to scale up the use of large-language models in marketing

The marketing industry will be transformed by the use of large language models and AI.

Christopher Penn, Chief Data Scientist at TrustInsights.ai and speaker at the MarTech Conference, explained that to stay competitive you will need to understand technology and its impact on our marketing efforts.

Discover how to scale large language models, the importance of prompt engineering, and what marketers should be preparing for.

The concept behind large language models

Since its release, chatGPT is a hot topic across most industries. It’s impossible to go online and not see everyone’s opinion. Penn said that few people are familiar with the technology.

ChatGPT, an AI chatbot based off OpenAI’s GPT-4 and GPT-3.5 LLMs.

The LLM is based on the 1957 premise of English linguist John Rupert Firth, “You will know a language by its company.”

The meaning of a particular word can be determined by the words that are usually used alongside it. Words are not only defined by their dictionary definition, but also the context of use.

Understanding natural language processing begins with this premise.

Take a look at these sentences, for example:

• I’m making the tea.”
• I’m spilling tea.

The first is slang used to refer to a hot drink, while the second is slang that means gossiping. In both cases, “tea” has a very different meaning.

The order of words is important.

• I’m making the tea.
• “The tea that I’m brewing.”

Even though the verb “brewing” is used in all of these sentences, they have different subject matter.

What is the best way to use large language models?

Below is the system diagram for transformers, which is an architecture model on which large language models can be built.

A transformer is a device that takes an input (i.e. “transforms”) and makes it into another.

LLMs are good for creating, but they excel at transforming one thing into another.

OpenAI, along with other software companies, begin by ingesting a vast corpus of data. This includes millions of documents, academic articles, news articles and reviews, as well as forum comments and product reviews.

Imagine how often the phrase “I am brewing the coffee” appears in these texts.

Amazon product reviews, and Reddit comments are just two examples.

This phrase is “in good company”, that is all the words that appear near the phrase “I am brewing the tea.”

These LLMs are based on “taste,” “smell,” coffee, “aroma,” etc.

Machines can’t read. To process this text, the machines use embeddings. This is the first step of the transformer architecture.

The embedding technique allows models to assign a numerical value to each word, which is repeated in the corpus of text.

The word position is also important to these models.

In the above example, the numerical values are the same but in a different order. This is called positional encoding.

Large language models can be explained in simple terms as follows:

• Text data is taken by the machines.
• All words should be assigned a numerical value.
• Take a look at the frequency and distribution of the words.
• Try to guess what the next letter in the sequence is.

This requires a lot of computing power, resources and time.

Prompt engineering: A must-learn skill

The more context we give LLMs and the more instructions we give them, the better they are likely to perform. The value of prompt engineering is this.

Penn views prompts as guardrails to what machines will produce. The machines will latch on to the context of the words we input as they create the output.

You’ll find that when you write chatGPT prompts detailed instructions are more likely to get satisfactory results.

Prompts can be compared to creative briefs. You won’t just give your writer one line instructions if you want the project to be done right.

You’ll instead send them a brief that covers everything they need to know and the style you prefer.

LLMs can be used to scale up the use

You might think of AI chatbots as a web-based interface, where users enter questions and wait for a response. Everyone is used to this.

This is by no means the end of these tools. This is the play area. Penn said, “This is the playground where humans can play with the tool.” “This isn’t how businesses are going to get this product to market.”

Consider prompt writing to be programming. You are a programmer who is writing computer instructions.

You can use APIs to get developers to wrap your prompts into additional code, allowing you programmatically send data and receive it at scale.

Here’s how LLMs can scale businesses and transform them for the better.

It’s important to remember that these tools are available everywhere.

Microsoft Office, including Word, Excel and PowerPoint and other services and tools we use every day will include this technology.

Penn added, “Because natural language programming is used, traditional programmers may not have the best ideas.”

LLMs are driven by writing. Marketing or PR professionals, not programmers, may find innovative ways to utilize the tool.

Search marketers: An extra tip

We are starting to see an impact of large-scale language models in marketing, and specifically on search.

Microsoft launched the new Bing powered by ChatGPT in February. Users can talk to the search engine directly and receive direct answers without clicking any links.

Penn said that these tools will take a big bite out of unbranded searches because they answer questions without clicking.

We’ve already had to deal with this as SEO experts, thanks to featured snippets. But it’s only going get worse.

He suggests that you check Bing Webmaster Tools and Google Search Console to see what percentage of your traffic comes from informational, unbranded searches. This is the most risky area for SEO.

The first post Scaling up the use of large-language models in marketing was published on MarTech.