I’m sorry all the images were borderline creepy that I generated
When was the last time you had a joyful experience using AI from a UI or UX perspective?
A lot of the joy that comes from these AI systems is the outputs. Joy could come from a sense of amazement, accomplishment, or imagination, but not from the actual AI interaction itself. They still feel largely devoid of personality, opinion, and edges.
Many in AI today are struggling with the idea that a large portion of AI-enabled software companies are very “thin” layers, with core technology moats accruing to the foundation models that power these application layer companies. This new cycle of AI has brought material tailwinds, resulting in early eye-popping signs of commercialization even at scaled revenue figures. With that said, we are starting to see cracks in these companies as revenue ramps materially.
Specifically, startups with high flying products can quickly experience death by a thousand cuts on the product side, and net negative churn on the commercialization side due to low barriers to entry and exit, incredible pace of innovation, and minimal product allegiance. I wrote about this dynamic at length in Company Building in the Curiosity Phase of AI.
Because of these complex dynamics, it’s likely that the next wave of AI-enabled software companies must be analyzed on different competitive vectors as we enter this stage of the AI cycle. Some of these nuances are more native to deep tech investors, understanding how to extrapolate milestones that don’t come from things cohort analysis or paying close attention to well-understood traction metrics. Others are more fuzzy to all market participants.
One example of the latter which we’ve been exploring has been the concept of personality-injected and opinionated AI-enabled experiences.
One of the amazing things that MidJourney was able to accomplish early on was an aesthetic on how they trained the models and what they wanted the model output to be like. Contrast this to OpenAI’s Dall-E which felt precise (at times) but somewhat lifeless in its outputs. What happens when you add personality or an opinionated aesthetic to a given AI tool is you create a community that feels allegiance to your product in a deeper way than just ROI, perhaps allowing for products to compete with larger, more general purpose models, or just better-funded organizations.
This community can be an immensely beneficial flywheel and perhaps a more important moat than we appreciated, as we’ve seen with the thousands of people contributing to the Midjourney and Runway communities over time, pulling out new experiences with prompts and showing everyone what is possible. These are one of the core things that compound within these startups outside of technical knowledge, data, and compute.
As we see base models continue to proliferate, and fine-tuned models begin to sit on top of them, our hypothesis is that these “thin layer” products should focus on having a clear opinion of how they form the output or how they shape the human to AI interaction layer. Put simply, the maximally viable product in AI is likely not where the vast majority of startups should focus today.
This can manifest itself in stylistically-tuned models, specific prompt injections, and a bunch of interaction steps surrounding these. Over time, this could enable these products to build deeper moats and expand outside of the “thin layer” narrative that allows it to accrue value regardless of the base model value capture that exists.
A thin layer of AI feeding into a thick layer of personality/opinion, resulting in a flywheel for increasing value capture as we continue to learn what parts of the stack enable compounding advantages and lasting moats.
A parallel I like to draw is fine dining. Some establishments perfect the experience with a specific take on what dining at their restaurant should *feel* like. Others put out incredible food in a cold, stark atmosphere. The feeling is often what we are chasing more than the food quality.