Building a Generative AI System You Can Trust

5 lessons from working with large language models

Mark your calendar for Mediaweek, October 29-30 in New York City. We’ll unpack the biggest shifts shaping the future of media—from tv to retail media to tech—and how marketers can prep to stay ahead. Register with early-bird rates before sale ends!

Working with artificial intelligence can be polarizing. The future possibilities of generative AI create a sense of hope … and fear.

We’ve been working with AI at our creative agency Omelet for a few years now. In early 2020, we started using AI to help automate the mundane parts of the banner localization process. Earlier this year, we released a natural language processing-powered web tool, Hikup, to help anyone be more mindful of the language they use. More recently, we spun up an in-house image generator and created a film primarily through AI.

We’ve also been working on the development of a gen AI-powered software product that will expedite elements of the agency process. As part of the development, we had the opportunity to build our own private large language model (LLM). This experience made me realize how critical LLMs will be for all of us moving forward.

Here are five key observations from working with LLMs to create a more helpful, ethical and impactful use of AI within advertising.

Ethics start with education

LLMs are deep learning algorithms that use enormous datasets to recognize, summarize, translate, predict and generate content; they’re the brains that power text-based generative AI. Though experts predict it will take many years to fully understand the inner workings of LLMs, we do have an understanding of the basic principles.

More education and understanding of how these models work is essential to ensure AI has an overall positive impact on society.

Word vectors, transformers, tokens and tuning are all terms your agency should start getting more familiar with if your brand wants to work with AI responsibly. More education and understanding of how these models work is essential to ensure AI has an overall positive impact on society.

Llama 2 democratizes the industry

Thus far, we’ve only had access to proprietary LLMs through OpenAI and Google. But everything changed when Meta released Llama 2 in July.

Llama 2—which has a commercial license—is a free, open source alternative to proprietary LLMs. It’s one of the most consequential AI advancements in the last few years, giving away the same foundational models as the major players that you can fine-tune and control. This transparency is paramount to having an ethical and trustworthy approach to building with AI.

Use a model you can trust

ChatGPT cannot always cite where an answer comes from and, even when it does, it’s sometimes wrong. But Llama 2 and other open-source LLMs enable us to have even more transparency and understanding of their training data.

OpenAI also recently released enterprise versions of ChatGPT that will enable companies more control over how their models are fine-tuned. The better brands and agencies can define the data their models use, the better protected they are from legal risk and unwanted outcomes.

Build for inclusion

Understanding the source and accuracy of your data helps you better trust the outputs of the generative model. But we also must align the training data with our human values.

Because Al algorithms are trained on existing human-generated information, we have to be cautious about the biases within the data. We also have to continuously tweak the models to be more inclusive. Ensuring the data used to train these models represents a diverse range of values and life experiences makes this tool and output more valuable to more people.

These days, more than ever, technology is part of our culture. But we must push to ensure our culture is represented in our technologies.

Private, open-source LLMs are the future

Bigger proprietary models are not always better. You’ll get sharper results solving specific problems with smaller, verticalized models.

For example, a smaller, open-source LLM trained on all your previous advertising copy and communications will perform a much better job of generating content in the correct tone and voice of your brand than a big, proprietary LLM would. The control and transparency that open-source models provide will become essential when creating a branded experience powered by generative AI.

Furthermore, combining first-party data with a branded LLM will supercharge personalized, one-to-one communications for that brand.

Brands will increasingly start deploying personalized and private LLMs, a more brand-safe, refined and responsive version of current gen AI tools.

This story is part of the “Building a Better Agency” special feature.

Adweek magazine cover
Click for more from this issue

This story first appeared in the Oct. 10, 2023, issue of Adweek magazine. Click here to subscribe.