5 things to know about the hottest new trend in AI: Basic Models

If you’ve seen tea pot photos like Avocado or read a well-written article that leads to some weird tantrums, you’ve probably come across a new trend in artificial intelligence (AI).

Machine learning systems called DALL-E, GPT and PaLM splash with their incredible ability to create creative work.

These systems are known as “basic models” and are not all hip and party moves. So how does this new approach to AI work? And will this be the end of human creativity and the beginning of a deep false dream?

1. What are the basic models?

Basic models work by training a large amount of general information in a single large system, then applying the system to new problems. The previous models wanted to start from scratch for each new problem.

DALL-E 2, for example, has been trained to scan hundreds of millions of instances with images (such as a photo of a pet cat) with the caption (“Mr. Fawzy’s boots relaxing in the sun”). Once trained, this model knows what cats (and other things) look like in photos.

But the model can also be used for many other interesting AI tasks, such as just creating new images from the title (“Show me a Koala Dunkin ‘basketball”) or editing images based on written instructions (” It looks like a sack that encloses with a drawstring.

2. How do they work?

The basic models run on “deep neural networks”, which are gently inspired by how the brain works. It involves complex mathematics and a large amount of computer power, but they fit into very complex types of patterns.

For example, by looking at pictures of millions of instances, the deep neural network can connect the word “cat” to pixel patterns that often appear in cats’ photographs – such as soft, fuzzy, hairy bulbs of texture. The more examples the model sees (the more data is displayed), and the larger the model (the more “layers” or “depths”), the more complex these patterns and relationships can be.

The foundation models are in a sense extending the concept of “deep learning” that has dominated AI research for the past decade. However, they show unplanned or “urgent” behaviors that can be both shocking and novel.

For example, Google’s PaLM language model seems to produce explanations for complex metaphors and jokes. It simply goes beyond mimicking the data types that it was originally trained to process.