There is a Massive Gap Between What We Want to Create and What We Can Describe
Why your lack of vocabulary is a real issue in the age of LLM-based creative tools
If you were a teenager (or chronically online) in the 2010s, you might recall the insurgence of the word “aesthetic”. It began with people using it to describe a certain vibe - popular aesthetics at the time included twee, grunge, and indie sleaze. As internet terminology evolved, the word “aesthetic” took on its own meaning. People dropped the descriptor and just called things “aesthetic”. If something was visually pleasing, it wouldn’t get called a certain aesthetic, it would just be described as “so aesthetic”. Coachella was so aesthetic, Lana del Rey was so aesthetic. We started to use what was previously a neutral noun as an adjective that was attached to a standard of liking decided upon by the masses. Despite having an era-defining visual identity, no one ever described the Supreme/sneakerhead/hypebeast subculture as “aesthetic”, perhaps because it didn’t align with the mass market’s bar of visual appeal.
This marks my first memory of the mainstream discussion eliminating the nuances of subcultures and transforming what was once a noun to be associated with a descriptor into an adjective that carries an implied standard of excellence.
Fast forward to the 2020s, and I’m seeing this same phenomenon happening to the word “taste”. As our Silicon Valley overlords have made progress in conquering workflow optimizations, they’ve begun to fixate on optimizing the next thing: taste. By now, taste is a tired discussion. We’ve mulled over why it is or isn’t important, who has it and who doesn’t. In my opinion, these conversations miss the critical nuance that taste itself means nothing without attribution to whose taste it is and how it was developed. Here lies what I think the actual problem is: you are not tasteless, you just don’t have the vocabulary to describe what you like.
In Virgil Abloh’s Harvard GSD lecture he states, “You also have to have mentors, dead or alive. You have to connect with a body of work or someone who formulated a thought and an aesthetic, and then build yours upon theirs.” The reason why you like the things that you like is because you are influenced by your environment, the media you consume, and the idols you look up to. In a conversation I had with (my favorite podcaster) Jackson Dahl, he brought up the idea that creators in art and fashion tend to be referential and draw on inspirations from the past, while modern technologists tend to be forward-looking and spend little time looking back. It is important to acknowledge that modern technology is, in fact, built upon the historical context that came before it, yet time and again, we see people striving to invent completely novel things. An example that comes to mind is technologists’ use of the words “disruption” and “first principles”, implying that things must be completely torn down and built from zero to truly push us into the future. However, this rarely applies to tastes and aesthetics as they aren’t simply pulled out of thin air. Artists are revered for being able to draw on niche references; look how Law Roach became the most famous stylist of our time for his god-tier archival pulls and historical references. Every artist draws inspiration from a web of experiences and inputs that are mentally sorted into preferences that shape their outputs. It’s difficult to create something new without building on the past, and having strong references is the foundation for pattern matching and ascribing vocabulary to your preferences.
I’d caveat that this isn’t to say that technology is not referential - the resurgence of 90s Apple marketing and 80s retrofuturism aesthetics have been key points of contention. This has even sparked a circular dialogue back to some people calling for more innovative and novel aesthetics. The same frustration has been echoed in other media like fashion and cinema.
The more you consume art, the better you can describe it. You begin to pattern-match certain visual themes, techniques, and compositions. You then learn to attach words to these patterns, which form the basis of your taste vocabulary. This is a critical consideration as AI unlocks a whole new universe of creative tools and enables people to experiment with new mediums. Consider image generation, where the model inputs are text and reference images. If I wanted to master wielding image models, I would first need the vocabulary and reference library to get the outputs I want. It is more direct to use words like “skeuomorphic” or “bento grid” than it is to string together sentences that broadly describe these things. Likewise, it is more direct to say “90s editorial candid digital camera shot” attached with a reference to a Calvin Klein BTS shot to get a very specific image than it is to describe the vibe with zero references. Jamey Gannon comes to mind as someone who has mastered this skill.
What the internet often calls tasteless are things made with very little intention and feel detached from their broader contexts. An inexperienced person making an AI-generated garment might be labelled tasteless, perhaps because they don’t yet know the best practices that have been refined through millions of prior iterations. However, a fashion designer using AI to develop new prints might catch and tweak subtleties like what balances a color palette. The difference is that an outsider who lacks the context and vocabulary to create an intentional output often ends up creating things that feel uncanny.
While we can attempt to calculate or develop a criterion for what makes something tasteful or not, we must remember that this is all subjective. Your taste may not be someone else’s taste, and that is because taste itself is a thing, not a descriptor. If we truly think it is important to add a dimension of aesthetics and intention to the technology we build, it is important for us to learn how to wield artistic vocabulary, especially as many of the tools we use are currently language-based. The outputs of an LLM are only as good as the inputs you give it, and without grounding our abstract expectations in usable terminology, it is extremely difficult for us to wield these tools to create the outputs we are looking for.