AI and Machine Learning
Everybody Wants to ChatGPT the World
It feels like we’ve been here before.
Just one year ago I wrote about how hype over AI in marketing had been overinflated by a bubble of exuberance. Nonetheless, there were meaningful and valuable AI use cases to leverage in marketing (which, incidentally, is why I have a job).
Now here we are one year later, talking about a new AI technology. This time, it’s generative AI that can create written and image content similar to that created by humans. Even though the hype has reached a crescendo just recently, this technology isn’t brand new: In fact, we released an AI Copywriting Assistant powered by GPT in May 2022 and an AI Image Generator powered by DALL-E 2 in December. So what is it this time? Just hype… or a revolution?
What Is Generative AI?
First, let's consider OpenAI’s GPT, a large language model that sparked the recent craze, as well as its offspring, specialized for realistic conversations, ChatGPT. It seems to do a lot of magical things and, as a result, a lot of clickbait has been written about how it can pass a law exam, supplant engineers, and write essays.. When considering the power of these tools it is important that note ChatGPT is not anything like a “general” intelligence or sentient AI that can grow up, go to college, and major in pre-med before finding its true calling to become a world-dominating supervillain.
As this excellent overview makes clear, GPT and large language models in general don’t actually “know” anything in any meaningful sense when it comes to the world, writing code, or history. They have not been engineered to use reason or logic of any kind or to capture the structure of the world and what goes on in it. For all intents and purposes, it’s a very elaborate autocomplete. It’s the same type of thing you get when you type “the world’s best” into Google search and it suggests completing the phrase with “chocolate,” “cat litter,” “soccer player,” or “lasagna.” All good options (that may or may not relate to my personal Google history).
But GPT does this in a much more sophisticated way. And it has been trained on huge chunks of natural language provided by the internet and sites like Wikipedia. You know those word clouds with bubbles that show how frequently words appear in a given body of text? Imagine having one of those clouds for nearly every term in the English language. What you’d get is a strong sense of what words tend to go together. So, if you start with a term like “Hedy Lamarr” and generate all the most likely words that follow, you’ll end up with snippets of phrases that tell a story about her life. By seeing millions of examples of natural language, GPT is able to generate output about related terms in a way that strings them together with grammatically correct pieces threaded in between.
Weighing the Pros and Cons of Generative AI
There are two critical consequences of GPT’s capabilities. First, with just one click, we can use GPT to generate a wealth of copy that seems like it was written by a human, meaning we can use it to add endless variety and spice to writing in an unparalleled way. Take it from Raffaella Accogli, Global CRM Manager, Activation at The Fork, who started using our AI Copywriting Assistant powered by GPT soon after it was released last spring:
“I felt like I always used the same words, the same expressions, the same concept. With the copywriting assistant, I can find a way to talk about the same concept using different words. And, it provides inspiration to develop new approaches.”
In the hands of marketers and creatives, this unlocks a whole new way to change up their content to keep things fresh and vibrant across their marketing efforts. By including controls for length and output language, GPT also makes it possible to adjust the content for the necessary language and message format. Live translation right in the dashboard!
However, the second critical implication is an important one—namely, that there’s no guarantee that what GPT says is factually correct. For one thing, the information and associations that the algorithm learned from the training data may contain factually incorrect statements. The randomized combinations of the word associations generated by the algorithm from that data may also not be factually accurate. As these large language models get even larger (have more “parameters” in the jargon), they will probably not create a model of how the world works and check against it for accuracy. They’ll just remember more sequences of associations between more terms. They won’t gain any facility in discerning facts from fiction. This cuts away a lot of potential applications where being incorrect is unacceptable. That’s why our generative AI tools are assistants – there’s still a very human need for judging when and how to leverage AI.
How Generative AI Is Evolving
ChatGPT, the successor to GPT, has been trained on those sources as well as refined on actual human conversations. Refining the extracted associations on real human conversations results in convincing dialogue as output. So while it does not “know” anything, ChatGPT can still provide a good simulation of naturalistic conversation. For example, it’s quite adept at following instructions to alter its responses based on context. See below for an example: It can appear to reason and take into account additional terms and statements introduced into the dialogue in an extremely human-like way. But, under the hood, it’s not “reasoning” in the way we would normally think of it. But it does create a very powerful illusion that it is!
What Is AI Image Generation?
Over the summer, a cadre of new image generation models also came on the scene. These truly impressive tools can synthesize images from simple textual descriptions to create entirely new images with ease. If you haven’t tried it out yet, hop over to the Braze dashboard and open up the AI Image Generator to give it a whirl. (Or, if you’re not yet a Braze customer, you can sign up for an account at OpenAI’s website.) Quite simply, it’s delightful.
Very specific descriptions of figures, actions, and style can yield impressive results. This tool allows marketers to enhance messaging with creative content with incredible ease and speed. Instead of shopping around for stock images or waiting for creative requests to be fulfilled, marketers can generate custom images within seconds. Different subjects and styles can be produced with stunning quality that would take a highly specialized and talented artist quite some time to create. Plus, the process of fiddling with the prompts to get just what you want can be fun in and of itself. Here are a few examples of Braze customer creations:
Now, even DALL-E is not perfect. It struggles with hands and faces, and it also created its own gibberish language that pops up often in place of text. But, we look forward to how it will evolve and what will become possible going forward.
The Future of Generative AI
Generative AI is evolving quickly, and we are just at the start of what this technology can do. Braze is evolving quickly, too, to bring the power of this technology into our customer engagement platform: The recently released GPT4 model will be powering the Braze platform’s AI Copywriting Assistant as soon as possible! As a result, it will deliver even higher-quality copy straight into your messages. And because this is a fast-moving space, Braze plans to quickly incorporate future GPT models as they’re available.
Keep your eyes peeled for more updates and improvements to this cutting-edge capability. If you have suggestions, let us know in our product roadmap portal. Meanwhile, I’m looking forward to advocating that we not freak out about next year’s new AI craze: A conversational home vacuum that criticizes your poor dietary and life choices by analyzing dust picked up from your floors. See you next time!