Mind Readings: Generative AI as a Programming Language

In today's episode, let's talk about a quote that I read last year that I loved from Andre Karpathi, who's a co-founder of OpenAI. He said in January of 2023, the hottest programming language in 2023 is plain English. This is specific to generative AI, large language models, to be really specific. And Karpathi, co-founder of OpenAI, worked at Tesla and built some of the vision software there. Really one of the pioneers in the space.

If you watch his YouTube channel, you can see him like, hey, let's build a language model from scratch. You're like, that is that's like you know, watching something, let's build an automobile from scratch. And they go out in the back over there, Forge and start forging pieces of metal. Like, oh, this this dude knows how these things work. This quote has stuck with me over time because of what it implicitly says, what it teaches us about using generative AI, using large language models, what capabilities it unlocks.

And here's why. We think correctly most of the time of large language models, like the ones that power Chat GPT and other tools as conversational models. We talk to them like people, right? We we ask questions, we get specific, we show examples, we do all sorts of things that you and I naturally do as humans. I I say in a lot of my keynotes um treat this like the world's smartest intern, right?

This intern's got 255 PhDs, really smart intern, but they don't know your company. So you still have to explain things, you have to still have to show examples, you have to still have a conversation with them. And that conversational style of working with language models works. Models understand that, they get that. They work with that because that's what they're trained on, vast amounts of human language.

And they're they're pretty good at that. That's why so many of us use these language models. But they're also trained on vast data sets of real programming languages, C and Python, Java, PHP, R, COBOL, right, DM Algol 2, like old mainframe code. They speak these computer languages exceptionally well because computer languages are unambiguous. Humans, humans speak and write with a great deal of ambiguity.

Machines don't do that. Machines, you uh you get creative in your use of language, and your code just doesn't work. It just you just get errors. Um, there's only so many ways you can word things and stay within the syntax of computer languages, and that's why that's why language models are so good at them, because they can produce results that are highly predictable. You know that there's not going to be a character after the left arrow in an R variable sign.

There just can't be. The statement will not work, the code will break. So no one does that. There are no examples of that that is learned, uh, and and things work just fine. But humans, yeah, we we have so much ambiguity in our language, especially in tonal languages, but even in non-tonal languages.

English is largely regarded as a non-tonal language, where it means that the the way you say something influences the meaning of it. Um English is technically non-tonal, but you can come up with different meanings in a sentence, just how you based on how you say emphasize things, right? You can say that's what she said, implying it's about what it is the person said. You could say that's what she said, not why. That's what she said.

Oh, the person. That's that's you're talking about the person. That's what she said. Um sort of disbelief about the content. Saying that four different tones creates four different sentences.

And when a machine looks at that text, it's like, I don't know what that means. Again, programming languages don't have this problem. Programming languages do not have that level of ambiguity. They are much less forgiving of mistakes and bad grammar, they're more precise. So, where am I going with all this?

Well, let's go back to Carpathi's quote. Hottest programming language in 2023 is plain English. What if English was a programming language? Well, in a generative model, it is. You can do things like control structures, if-then statements, while while statements and for loops.

Familiar friends from computer programming languages used to say it in English instead of computer code. You know, if you're making a dad joke tool, you could say, if the user says this, do that. If the user doesn't include this, do this. Until the user says stop, do this. You just write out these instructions in plain language, maybe a formatum and do some indenting and things.

Do you see how powerful that is? You can make your prompts, your use of large language models incredibly powerful by adopting these programming structures, the logic of programming, into regular language. You don't have to write in computer code, but you're thinking in computer code. You're thinking, how do I set up this set of statements in plain English in a prompt? And because large language models understand both human language and computer language, they can operate human language formatted like computer programming languages to the same good effects.

You can program in human language. You can program in Danish, you can program in Japanese or Ukrainian by adopting the structures, the logic of computer programming into your language. This when I learned this was mind-blowing. It's mind-opening. You can program today's AI with the same programming skills used in computer languages, only with language you speak and write every single day.

And you will be amazed at how much more you can accomplish with AI. Because you're now bringing that structure and that logic to generative AI, to prompting, in language that everyone can understand. So give that some thought. That's going to do it for this episode. Thanks for tuning in.

I'll talk to you on the next one. If you enjoyed this video, please hit the like button. Subscribe to my channel if you haven't already. And if you want to know when new videos are available, hit the bell button to be notified as soon as new content is live.