarithmic rather than linear, which means it's exponentially faster because it stops rereading the whole worksheet. Exactly. And they proved this works in practice by running a web assembly interpreter directly inside vanilla pie torch weight. In some vanilla weights, that's insane. Right, meaning they built this using the standard building blocks of AI with zero external tools. They used it to solve the Ardo incolisodoku, which is widely considered the world's hardest. And it solved it with 100% accuracy in under three minutes. Yeah, streaming 30,000 tokens per second on just a standard CPU. Even Andre's carpathy praised the research as inspiring. The crazy part to me isn't just the speed. It's that they didn't need some massive server farm. The AI wasn't just guessing the next word. It was running a rigid, highly efficient,

4:06mathematical program directly inside its own neural pathways. Exactly. It completely bridges the gap between human-like intuition and machine-like precision. I mean, in the future, we won't just train models with gradient descent, you know, feeding them mountains of data and hoping they learn the patterns. It will literally compile C code directly into their weights. We'll be growing AI systems like modular software. Housing that kind of compiled logic inside a flexible neural net, changes our entire approach to problem solving. It's just incredible. Hey, if you are enjoying this deep dive, please subscribe to the show and leave us a five-star review if you can. It really does help get the word out. Thanks for tuning in. It definitely changes everything. It really does. And I want to leave you with a wild, genuinely optimistic thought to mull over. By merging this flexible creative intuition with flawless, compiled logic, we are unlocking an unimaginably bright future. I mean, this hybrid intelligence is exactly what will effortlessly solve our greatest supply chain challenges and completely revolutionize

5:07healthcare. The limitless potential of human ingenuity is really just getting started. But it leaves you with a fascinating question. If we are moving toward a world where AI doesn't just write our software, but its very brain is made of compiled code. Where does the software end and the AI begin?