With companies like Microsoft investing over $10 billion into chat GPT, it kind of begs the question

how valuable is one of these large language learning models.

So today on the podcast, we're going to talk about something very interesting that recently

came out, which is the fact that some researchers at Stanford said that they have essentially been

able with some new technology that came out to clone chat GPT for less than $600.

Now, when we're looking at, you know, Microsoft, you know, in 2017 investing a billion dollars,

getting this kind of technology kicked off, and then just recently investing $10 billion into

the company to further

its efforts. It's pretty interesting that, you know, people are claiming they're able to duplicate

this for $600. So on the podcast today, we're going to dive into what exactly happened and what

this is going to mean for AI in the future. So first off, Stanford is calling it Alpaca AI.

And essentially, what they did was Facebook recently, or meta, they recently came out with what's called Lama model. So essentially it's Facebook's like open source language model. It's pretty much the smallest and cheapest of all the Lama models available.

And it's pre-trained on a trillion quote-unquote tokens, which tokens is not necessarily words.

It's kind of like bits of words or half. I don't know. It's just like how AI models measure what is being spit out and what it was trained on. So like one long

word might be like three tokens and like a short word might be one token. Anyways, not that important.

So pretty much this model has a certain amount of capacity baked into it, but it would leg

significantly behind chat GPT with most tasks. And the number one area is cost.

So the biggest competitive advantage in the GPT model comes largely just from its

the enormous amount of time and manpower that OpenA has put into the post training.

So it's one of the things, for example, it's just like one thing to have read a billion books,

but it's another to have, you know, chewed through large quantities of questions and answers,

conversion pairs that teach an AI what their actual job will be, right?

...