A hitchhikers guide to GPT-3. Part 3

What is GPT-3 good at?

In the first two parts, we looked at some fun stuff using GPT-3. But what is GPT-3 really good at? We know it’s good at NLG. But can we use it’s features to improve some business aspects. Can it be applied in some products?

So, in this blog post we will look at some experiments which I did to see where we can fit in GPT-3 in production systems. GPT-3 is really good at slot filling. It can take natural language English text and output json as we saw in our movie database questions. Expanding on that I did a few experiments.

Experiment 5:

Title: Build charts from English text:

Result:

nutanc @nutanc

Starting the day with a chart building demo. Primed GPT-3 with Chartjs scripts to generate the below.

Analysis:

It works really well with very little priming. Others have show similar results with other Javascript libraries like plotly. I have also checked it with Highcartsjs and other libraries. It generally performs pretty well.

Experiment 6:

Title: Text to Google slide generator.

Result:

nutanc @nutanc

Letting GPT-3 do my presentations from now on. Just copy a bunch of text and let GPT-3 generate the presentation. Using Google slides scripting. I am not good at presentation skills. This is for proof of concept. So be gentle :). cc @gdb

Analysis:

As expected, this too performs well. This whole demo was built in a couple of hours which captures the power of all the APIs we have access to.

Experiment 7:

Title: Text to IVR

Result:

nutanc @nutanc

One of the first innovation we are doing is to rethink the age old process of IVR design. Make the boring IVR development exciting. Just tell the flow of your IVR in natural language and get a phone number with the IVR flow designed. Prepare to have your MIND BLOWN :) cc @gdb

Analysis:

Surprised to see it perform well. It could not directly learn our KOOKOO markup up language directly. So I used an intermediate JSON representation to make it work. But mostly it works real well and this could be our first production deployment on GPT-3 in one of our products.

Miscellaneous:

I did multiple other experiments of mashups and in most cases GPT-3 performs well enough.

That’s it for the experimentations now. In the next post I will be going more in depth about the ML piece and the philosophy.