ChatGPT and comparison to Watson Assistant

I started working with Virtual Agent solutions in 2017. Since then I had my head buried deep in virtual agent innovation as well as integration.

Because OpenAI's ChatGPT is such a hot topic right now, it's a must to check it out and understand the difference between ChatGPT and existing technologies like Watson Assistant (similarly Lex, DialogFlow, Luis).

?

Let's start on what ChatGPT is. It is an advanced model trained with data from all over the public internet. Training data contains billions lines of code from public GitHub and human text from the rest of public internet. It has an API so you could utilise this model. You can set probability of the responses. Meaning highest probability answer will be more deterministic and repeated. Lowering the probability the model will take more risks and will give more random answers.

You can send a "prompt" which is basically an utterance and get response.

You can also send an instruction and an input on which you want the instruction to be applied. Something like asking to fix grammar mistakes on text provided and so on. Or you can provide a text and ask to explain it to a small child. Asking to explain a complex physics definition like?Schr?dinger's?equation?in quantum mechanics?to a 2nd?grader is fun!

?

Main set of features:

• 4 models of different capabilities, speed and pricing;
• Moderation endpoint. This model is trained only by OpenAI, no custom training available;
• Text completion and editing with instructions;
• Code completion and editing with instructions;
• Image generation, editing and creation of image variations;
• Fine-tuning endpoint to add custom training data to the existing ChatGPT model;
• Embeddings that allow to measure relatedness of text. Might be commonly used for: classification, topic clustering, search, recommendations.

?

ChatGPT versus Watson Assistant??

What you really should say is apples versus oranges. But we can try to compare between the two. Okay, let's try putting them side by side.

But what really does ChatGPT do?

Well, if you already know what your client wants, where is he coming from, you don't really need a complicated AI or an AI at all. Bunch of buttons will do the job well.

You need Watson Assistant if you prefer free text and you also have 10-20 or more directions that user might go from the start.

Now the ChatGPT could try answering to every question provided. But the results might not be what you want.

?

So ChatGPT is good for general text understanding and it got most excitement from developers for code generation. Let's look at this example:

I provide the text as "prompt" for ChatGPT:

?

"Write a program that simulates John Conway's Game of Life."

?

ChatGPT:

Using the pattern or context you gave it, it will:

Create a program. Give you the code.

?

I again provide the instruction on the previous context. "Refactor it". Or "Use this library instead". Or "Write documentation". Or "Create unit tests".

It is important to know that code needs to be checked as it might be wrong. Documentation and tests too. But it's a pretty good starting point.

You can also ask to write a harmful code that is malware. But that is the responsibility of moderation model.

?

It works best if you provide a pattern you want ChatGPT to complete.

?

"Decide whether a Tweet's sentiment is positive, neutral, or negative.

?

Tweet: I loved the new Batman movie!

Sentiment:"

?

ChatGPT actually has an understanding of what sentiment is and the concept of a tweet. But if you asked something that it does not know, the example in this case will help.

??

Where can it be used in customer support, containment, etc.?

Well, it is not replacing Watson Assistant. But maybe it could prove useful if used to understand sentiment based on a client's feedback. Or other use cases could be thought of, but for now it does not look a like good idea to let the user go wander around the wildlife that is World Wide Web encyclopaedia in a chat. So I guess there is a short-lived excitement that everyone might experience.

?

You can find more about what can ChatGPT do in these examples:?https://beta.openai.com/examples

??

ChatGPT API is not free, but you can register and get $18 in free credit as of writing this article.

The opinions stated here are my own. I do not speak for my primary workplace (IBM).