Bringing specificity and uniqueness to AI models
Scarce, individualized AI models that create unique value through personality differentiation, persistent memory, imperfection, and exclusive ownership mechanisms represent a novel approach to building moats that can accrue at the application layer of AI, where traditional switching costs remain persistently low.
A core belief we have at Compound is that in the near-term, value capture will shift from the model layer towards the application layer of AI as we approach scaled minimum useful performance for a collection of LLMs. With this in mind, we are thinking about how application layer companies can build higher barriers to exit that also allow for novel product experiences that aren’t purely solved by memory + frontier model UI. This personalization has manifested itself in AI models today with ChatGPT using a form of memory, and Claude allowing personalization in writing styles. Both have also implemented Projects which effectively use the context windows in order to bring “personalization”. While context windows may get obliterated, accurate retrieval still remains an open (but increasingly solving) question, as shown by recent performance of Gemini Flash 2.0.
What if instead of implementing these tasks at the model layer one were to build them at the application layer either with explicit fine tunes, better memory techniques (highlighted in the Related Reading), and specific use-cases. What if we then took it a step further and allowed models/agents to be 1/1 or scarce such that only a core group of people can access and interact with them, perhaps attributing value to the lore of the training of the model.
We can imagine that some of these dynamics can be unlocked further with the rise of performant distilled models that can be re-trained or fine-tuned locally on consumer GPUs over time.
A more structured view:
Other Thoughts There are likely multiple use-cases that one could imagine but the core idea is bringing some form of novelty/uniqueness and personality to models such that the core underlying pre-trained model is abstracted/commoditized and the layer on top creates lock in.
The evolution path might mirror other digital-to-physical transitions:
There are a lot of technical challenges still left to figure out (as can be seen in Related Reading below), but also there are lots of fun experimental apps one could build on top of this infrastructure.
An example we came up with was an app called “terminal” in which you loaded all sorts of context for a character at end of life that enabled them to have views and think differently and the character gets asked a single question and vice versa and eventually the AI bot dies.
Alas, we are not the smartest people but it’s a fun idea.
Related Reading
Comparable Companies
