I'll take an initial stab:
- For the "AI" part:
- Makes use of machine learning techniques (probably gradient descent on a neural network) to model the patterns/structures of a given data set
- For the "generative" part:
- Outputs new content in a relatively open-ended domain (such as text, images, or audio)
- Task's output is underdetermined by input, such that there'd be a wide range of acceptable outputs rather than one/few "correct" results
Some examples, including tools mentioned:
Tool |
GenAI? |
Reasoning |
GPT-3 |
? |
AI: trained on web text. Generative: text is a broad domain, output of text continuation task is not fully defined by input text |
Stable Diffusion |
? |
AI: trained on web images. Generative: images are a broad domain, input prompt doesn't fully define a generated image |
Gaussian filter |
? |
Not AI: not trained on data. Not generative: output domain is broad, but output is fully defined by user's input |
Github Copilot |
? |
AI: trained on web text, then public Github repos. Generative: code is relatively open domain, and there are many ways to approach a given problem |
Scigen |
? |
Not AI: uses a hand-written context-free grammar rather than machine learning from data |
Clippy |
? |
Not AI: probably hand-tuned criteria. Not generative: Selects from a set of pre-defined suggestions, rather than an open domain |
Cleverbot |
? |
Not generative: assuming output is selected from previous user responses, rather than new text content |
Google Translate |
?? |
Maybe not generative: output could be considered too tightly defined by the input text |