Reading Jack Smyth’s reflections on the recent BRXND Marketing x AI conference in LA I came across the idea of ‘latent space’ as it relates to brands and LLMs. I confess I’d not heard the term before so I did some digging to find out more about what it means, and I think it’s a rather fascinating concept.
It’s useful here to think of a Large Language Model as a massive, well-organised library that has read almost everything ever written on the internet but instead of the books in the library being arranged alphabetically, they’re grouped by meaning and relationships. This is kind of like how bookstores group together books into genres or categories (like ‘Science Fiction’, or ‘Self-Help’, or ‘biographies about famous inventors’).
And if we now imagine that the library has a giant map that shows how different topics, words, and ideas relate to each other. This map isn’t based on exact definitions but rather on how words and concepts naturally cluster together based on context. An example here: if we looked for something related to ‘luxury cars’, the model wouldn’t only pull up brand names like Mercedes or Lexus, it would also recognise that the term is also associated with other terms like ‘high performance’, ‘sleek design’, or ‘status symbol’. So the latent space is like the invisible map in which all of these ideas live closely together. Jack links to a talk given by Pip Bingemann, co-founder and CEO of Springboards, showing a simple visualisation of a latent space.
I think this is a fascinating idea which has all kinds of implications as the role of AI models and agents in marketing rapidly increases. Say you’re a marketer wanting to understand how your brand exists in it’s own latent space, visualisations of the kind that Pip demos can be used to understand how LLM tools are understanding your brand and it’s associations. If you then run a campaign focused on particular brand attributes or angles then you may be able to see whether the way in which the brand is represented in its latent space has changed. If you’re a strategist wanting to understand a concept or an idea, looking at it’s position and associations in its latent space may throw up interesting connections or new lines of thinking. Remember, LLMs are (for now at least) trained on content that humans have created and so they can reflect how people, and not just machines, see the world (with the usual caveats about automatically assuming machines are truly representing the information they are trained on).
This matters because as people use LLMs more and more, how you show up in that space will become an increasingly important predictor of visibility but also association. The invisible map that the LLM is using to generate its responses will determine the words and ideas that your brand is associated with and therefore how it is represented. It’s a bit like thinking of LLMs as a huge Spotify playlist generator that, rather than simply organising songs by artist, clusters songs by genre or vibe (sidenote: ‘The Sound of Everything’ is a curated playlist on Spotify which has one track from every genre that Spotify tracks and there are 6,430 songs on it). If your brand were a song, where would it be placed? And if you wanted to shift your brand’s perception, what other songs (ideas, keywords, associations) would you need to be near?
I write a weekly Substack on the latest in AI, digital strategy and trends, transformation insights and the best links on the internet. To join our community of over 10,000 subscribers you can sign up to that here.
To get posts like this delivered straight to your inbox, drop your email into the box below
Photo by Jeremy Thomas on Unsplash
Leave a Reply