The moment we stopped understanding AI [AlexNet]

The moment we stopped understanding AI [AlexNet]

Written by Massa Medi

Dive deep into the secretive, high-dimensional spaces that power modern artificial intelligence models like AlexNet and ChatGPT. Discover how these revolutionary neural networks make sense of our world, organize knowledge in ways humans struggle to visualize, and why scaling up simple mathematical operations can lead to stunning breakthroughs in machine intelligence. Whether you're an AI enthusiast, student, or just plain curious, this exploration of activation atlases and embedding spaces is your ticket inside the machine's mind.

The Hidden Geometry of AI: Enter the Activation Atlas

Imagine peering into a map—an atlas—not of countries or continents, but of the way AI models comprehend reality. This is the essence of an activation atlas: a tantalizing, visualized glimpse into the high-dimensional “embedding spaces” where artificial intelligence organizes everything it learns. It’s in these mathematical landscapes that models like OpenAI’s ChatGPT and the legendary AlexNet arrange their knowledge of language, images, and the world into intricate webs of similarity and meaning.

A Quantum Leap: AlexNet and the Dawn of Deep Learning

The paradigm-shifting moment came in 2012 with the debut of AlexNet. In just eight pages, this academic paper sent shockwaves through the computer vision community by achieving feats most thought impossible. Its revolutionary idea? Resurrecting older neural network concepts and simply scaling them up with more data and computational muscle. A breakthrough so potent that its co-author, Ilya Sutskever, would go on to co-found OpenAI—and push the same approach into the stratosphere with ChatGPT.

Under ChatGPT’s Hood: Stacks on Stacks of Transformers

You might expect to find something resembling human intelligence inside ChatGPT. Instead, you’ll find simplicity—layer upon layer of blocks called transformers (that’s the “T” in GPT for you acronym aficionados). Each transformer block does little more than churning through a set of fixed mathematical operations called matrix multiplications.

Here’s how it works: When you ask ChatGPT a question, it first splits your text into words and word fragments (“tokens”), then maps each of these into a mathematical object called a vector. These vectors are stacked into a giant “input matrix,” which gets passed through the stacked transformer blocks—96 times for ChatGPT 3.5, and reportedly up to 120 for ChatGPT 4. Every pass produces a new matrix, inching the AI’s understanding forward.

The absurdly simple secret? The next word ChatGPT generates is literally just the last column of the final output matrix—remapped from numbers back to text. The process then repeats: the new output text is appended to the prompt, transformed back into vectors, and run through the stack of blocks all over again, churning out a single new token or word fragment each cycle, until a special stop signal is reached. This relentless mathematical machinery, with no inherent understanding or intent, slowly morphs input into the elaborate outputs we see.

Where Is the Intelligence?

How can such a mechanical sequence of matrix multiplications generate everything from essays to code to poetry? The answer isn’t in the math, but in the enormous training datasets. These models are not designed by hand. Instead, they learn to represent knowledge by being trained on massive oceans of examples—trillions of words, millions of images—so that their vast, layered parameters encode the statistical structures of language and vision.

AlexNet: A Milestone in Visual Intelligence

While ChatGPT works with words, AlexNet’s realm is the pixel world. It takes an image, translates it into a three-dimensional grid (a tensor) representing RGB values, and ultimately spits out a one-dimensional vector with a thousand entries—one per possible class in the famous ImageNet dataset (think: ‘tabby cat,’ ‘hot dog,’ ‘aircraft carrier’).

What’s stunning is that AlexNet, like ChatGPT, strings together layers of computation (convolutional blocks) and learns to map inputs to desired outputs, all just by processing mountains of data. Yet, unlike language models, we can peek into these vision networks and see what they’re learning, layer by layer.

First Layer Revelations: Edge and Blob Detection

Early on, AlexNet develops a knack for visual patterns. The first five layers, known as convolutional blocks, use tiny learned filters (kernels) that slide over the image, scoring how similar each patch is to a pattern the model’s discovered. Visualize these 96 RGB kernels in the first layer: many have morphed into edge detectors (highlighting rapid color changes at different angles), while others become color “blob” detectors. None of this is coded by hand—every kernel begins as random noise and learns purely from exposure to data.

When these kernels interact with an input image, they generate activation maps, revealing which parts of the image “excite” or activate the kernel—bright glow for matches, darkness for disinterest. For instance, if you show AlexNet a pattern reminiscent of one of its edge kernels, the corresponding activation map will light up, but twist the pattern 90 degrees, and the activation fades: the “alignment” is gone. Beyond edges and blobs, stacking up the kernels’ outputs enables the network to perceive increasingly complex features as you move through the layers.

From Edges to Concepts: Deep Feature Stacks

In higher layers, the complexity compounds. The 96 activation maps from the first layer become the 96 “color channels” for the second layer, where new kernels can only be visualized abstractly. Still, we can study what activates these deeper kernels: they may light up for basic corners, outlines, or, further up, faces—even when the dataset never directly tells the model what a face is! By layer five, AlexNet might have a kernel lighting up for faces despite never seeing an explicit “face” label. It’s pattern abstraction at its finest.

Probing the Brain: Feature Visualizations and Embeddings

Want to know exactly what a kernel recognizes? Examine which training images most maximally activate it—for a “face” kernel, you’ll consistently find images with human faces. Or flip it: use optimization tricks to generate synthetic images that best excite a particular kernel, painting a kind of neural Rorschach test showing the pure essence of what excites that part of the network.

By the model’s penultimate layer, the data is distilled into a 4096-dimensional vector, sometimes called an embedding. Each image is now a point in this hyperspace, and finding “nearest neighbors” in this space reliably retrieves conceptually similar images (elephants cluster with elephants, tigers with tigers), even when their pixel-level patterns are vastly different. This clustering reveals that AI learns meaning by organizing concepts in high-dimensional, geometric spaces.

Directions within these embedding spaces are imbued with unexpected meaning. For example, morphing along an “age” or “gender” direction in a face-embedding space can alter a photo’s perceived age or identity—a trick now famous in many face-manipulation demos.

Beyond Images: The World of Activation Atlases

The magic really unfolds when researchers blend these embedding spaces with synthetic feature visualizations, compressing them into two-dimensional layouts—a feast for the eyes known as activation atlases. On these maps, nearby points often represent visually or conceptually similar entities: zebras smoothly morph into tigers, then leopards, then rabbits, showcasing the continuity of learned visual concepts. In intermediate layers, the model might cluster images by abstract characteristics, such as the number and size of fruits.

And it’s not just vision: in language models, words and fragments are embedded in similar high-dimensional spaces, where proximity signifies semantic similarity, and even the “directions” can denote analogies (“king” minus “man” plus “woman” lands you at “queen”).

Manipulating Meaning: Language Embeddings in Action

Recent research from Anthropic has shown we can tweak these “concepts” directly within language models. By artificially boosting activations connected to a phrase like “Golden Gate Bridge,” not only do models focus on that concept, but they might even hilariously start identifying as the Golden Gate Bridge—highlighting how deeply these directions are tied to meaning.

AlexNet: From Black Box to Breakthrough

AlexNet clinched the 2012 ImageNet Large Scale Visual Recognition Challenge in resounding fashion, leapfrogging approaches that, on the surface, seemed far more “intelligent.” Until that point, AI models often relied on suites of hand-engineered algorithms, such as the 2011 champion’s highly specialized SIFT technique. AlexNet, by contrast, ran on a simple principle: let the network learn everything from scratch, shaped only by millions of examples.

This approach dates to the 1940s’ earliest artificial neuron models by McCulloch and Pitts, and the physically built “perceptron” machines of the 1950s. In the 1980s, Geoffrey Hinton (one of AlexNet’s creators) helped pioneer a vital learning algorithm called backpropagation, enabling multilayer networks to efficiently learn from their own mistakes. Even then, neural networks rarely went deeper than a few layers, and successes (such as early self-driving cars or handwriting recognition) were considered outliers rather than a revolution-in-waiting.

The Real Difference: Scale Up, Blow Minds

So what changed in 2012? Scale. The ImageNet dataset provided more labeled images than any previous attempt (1.3 million!). Thanks to modern Nvidia GPUs, Hinton's team wielded roughly 10,000 times more computation than pioneers like Yann LeCun had a decade earlier. AlexNet’s 60 million parameters dwarfed prior efforts and paved the way for today’s titans—ChatGPT’s parameter count now soars above a trillion.

The lesson is profound: simply making old ideas bigger—deeper networks, more parameters, more data—can flip performance from mediocre to magical, with abilities so complex that even researchers struggle to interpret how the machines do what they do.

And while tools like activation atlases shed light on a portion of these mysterious spaces, for every “face” or “Golden Gate Bridge” neuron we recognize, there are thousands more encoding concepts that defy easy explanation (and may not even have words in human language).

Hands-On Learning Spotlight: Kiwico

Before we venture even deeper into embedding spaces, let’s give a grateful nod to this exploration’s sponsor: Kiwico. Kiwico crafts imaginative, hands-on project crates for children and learners of all ages, boasting nine distinct monthly subscription lines that nurture curiosity in everything from science to engineering. You don’t have to subscribe blindly—they also offer individual projects, perfect for experimenting or gifting.

The creator recalls their own obsession with building things as a child—constructing towers, tinkering with electronics, and the thrill of deep, self-driven project-based learning. Now, as a parent, they cherish Kiwico’s thoughtful approach. Projects like the Eureka crate’s pencil sharpener ignite passion for STEM, while the Panda crate invites even toddlers to explore motor skills with inventive crayons (which, as any parent knows, quickly become coveted treasures in the backseat).

If you want to nurture hands-on learning and support future videos like this, consider exploring Kiwico’s offerings. Use the code Welch Labs for 50% off your first month!

Visualizing the Unknown: The Road Ahead for AI

Activation atlases are a window into spaces so vast and intricate that our human intuition falters. The distances and directions in these spaces encode relationships, analogies, and attributes we can sometimes recognize—but often, we’re only scratching the surface. Each projection or visualization collapses dizzying dimensions down for our inspection, but the deeper organization remains elusive and, frankly, awe-inspiring.

A History of Surprising Leaps—and What Comes Next

No one in the early 2000s predicted that decades-old neural networks, scaled up by a few orders of magnitude, would leave handcrafted algorithms in the dust. Even fewer foresaw that the same building blocks could generate AI models like ChatGPT, powerful enough to write essays, summarize books, and generate code—yet often mystifying in their inner workings.

Could the next AI breakthrough be waiting, once again, at a new scale? Or perhaps, hidden in a long-overlooked technique ready to surge back into the spotlight, just as AlexNet did in 2012? The only certainty is that the story of AI’s evolution is far from over.

Are AI’s compute blocks “dumb?” Not at all. Calling them “dumb” only emphasizes just how impressive it is when such simple, mechanical pieces combine—guided by oceans of data and skillful algorithms—to create the illusion (and, in some tasks, the reality) of intelligence.

Recommended Articles

Geoffrey Hinton: The “Godfather of AI” Sounds the Alarm: From Neural Nets to Nobel Prizes and the Uncharted Future of Artificial Intelligence

AI

Geoffrey Hinton: The “Godfather of AI” Sounds the Alarm: From Neural Nets to Nobel Prizes and the Uncharted Future of Artificial Intelligence

Geoffrey Hinton: The “Godfather of AI” Sounds the Alarm: From Neural Nets to Nobel Prizes and the Uncharted Future of Artificial Intelligence

The Rise of Model Context Protocol (MCP): Why Every Developer Is Talking About It

The Rise of Model Context Protocol (MCP): Why Every Developer Is Talking About It

Inside the Magic of Large Language Models: How AI Autocompletes Human Thought

Inside the Magic of Large Language Models: How AI Autocompletes Human Thought

Visual representation of AI evolution from rule-based systems to cosmic intelligence

The Evolution of Artificial Intelligence: From Rules to Cosmic Consciousness

A Hands-On Review of Google’s AI Essentials Course: 5 Key Lessons, Honest Pros & Cons, and Is the Certificate Worth It?

A Hands-On Review of Google’s AI Essentials Course: 5 Key Lessons, Honest Pros & Cons, and Is the Certificate Worth It?

Why Human Connection Still Beats Technology—Even in the AI Era

Why Human Connection Still Beats Technology—Even in the AI Era

The Future of Jobs: Which Careers Will Survive the AI Revolution?

The Future of Jobs: Which Careers Will Survive the AI Revolution?

The Truth Behind Those 'Handcrafted' Leather Bags and Watches

The Truth Behind Those Handcrafted Leather Bags and Watches: How AI, Actors, and Cheap Goods Are Fooling Shoppers Online

Will AI Replace Programmers? A Veteran Engineer on the Future of Software Jobs

Will AI Replace Programmers? A Veteran Engineer on the Future of Software Jobs

A split screen showing a programmer in a remote interview with hidden cheating methods illustrated

Inside the Secret World of Technical Interview Cheating: Tactics, Temptations, and Terrible Consequences

 AI Agents Demystified: The Step-by-Step Guide for Non-Techies Using Real Life Examples

AI Agents Demystified: The Step-by-Step Guide for Non-Techies Using Real Life Examples

Is AI Making Us Dumber? Navigating the Cognitive Costs of Automation in the Knowledge Age

Is AI Making Us Dumber? Navigating the Cognitive Costs of Automation in the Knowledge Age

The Death of Coding: Why Chasing Tech Jobs Might Keep You Broke in the Age of AI and Bitcoin

The Death of Coding: Why Chasing Tech Jobs Might Keep You Broke in the Age of AI and Bitcoin

Beyond the Nobel: Demis Hassabis, DeepMind, and the Race Toward Superhuman AI

Beyond the Nobel: Demis Hassabis, DeepMind, and the Race Toward Superhuman AI

Microsoft’s Majorana One Chip: The Topological Quantum Leap That Could Change the Future of Computing

Microsoft’s Majorana One Chip: The Topological Quantum Leap That Could Change the Future of Computing

All Machine Learning algorithms explained

All Machine Learning algorithms explained