When I first started looking to use GPT-3 as part of Aampe, I was skeptical. Two reasons:

1. Asking GPT-3 to write your customer communication is like asking Twitter to write your customer communication.

Most of the time, it'll be fine. Sometimes it'll be surprisingly good - maybe even brilliant every once in a while. But sometimes it'll be pretty weird - and, every once in a while, it'll be offensive.

2. All models like GPT-3 suffer from "drift".

If you write five sentences, sentence 2 will follow reasonably well from sentence 1, and sentence 5 will follow reasonably well from sentence 4, but sentences 1 and 5 don't make any sense in the same paragraph. It's kind of like that game of "telephone" many of us played as kids - each step introduces a little distortion, and it builds up.

Both those problems together are a big problem for any business - your communication with your customers is too valuable to risk nonsense, or worse.

Then I realized that we'd already built what we needed for GPT-3 to work for our customers:

◼ Aampe lets you build meaningful pieces of a larger message, then mix and match those pieces to scale.

◼ GPT-3 does really well at individual sentences. Amazingly well. Like...it constantly surprises me just how good it is at it.

So combine the two:

1. Have GPT-3 write ten different ways of saying you sell clothes, or food, or hotel bookings.

2. Pick 5 different reasons people should want what you’re offering (for travel: are they looking for an escape? or for a cultural experience?) and have GPT-3 write 10 different ways of stating each value proposition.

3. Then have it restate the 10 offering statements and the 5 * 10 = 50 value propositions in 5 different tones (urgent, sophisticated, inspirational, easygoing, etc.)

GPT-3 generates the messages. Aampe labels them and distributes them intelligently.

👉 That's 10 + (5 * 10) = 60 pieces you need to enter.
👉 That's 60 * 5 = 300 one-sentence tonal variations you need to review to make sure it's on brand and not confusing.
👉 That's 10 * (5 * 10) * 5 = 2500 unique sendable message variants you can send. (5000 if you create variations that switch up which sentence appears in the notification title or email subject line.)

It adds up quickly. And that's assuming you only sell one thing. Most of our customers do a lot more than one thing.

Aampe + GPT-3 allows our customers to write copy at scale while still ensuring quality and brand consistency.

And all those messages are tagged, so you can learn which customers are looking for escapes, and which are looking for cultural experiences. You can learn which prefer a friendly, easygoing voice to their messages, where others prefer a just-the-facts tone. Aampe already does all that learning (plus learning when users prefer to be contacted).

GPT-3 is allowing us to do that learning fast, at a larger scale, with more consistently labeled data, and extremely little effort on the part of our customers.