For brands figuring out how to utilize AI, one option that could present a lower barrier of entry is small language models or SLMs. 

Ever since OpenAI launched ChatGPT almost two years ago, large language models (LLMs) have been all the rage, capable of yielding a wide range of knowledge thanks to their huge appetites for data. SLMs, meanwhile, channel all of their processing power toward a much smaller body of tasks. Not only are these models cheaper to operate than LLMs because they require less training data, but they can also perform better across a number of enterprise use cases, such as summarizing large volumes of text, providing technical support and answering customer questions, according to research conducted by Salesforce.

Last month, Google introduced a new feature for Gemini, called Gems, that allows users to program an AI assistant for a single purpose. Gems is not technically powered by an SLM, but its emphasis on custom, single-purpose AI applications could help ready marketers for the use cases afforded by smaller models, said Damian Rollison, director of market insights for marketing platform SOCi.

For example, a big box retailer could train a Gem on its brand style and apply it towards an internal task, such as cataloging and describing products online, said Rollison. That same retailer could train another Gem for the sole purpose of responding to the plethora of Google reviews left by customers that may otherwise go unanswered. 

A general-purpose AI system would require far more resources to continually train for these tasks, and its outputs could still be less precise, according to Rollison. These systems may have served marketers well over the past two years, but one of their primary shortcomings is that they aren’t subject-matter experts in anything, he said. CMOs have struggled to advance the use cases afforded by these systems beyond low-hanging fruit such as social media-related content creation, and are now facing the threat of shuttering their AI projects because of a lack of financial returns.

“LLMs are starting to average out, especially for enterprise clients,” said Augustus Cook, VP and head of design for R/GA’s Americas business.

AI providers may be aware of this, for Google is not the first to shift towards single-purpose offerings. OpenAI released custom GPTs last November, which allow users to tweak a version of ChatGPT toward a specific expertise. And Anthropic launched a similar feature for its Claude model earlier this year, dubbed Artifacts.

Meanwhile, the launch of Gems comes as Google tries to move past a series of public flubs that have tripped up Gemini. The tech giant last week pulled a demo video that inaccurately portrayed its platform's abilities. Before that, an AI ad for the Olympics stirred backlash and Gemini went off the rails with insensitive misinformation. 

Programming a Gem 

Using a Gem is a bit more elaborate than using a general-purpose chatbot because a user will first need to program the system to complete a certain task. This process resembles an advanced method of prompting, said Christopher Penn, co-founder and chief data scientist of Trust Insights, a marketing analytics consulting firm. Penn demoed the Gems platform for Ad Age by building an assistant that writes ad copy for Google Ads. 

A user can write instructions for their Gem in any way they see fit, but one hack that Penn illustrated is to develop them inside Gemini’s general-purpose chatbot. Penn queried Gemini to first explain everything it knew about best practices when it comes to creating copy for Google Ads copy. This is the point at which a user primes the AI in terms of the subject matter at hand.

Next, Penn asked the bot to create a scoring rubric based on this information, such that any copy could be critiqued in a standardized way and subsequently improved. The function of this step is to give the Gem a way of comparing its own outputs, so that a user can have a headstart when processing which results are most optimal.

Finally, Penn told Gemini to formulate the best practices and scorecard into a body of instructions for creating ad copy. The only thing left to do was to simply copy the resulting text, open the Gems platform and paste them into the “Instructions” field. After a few moments of processing, a Gem titled “Google Ads Creator” was ready to use. 

As with Gemini’s general-purpose chatbot, the Gems interface allows users to upload documents. These could be files of client data or brand style guides—whatever may be useful  for the Gem in order to produce content that is precise and up-to-date. File uploading is a strength of Gems compared to OpenAI’s custom GPTs. Gems can maintain up to eight times more data from text and file uploads, or up to 700,000 words in a single session before it starts forgetting bits and pieces, according to Penn.

“And with all that extra context and all those extra instructions, you’re giving the model more guardrails,” Penn said, referring to protections that prevent AI from going rogue.

Generating and optimizing outputs

Once the Gem has been programmed, a user will need to prompt the bot with the task at hand. The more precise the better, so deconstructing a single task into its constituent parts can be helpful. For example, Penn’s query gave the bot a product (in this case, a promotion for an AI consultation from his company, Trust Insights), relevant information (his company’s landing page URL) and step-by-step actions to complete (“write some copy for Google Ads, score those ads, pick some winners and then refine the winners”).

The Gem spit out five headlines for the promotion including “Leverage AI for Business Success” and “AI for Business Growth,” along with two description lines and a pathway to sign-up for the promotion for each. Then came the evaluations, applying numerical scores to different criteria and sub-criteria, such as “Keyword Relevance and Targeting” and “Match Keyword Intent,” respectively. Based on the total scores across the criteria, the Gem chose the two highest scores: “Generative AI Consultation,” which received an 80 out of 100, and “Leverage AI for Business Success,” which received a 78 out of 100.

The bot then automatically refined these two options, adjusting the headline itself or its two lines of descriptions. It created two more variations for each based on these adjustments, then produced a final ranking of its top 5 choices. “AI for Business Success” was the winner, with a total score of 90 out of 100. Its strengths, according to the Gem, were a strong call to action, the fact that it highlights expertise and that it clearly states a benefit for the audience. And if it were to improve anything, it reported that it would further emphasize unique selling points.

A user can continue the session, such as by requesting the Gem for new variations or more adjustments on a select few variations. One can even build additional Gems supporting adjacent tasks. For example, Penn said it may be useful for his demo to build a Gem that can analyze pre-existing ad copy and determine the qualities that make it compelling. 

A whole group of Gems could ideally be distributed throughout an entire company, effectively forming a new layer in the tech stack, accessible to all employees. But for the moment, Gems are not shareable, Penn said, which means the only way another employee could gain access would be by copying the bot’s instructions and re-creating it from scratch. 

Despite the need for further development, Gems already may be worth the experimentation. In all, Penn’s work took roughly 22 minutes, from asking the general-purpose chatbot to gather information to receiving the Gem’s top 5 pieces of ad copy. Of course, Penn is a Gemini expert, so many marketers will require more time to test the tool, determine its strengths and weaknesses and point it in the right direction.

“The technology is very impressive, but it is gated by your experience with it,” Penn said