Language models like GPT-3, OPT, BERT, and BlenderBot have changed the machine learning and application development landscape. Today, we can build applications in a natural, user-friendly manner like never before. Unfortunately, language models don't always get it right. It's been well documented that language models are capable of biased responses that can be harmful if not tracked correctly. In light of this, many companies have implemented something called a toxicity filter for their respective services.

Our project outline is simple: we're going to build a chatbot for a fictional storefront that sells ice cream. Side note: While this is simple in nature and somewhat of a fun example, this project outline could be used in many applications. Fusing together visual and conversational abilities into a single platform could help automate a number of different tasks, like automated customer service. To continue with the rest of this tutorial, let's install the Humingbird package with the command: First, we need to start by building an intent recognition system. For those not familiar, intent recognition is the task of predicting what a query "means".