Presented by:

Ken Kahn

from University of Oxford

One example of many. Here spoken commands become movement scripts
Open in a new tab

Using the web services library one can create blocks that send prompts to GPT-3, GPT-4, Cohere, Jurrassic 1, and other large language models. These blocks report the "completions" returned by these API calls. I'll present five sample projects using these blocks:

Conversations with and between personas using language models

Using Large Language Models to simulate debates

The Automated Generation of Illustrated Stories

Using large language models to control a Logo turtle

Demonstrating that GPT-3 can play Tic Tac Toe

Unlike using a large language model "playground" where you can enter text and receive completions, a programmatic use of large language models enables one to integrate the completions into a larger task. For example, the language model can be asked to generate a story and then asked to generate suggestions for illustration prompts for each paragraph and then contact a text-to-image model to receive costumes to use in an illustrated version of the story. And the story and illustrations can be co-developed by the user and the model. For example, the model can be prompted to suggest changes to a story, the user can approve or reject, and then the changes are implemented by another request to the model.

Here is a Snap! project for working with language including large language models.

This abstract is an abbreviated version of this brief introduction to using large language models in Snap!.

20 min
Auditorium (Online)
Snap!Con 2023
Presented via: