This is the OpenAI API call for embeddings—you send it text, it returns those floating point numbers. It’s incredibly cheap. Embedding everything on my site—400,000 tokens, which is about 300,000 words or the length of two novels—cost me 4 cents.

The trick instead is to take the user’s question, search for relevant documents using a regular search engine or a fancy vector search engine, pull back as much relevant information as will fit into that 4,000 or 8,000 token limit, add the user’s question at the bottom and ask the language model to reply.

A simple Python implementation of the ReAct pattern for LLMs.

It’s called the reAct paper, and it describes another one of these prompt engineering tricks. You tell a language model that it has the ability to run tools, like a Google search, or to use a calculator. If it wants to run them, it says what it needs and then stops. Then your code runs that tool and pastes the result back into the model for it to continue processing.

What this adds up to is that these language models make me more ambitious with the projects that I’m willing to take on. It used to be that I’d think of a project and think, “You know, that’s going to take me two or three hours of figuring out, and I haven’t got two or three hours, and so I just won’t do that.” But now I can think, “Okay, but if ChatGPT figures out some of the details for me, maybe it can do it in half an hour. And if I can do it in half an hour, I can justify it.”

There’s this idea of T-shaped people: having a bunch of general knowledge and then deep expertise in a single thing. The upgrade from that is when you’re pi-shaped (actually a real term)—you have expertise in two areas. I think language models give us all the opportunity to become comb-shaped. We can pick a whole bunch of different things and accelerate our understanding of them using these tools to the point that, while we may not be experts, we can act like experts.

LLMs have started to make me redefine what I consider to be expertise. I’ve been using Git for 15 years, but I couldn’t tell you what most of the options in Git do. I always felt like that meant I was just a Git user, but nowhere near being a Git expert. Now I use sophisticated Git options all the time, because ChatGPT knows them and I can prompt it to tell me what to do.

We call on the field to recognize that applications that aim to believably mimic humans bring risk of extreme harm. Work on synthetic human behavior is a bright line in ethical AI development. This has been ignored by essentially everyone! These chatbots are imitating humans, using “I” pronouns, even talking about their opinions.

Code-wise, I will never commit code if I can’t both understand and explain every line of the code that I’m committing. Occasionally, it’ll spit out quite a detailed solution to a coding problem I have that clearly works because I can run the code. But I won’t commit that code until I’ve at least broken it down and made sure that I fully understand it and could explain it to somebody else.

I read academic papers now. I never used to because I found them so infuriating—because they would throw 15 pieces of jargon at you that you didn’t understand and you’d have do half an hour background reading just to be able to understand them.

The question I always ask myself is: Could my friend who just read the Wikipedia article about this answer my question about this topic? All of these models been trained on Wikipedia, plus Wikipedia represents a sort of baseline of a level of knowledge which is widely enough agreed upon around the world that the model has probably seen enough things that agree that it’ll be able to answer those questions.

The obvious question then is how on earth do these things even work? Genuinely all these things are doing is predicting the next word in the sentence. That’s the whole trick. If you’ve used an iPhone keyboard, you’ve seen this. I type “I enjoy eating,” and my iPhone suggests that the next word I might want to enter is “breakfast”.

There’s a thing called the system prompt, where you can provide an additiona prompt that tells it how it should behave. I can run the same prompt with a system prompt that says “You are a poet”—and it writes a poem!

This paper showed that if you give a logic puzzle to a language model, it gets it wrong. But if you give it the same puzzle and then say, “let’s think step by step”, it’ll get it right. Because it will think out loud, and get to the right answer way more often.

They were plotting to rob Squire Trueman,

They were plotting to rob Squire Trueman

“I like to have faith that mankind still believes,” Gabriel said. “Not the ones you want to believe,” Seppo muttered.