Artificial intelligence has been promised and overhyped for decades. But the current wave — large language models, generative AI, multimodal systems — is genuinely different from what came before. This lesson gives you a grounded foundation before we go deeper.
The Honest Definition
AI, at its core, is software that performs tasks we previously thought required human intelligence. That's it. There's no magic, no alien consciousness — just mathematics applied at scale.
The key insight: intelligence is a spectrum, not a binary. A thermostat is a tiny bit intelligent. A chess engine is more so. A large language model like Claude or GPT-4 sits much further along that spectrum — but the fundamental nature of what it is (software running on hardware) hasn't changed.
Three Waves of AI
Wave 1 — Rule-Based Systems (1950s–1980s)
Early AI worked by encoding rules explicitly. "If the patient has a fever AND a cough, suggest flu." These systems were brittle — they broke whenever reality didn't match the rules. But they worked remarkably well in narrow, well-defined domains.
Wave 2 — Machine Learning (1990s–2010s)
Instead of writing rules by hand, we let machines learn rules from data. Feed a system thousands of spam emails and normal emails, let it find patterns, and it learns to distinguish them. This was the era of support vector machines, decision trees, and eventually deep learning.
Wave 3 — Foundation Models (2017–present)
The current era. Models trained on internet-scale data using the transformer architecture. Instead of narrow specialists, these are generalists: the same underlying model can write code, analyze images, translate languages, and reason through complex problems.
Why This Wave Is Different
Three things changed simultaneously:
-
Data: The internet created an unprecedented corpus of human knowledge and communication — trillions of words, images, and code.
-
Compute: GPU clusters gave us the processing power to train models at previously impossible scales.
-
Architecture: The transformer (introduced in "Attention Is All You Need," 2017) proved dramatically more effective than prior approaches.
The result: a qualitative shift. These aren't just better spam filters. They're systems that exhibit genuinely surprising emergent capabilities — abilities that weren't explicitly trained for, that emerge from scale.
What AI Is Not
Before we go further, let's clear up common misconceptions:
AI is not sentient. Current models have no experiences, desires, or consciousness. They're pattern-matching engines of extraordinary sophistication — but there's nobody home.
AI is not infallible. Models hallucinate confidently. They can reason incorrectly. They have training cutoffs. They reflect biases in their training data.
AI is not replacing human judgment. At least, not yet — and perhaps never fully. The most effective use cases pair AI capability with human oversight and domain expertise.
The Key Distinction: Narrow vs. General AI
Today's AI is narrow — even the most capable models have significant limitations. They're not persistent (no memory across conversations unless engineered), they can't take long-horizon actions autonomously without careful scaffolding, and they can fail on tasks that seem trivially simple.
General AI — systems with broad, flexible, human-like intelligence across all domains — remains a research goal, not a product you can use today.
Understanding this distinction matters for how you use AI tools: as powerful, fallible assistants that augment your thinking, not as oracles.
In the next lesson, we'll go inside the transformer architecture to understand how large language models actually work — and why that matters for using them effectively.