Straive’s Post

Transcript

There is another way of using these models and that happens to be using embeddings. What embeddings tool is convert text to numbers? So each of these little dots, for instance, is a piece of text. So this is MA, this is homepage, this is Dewey, and they have been converted into numbers and they've been shortened to just three numbers so that you can view them on the XY&Z axis as a point in space. Now the interesting thing about these numbers is that similar numbers mean similar words. So if I, for example, search for the word. Singapore. And look at what comes, uh, closest to Singapore in space. Now, the closest words to Singapore you'll also see on this list here happen to be Malaysia, Philippines, Hong, Taiwan. It's not selecting the words. It is just discovering the words that happen to be closest in this multidimensional space and automatically when we select Singapore, Malaysia, Philippines. Hong and Kong, Taiwan, Indonesia happened to pop up as having the nearest, uh, the lowest distance or the highest similarity. If that is the case, what it means is that we can use these embeddings to identify similarity and not just similarity in a vague way, but in a quantitative way. We can say that the distance between Malaysia and Singapore is .38, which is lower than that of Philippines, which is even lower than that of Australia. And the distances may be a factor of 2 lower for Malaysia than for Australia. That's sort of quantitative computation becomes possible, making the reliability a little bit. Higher than just with chat where you get an answer, but not necessarily with clarity on why that is the answer. The other interesting property is that these embeddings are even cheaper. For example, plot 3 high Q which we saw was pretty cheap, cost only $0.25 for a million tokens. But the same thing with open AI's text embedding three small would have cost only two sets. And if you were running Baidu A Beige or Microsoft's E5 or any one of the open open source text embedding, open wait text embedding models, you can run this on your GPU machine or even non GPU machine. Practically at no cost more than electricity. You could run this on Google Colab. You could run this on any environment. You don't need to provision something separately, which means that the cost is almost zero. And even if you didn't and just used an existing API, that $10, 000 budget would shrink to just $500.