Building Intuition: Generative AI and Predictive (“Traditional”) AI
TL;DR: generative AI produces new content in some form, predictive AI produces predictions about its input
2023 was undoubtedly THE year of generative AI, but what is generative AI? How is it different from what we more commonly know as AI? And since when did neural networks get termed “traditional” AI despite these techniques only being viable within the last 10 years? Let’s take a look at how all this happened.
What is Generative AI?
Generative AI consists of models focused on generating new output. Examples of techniques that are considered generative AI include generative adversarial networks (GANs), variational autoencoders (VAE), and transformer models.
Generative models are trained on data in an “unsupervised” manner. They learn the patterns and structures in their training data. When we interact with these models, they give us predictions based on the data they were trained on and the input we give.
Currently, the most common use cases for generative AI are chatbots through retrieval augmented generation (RAG) and image generation via diffusion models. We’ll touch more on RAG, which may use both traditional models and generative techniques below.
What is Predictive/Traditional AI?
Predictive AI has technically been around since the 1970s, but in 2024, when people say “traditional”* or “predictive” AI in the industry — they typically refer to the use of neural networks to drive predictions. Examples of what is referred to as predictive AI include recurrent neural networks, convolutional neural networks, and often ML techniques like K Nearest Neighbors, SVMs, and XGBoost.
Predictive AI is trained via “supervised” learning. Supervised learning requires labeled data. Each data point must be labeled with a “class” to predict on. These models learn patterns of how data is related to its labeled class.
The most common use cases for predictive AI include things like image segmentation, entity recognition/resolution, and predicting the weather. An example of predictive AI you’re probably familiar with is when you see videos getting recommended to you on YouTube or items getting recommended on Amazon.
Key Differences
Predictive AI:
- Learns patterns to relate input data to an output
- Primarily used to predict or automate processes
- Industry uses this to refer to pre-transformer neural network architectures
Generative AI:
- Learns patterns and structures within the input data
- Primarily used to “create content”
- Industry uses this to refer to things like GANs, VAEs, and transformers
Using Both — Retrieval Augmented Generation
RAG is a type of LLM app architecture that consists of four pieces:
- An embedding model
- An LLM
- A vector database
- Prompts to interface with the LLM
This is basically the new way to make chatbots. This is also what I’ve been primarily focused on building over the last few months. Here are some tutorials for the curious:
Like this article? Follow me, Yujian Tang, for more posts about NLP, Software, and Growth. Make sure to follow Plain Simple Software for more software articles too!
Thank you to my reviewers for this article — Kristen Kehrer, Adam DeJans Jr, Megan Lieu, Daliana Liu, and Jess Ramos.
*Note — this article offers a GENERALIZED view, and there is much more nuance once you dive deeper.
