Why This Time Is Different
Every decade or so, AI gets a moment. Researchers publish something impressive, funding pours in, magazines declare the future has arrived — and then it doesn't. The systems turn out to be brittle, the use cases narrow, the promises hollow. The pattern has happened enough times that "AI hype" is its own genre.
So it's a fair question: is this one of those moments? Or is something genuinely different?
The hype cycle is real
The history of AI is a history of overconfidence followed by disappointment.
Symbolic AI
Founders of the field predicted human-level intelligence within a generation. They built systems that could solve logic puzzles and play chess — and concluded, incorrectly, that general intelligence was just around the corner. The first AI winter followed.
Expert Systems
Rules-based software encoded human expertise in logic trees. Briefly celebrated, widely adopted — then brittle reality set in. Real expertise doesn't compress neatly into rules. A second winter followed.
Deep Learning
Neural networks trained on massive datasets outperformed everything before them on image recognition, speech, and translation. Real progress — but also real hype. Self-driving cars were 'five years away' for a decade.
Generative AI
Transformers at scale. ChatGPT. 100M users in two months. Something structurally different — or so it seems. The rest of this lesson makes the case.
What actually changed
The current wave isn't just bigger hype. Three things converged that hadn't before.
The transformer architecture. In 2017, a team at Google published a paper called "Attention Is All You Need." It introduced a new way of processing sequences of text — one that could handle much longer-range relationships between words, and that scaled far better with more data and compute. Every major language model today is built on this foundation.
Scale. It turns out that making neural networks bigger — more parameters, more training data, more compute — produces qualitatively different capabilities, not just quantitatively better ones. GPT-2, released in 2019, could write coherent paragraphs. GPT-4, released in 2023, can pass the bar exam, write production code, and reason through novel problems. Nobody fully predicted this. The field has a term for it: emergent capabilities — abilities that appear at scale without being explicitly trained.
Accessibility. Every previous breakthrough stayed inside labs and enterprise software. AlexNet won an image recognition competition in 2012; most people never heard of it. ChatGPT launched as a free website anyone could use. That's not a minor distribution detail — it meant hundreds of millions of people immediately discovered what the technology could and couldn't do.
The part that surprised even the researchers
Here's what makes this moment genuinely uncertain: the people who built these systems didn't fully predict what would emerge from scaling them up.
OpenAI's researchers were surprised by GPT-4's performance on standardized tests. Anthropic published research on emergent capabilities that appeared in larger models but not smaller ones, with no clear explanation for why. Demis Hassabis, CEO of Google DeepMind, has described being genuinely surprised by what large models can do.
When experts are surprised by their own tools, it's a signal that this technology is worth paying careful attention to — in both directions. The upside may be larger than expected. So may the downside.
The honest qualifier
None of this means every AI claim is true or that every predicted use case will pan out.
These models still make things up confidently. They struggle with precise arithmetic. They can be manipulated with surprisingly simple tricks. They encode the biases of their training data. Deploying them responsibly — in medicine, law, education — is genuinely hard, and the field is still figuring out how.
The goal of this course isn't to convince you that AI will change everything, or that it's overblown. It's to give you enough grounding to evaluate those claims yourself. That starts with understanding what these systems actually are — which is what the next module covers.