"Although these transformer models were trained to predict the next word in a text, they do a surprisingly good job at other language tasks"... "it may be that transformer models and their relatives are learning latent representations that capture the same basic ideas as grammars and semantic information, or it may be that something entirely different is happening within these enormous models; we simply don't know."
Artificial Intelligence, A modern Approach, Stuart Russell and Peter Norvig
The most popular textbook on AI, used by more than 1,500 universities, can’t explain why large language models work so well.
These models were designed to predict the next word in a sentence. Then we fed them more data and trained them with faster computers. Now, they’re passing medical licensing exams.
That insight blew my mind.
The unexpected intelligence of large language models has changed how I’m approaching AI implementation. Instead of purely using first principles, like an engineer calculating the load on a bridge, I must be more experimental. Given we don’t fully understand how these models are so capable, it’s better to hypothesise their abilities and test to see if I’m right.
I was going to write “building with LLMs is more of an art than a science.” But that’s wrong. What I’ve described—making hypotheses, testing, refining—is literally the scientific method. So it’s more accurate to say, “building with LLMs is more of a science than engineering.”
That’s why it’s time for me to close the textbooks and start experimenting.
“The world was to me a secret which I desired to divine.”
— Mary Wollstonecraft Shelley, Frankenstein
What’s next:
Given the rapid pace of AI development and fire hose of new research, I’m going to narrow my focus to make progress mastering AI implementation. I’ll tackle one area at a time (e.g. prompting, agents, image generation).
I’ll focus the investigation of each theme around a project which will allow me to experiment and learn.
I’ll share learnings on Substack, but also keep my AI Implementation guide updated as a reference wiki.