AI is more than ChatGPT

I recently wrote about how ChatGPT lit a fire of curiosity within me that I haven't experienced in years. (Obviously I'm not alone, since ChatGPT grew to 100 million users in two months, making it the fastest growing consumer app in history.)

After a few months of playing with it, I started talking about how "AI is going to change X" or "AI is going to be huge for Y," etc. I'm not alone there either, as everyone who experiences ChatGPT starts to form opinions about what AI can or cannot do, or how AI will or will not change the world. But forming opinions on "AI" based on our experience with ChatGPT is dangerous, because:

ChatGPT is not the same as "AI".

Rather, ChatGPT is one specific implementation of one specific type of AI. But the overall AI universe is much, much bigger. To understand the relationship of "ChatGPT" to "AI", here's the path starting from the highest level universe of AI, drilling down through the layers to ChatGPT:

1. Artificial Intelligence (AI) is an academic discipline which involves machines mimicking cognitive functions.
2. Machine Learning (ML) is a subset of AI where systems learn patterns from data.
3. Generative Models are a type of ML that generate new outputs from given inputs.
4. Large Language Models (LLM) are a type of generative model specifically designed for text data.
5. Autoregressive Models are a type of LLM that generate sequences (e.g. words and sentences) one element at a time.
6. Generated Pre-trained Transformers (GPT) are a type of autoregressive model developed by OpenAI. (Generative: model generates output. Pre-trained: model was trained on a huge amount of text before it can be used. Transformer: model processes sequential data by paying attention to the different elements to generate an output.)
7. ChatGPT is a specific application of a GPT which interacts with users in a conversational way.

In other words, from broad to narrow:

AI → ML → Generative Models → LLMs → Autoregressive Models → GPTs → ChatGPT

I put together the following diagram (myself, AIL 0) to show this conceptual model. Note that the circles are not in scale relative to each other, rather they're just showing how the various technologies relate to each other:

I should note that this progression is my own interpretation and representation. The boundaries and layers are fuzzy, especially when it comes to the specific implementation and application. But the conceptual mental model applies regardless of the labels and boundaries.

Fun sidebar: Maybe this is obvious, but the deeper you go, the newer the technologies are. The modern concept of AI is from the 1950s, ML is from the 1960s, generative models are from the 2000s, GPTs are from 2018, ChatGPT is from 2022, etc.

Dive deeper: There are lots of GPTs too!

The diagram above is obviously oversimplified and doesn't include all the other models and technologies at each layer that aren't in the drill-down to GPT. But even the "GPT" portion is simplified too. For example, OpenAI has lots of GPT models. Their famous app, ChatGPT, uses either the gpt-3.5 model (for free users) or gpt-4 (for paying users). But they have lots of other models. Some are constantly upgraded, others are "locked" (like the long term support versions of enterprise software). Some are tuned to be faster, or to support longer inputs, or more complicated inputs, or can run longer, or... (seriously, lots and lots of options, just like VMs on a cloud provider. It's the AI version of that balance of speed, power, and cost. More to write about this in the future!)

If we zoom further into the diagram from above, we can show the various GPT models. The drawing below shows this, though it's only showing 5 of the dozens of GPT models OpenAI has created. (Also I added some more generative models, just to show there are others at that layer too, though those shown are just a subset of all the types of generative models that exist.)

There are lots of LLMs besides the GPTs

And of course, the GPT models from OpenAI are only a small subset of all the LLMs in the world. You may have heard of Google Bard, which uses PaLM 2, Meta's LLaMA, Anthropic's Claude, etc. All LLMs have the same speed/power/complexity/cost tradeoffs as the GPTs of course.

Here's a chart from earlier this year which shows a bunch of different LLMs, along with their relative sizes to each other:

So what? What's my takeaway?

Even though some of this might seem like trivia, there are a few keys to remember that will make you smarter about AI:

• Despite the crazy popularity of ChatGPT, it's just one specific application of one specific type of AI. The overall AI world is much broader and diverse.
• Every AI "layer" covered here is also much bigger and includes more than we was covered in the drill-down to ChatGPT.
• When you think about "AI", don't form your opinions about it based on using one model from one vendor. Think of AI like a toolbox. GPTs are one type of tool (like a hammer), and ChatGPT is one type of hammer. There are lots of tools out there. Let's build something great!