Cookbooks are guaranteed bestsellers for the book trade (that’s what the bookseller I trust at the Holzbaum bookstore in Vienna said). They always sell and are often bought as gifts. But what exactly, on closer inspection, are cookbooks?
It all starts with the idea of a cook who wants to write down the recipes for her favorite dishes and publish them in bound form with appetizing pictures. To do this, she buys ingredients, rents a professional kitchen or an idyllic one in the countryside, hires a photographer, enlists a few other chefs for advice, gets a secret recipe or two from her mother, and starts cooking. Recipes are recreated several times, varying the amounts of ingredients and the cooking or baking time, until the perfect meal lands on the plate. The neatly noted recipes are accompanied by photos of the delicious dishes and go to print.
German Treats for the Hungry, 40 recipes with Swiss cheese, Chef Cuisine For Fifi, Cooking with Tic Tac. These and similar are the titles of the cookbooks.
In bookstores you can buy these masterpieces, but they come only with the instructions, the ingredients are not included. And if you want to try Korean food instead of European cuisine, but can’t find a cookbook on the subject at your local bookseller, you’ll have to gather the information yourself from various sources on the Internet.
If you are now hungry and wondering what this has to do with artificial intelligence, I would ask you to get a snack and be patient. This is an analogy to make it easier to understand a Generative Pre-trained Transformer, or GPT for short. GPT is a software fed with large amounts of data and trained, thanks to human help, to reassemble this data and generate, for example, texts, images, videos, or music. It has been pre-trained, transforms the data and generates something new.
Such an AI is like a cookbook. It contains conversion instructions and parameters, which are recipes comparable to the mixing ratios and cooking instructions. Someone has ‘pre-cooked’ the recipes for us before, that is, ‘pre-trained’ us, and relieved us of the work of creating the recipes ourselves.
I can then use a GPT in two ways: either I use an app or a web page in the browser, or I download the model and install it locally on my servers or in the cloud. The former would be like me going to a restaurant and the chef preparing it for me, and with the latter it would be like me buying the cookbook and recreating the recipes at home.
I can’t always just use an AI like that for myself. If I have special data that I want to use in such a model, then I have to do a so-called ‘finetuning‘, i.e. make adjustments. In the cookbook, I may have to change some ingredients because they are not available locally or I and my guests have intolerances or special dietary needs. Or I just add recipes to South Korean dishes.
The chef in AI is called a ‘transformer’, which is a special neural network. A GPT not only adds one word (like Google search) or pixel at a time to a sentence or image, but holds a larger context over multiple sentences or even pages of text or the entire image to be generated. Just as a chef must keep track of the entire menu for the evening.
This is a small excerpt from my book Creative Intelligence: How ChatGPT and Co will change the world, which will be published in the fall. It can already be pre-ordered here.