The ad industry is always looking for new ways to leverage generative artificial intelligence (AI) to create more engaging and personalized content for their audiences. One of the latest trends in this field is the use of small language models (SLMs), which are scaled-down versions of large language models (LLMs) that can perform more specific and efficient tasks.
What are SLMs, and why are they useful?
SLMs are AI models that are trained on smaller and narrower data sets than LLMs, which are trained on massive amounts of public data. This makes SLMs more focused and relevant for certain tasks, such as generating dynamic creative assets, summarizing documents, or answering questions. SLMs also reduce the risk of inappropriate or biased outputs, which can be a problem for LLMs that learn from unfiltered data.
Another advantage of SLMs is that they are cheaper and faster to train and deploy than LLMs, which require a lot of computing power and resources. SLMs can also run locally on devices, such as mobile phones, which can improve the user experience and privacy. SLMs are therefore more accessible and adaptable for agencies and brands that want to experiment with generative AI for their specific needs.
Who is using SLMs and how?
SLMs are not a new concept, but they have gained more attention and popularity in the past few months, thanks to some major developments and announcements in the AI industry. For example, Microsoft launched its own SLM, Phi-2, in November 2023 and revealed that some of its customers, such as Anker, Ashley, AT&T, EY, and Thomson Reuters, are exploring Phi for their AI applications. Meta, formerly Facebook, also open-sourced its LLM, Llama-2, which can be used to create SLMs for different domains and languages.
Some of the use cases for SLMs in the ad industry include:
- Generating dynamic creative assets: SLMs can be used to create personalized and relevant ads, banners, headlines, or slogans based on the user’s profile, preferences, or context. For example, R/GA, a global innovation company, is testing SLMs for this purpose as part of its strategy and innovation services for its clients.
- Summarizing and classifying documents: SLMs can be used to extract the main points and insights from large and complex documents, such as reports, contracts, or articles, and categorize them according to certain criteria. For example, AT&T is using SLMs for subdocument summarization and classification within its internal question-and-answer chat applications, such as Ask AT&T.
- Answering questions: SLMs can be used to provide accurate and concise answers to user queries, either through text or voice, based on a specific knowledge base or data source. For example, EY is using SLMs to power its virtual assistant, EY Helix, which can answer questions about accounting and auditing standards and practices.
What are the challenges and opportunities for SLMs?
SLMs are not a silver bullet for the ad industry, and they come with their challenges and limitations. For instance, SLMs may still require a lot of data and expertise to train and fine-tune, depending on the task and the domain. SLMs may also have lower accuracy and diversity than LLMs, as they are more constrained by the data they are trained on. Moreover, SLMs may still pose ethical and legal issues, such as data privacy, intellectual property, or social responsibility, which need to be addressed and regulated.
However, SLMs also offer a lot of opportunities and potential for the ad industry, as they enable more creativity, efficiency, and customization for content creation and delivery. SLMs can also help agencies and brands differentiate themselves from their competitors and connect with their audiences in more meaningful and engaging ways. SLMs are therefore a promising and exciting area of general AI that is worth exploring and experimenting with in 2024 and beyond.