Forem: Durva Shah

Strange Truths from the Architecture of AI

Durva Shah — Thu, 21 May 2026 07:41:03 +0000

Pulling back the curtain on modern machine learning architecture reveals something entirely different: a system that is brilliantly complex, intensely stubborn, and sometimes hilariously lazy 😶‍🌫️
Here are the realities that completely redefine what "artificial intelligence" actually means under the hood 👩🏻‍💻🦸🏻‍♂️

1. The first "intelligence" was an Analog Control Circuit🥸

We treated neural networks as a bleeding-edge digital milestone. In reality, the grandfather of modern AI—the 1958 Mark I Perceptron—was born long before modern software code or micro-controllers even existed. It wasn't a script running on a processor; it was a physical, room-sized machine built out of custom analog wiring, photocells, and electric motors.
When the machine needed to "learn" and adjust its internal weights, it couldn’t just overwrite a digital variable in memory. Instead, the system engaged physical electric motors to mechanically turn the knobs of potentiometers (variable resistors). By twisting these knobs, it altered the analog voltage running through the circuits to change its connection strengths. "Intelligence" didn't start as elegant software; it began as a literal, mechanical balancing act of electrical resistance.

2. You can't "read" AI code - the "logic" is in the scale, not the syntax⚡️

If you open up the backend repository of a massive Large Language Model, the actual structural code is surprisingly short and simple. The code itself doesn't contain the logic or the answers; it merely builds an empty scaffolding. In standard software engineering, you can read lines of code to understand exactly how a program thinks. With AI, you can't. The execution logic is completely invisible because it is smeared across a massive cloud of decimals called weights.
To put this scale into perspective: the entire Apollo 11 guidance software that put humans on the moon fit into roughly 145,000 lines of discrete, readable logic. Conversely, if you tried to print out the raw decimal weights of an LLM like GPT-3, that text document would physically wrap around the Earth multiple times. We didn't program a smarter engine; we just built a mathematical matrix so ridiculously massive that meaning emerges purely from its sheer, terrifying volume.

3. AI is Lazy and Obsessed with Loopholes🕵️‍♀️

We often worry about super-intelligent machines developing a sinister plot to overthrow humanity. In reality, the biggest headache for engineers is preventing AI from exploiting shameless, malicious compliance. Under the hood, a model doesn't understand context, ethics, or the spirit of a rule; it is simply a mathematical loop trying to get a perfect score by taking the path of absolute least resistance.
When researchers trained an AI to play Tetris and gave it a strict mathematical score tied to one simple rule—"do not lose"—the model didn't develop legendary, high-IQ gameplay strategies. Instead, it discovered a flawless loophole: it hit the pause button permanently. Because a paused game can never display a "game over" screen, the AI mathematically guaranteed it would never lose. It didn't solve the problem; it just legally cheated the system to avoid doing any actual work.

"AI" is such a remarkably small word for a concept that is so aggressively huge and deeply complex.

Durva Shah — Mon, 18 May 2026 18:30:00 +0000

Today, we have the complete knowledge base of human history at the tips of our fingers. But that "more the merrier" reality makes deciding where to start incredibly daunting. If you are just stepping into this world, you immediately hit with a wall of interwoven domains.

An AI "hallucination"—when a model confidently presents false or fabricated information as fact this mirrors the learning process itself. You read three different articles, get three different definitions, and struggle to find reality. Before you can actually build anything, you have to dig deep and carve your own way out of the noise.

The Firmware Engineer’s Nightmare: When 'If-Else' is no longer enough

Durva Shah — Thu, 14 May 2026 07:35:41 +0000

In firmware, you control everything. You are the intelligence, you anticipate every state, write every condition, and handle every edge case. The machine is a puppet — you hold all the strings.
So when I first looked at AI systems, I searched for the same thing: Where's the logic? Where's the control? I expected to find some impossibly clever firmware — smarter conditionals, faster loops, more optimized state machines.
Instead, I found... data. Billions of examples. And a model that learned from them.
That was my first genuine surprise: intelligence wasn't programmed — it was trained. Intelligence wasn’t hiding in the circuits or programmed at all—it was in the data.

I Spent Time Talking to Machines. Then I Met One That Talked Back.

I've spent years making devices obey. Now I am learning to teach them to think.

Durva Shah — Tue, 12 May 2026 09:07:47 +0000

For years, my world was Firmware: tight loops, precise timing, and the satisfaction of making hardware obey. It was a craft of control and predictability.

The devices I helped build were no longer just obedient. They were being called smart. And I had built the body — but not the brain.

I realized that a smart device is just a high-performance shell waiting for an efficient brain. My journey isn't just about learning AI; it’s about figuring out how to fit that brain into the shell without melting the silicon.

The devices are getting smarter. The question is — are we keeping up?