LESSON 1

Complete this sentence:

For example, open a ChatGPT window and type the first sentence of a story.

Press the play (►) icon below to see what happens...

How do these systems do all the amazing things they do?

And why have then been developed just recently, rather than twenty years ago, or twenty years in the future?

The key ingredients seem to include the following:

• a novel algorithmic architecture known as a transformer, developed in 2017.
• the massive amounts of training data—information used to teach the machine to predict what will come next in a stream of text—that can now be scraped from the internet. Current models are thought to be trained on terabytes of text and code.
• The process of reinforcement learning from human feedback, whereby an LLM interacts with human trainers to "learn" what humans are seeking when they query the model and what constitutes a good answer to a question.
• the scale of contemporary computational power, which only recently became great enough to train a model of appropriate scale.

These developments, coupled with the weirdness of the trillion-dimensional spaces in which these algorithms operate, the remarkable utility of machine learning, and the surprising predictability of most linguistic activity, made it possible to create the powerful LLMs that we use today.

By thinking about what LLMs do, we can better understand what they don’t do. They don’t engage in logical reasoning. They don’t have any sort of embodied understanding of the world. They don’t even have a fundamental sense of truth and falsehood.

They simply generate plausible dialogue that resembles the text they were trained on.

That’s a key point. LLMs generate text based on their training data. Ask an LLM a question. It doesn’t ‘think’ about your question the way a person does. It tries to guess what words would have come next if your question had been in its training data. The wonder of LLMs is how well this simple trick approximates a thoughtful response.