Jean - Software for the Individual
My Life's Work - Top-Down and Inside-Out
Your Life’s Work
It feels important to make sure this is my life’s work, before spending my life on it. Part of the difficulty with the ambiguity in startups is you may have 20 paths forward with 20 angles to attack each from. This can be paralyzing if you approach the world in a purely rational way.
I left finance because I view value as more than money. I view my work as a way to be the change I want to see in the world.
Irreverent is a cultural vehicle. It will shift the culture towards individualism by proving that you can leave existing structures, believe in yourself, and follow your instincts to success. Then, I will invite others to join me in solving important societal problems.
Jean
Similarly, I want our first portfolio company, Jean, to be aligned with this vision. It is most important that Jean is successful, but I also want Jean to be aligned with the change I want to see in the world.
Software fundamentally now understands the world and us. This enables interactive and dynamic experiences that are shaped to the individual. Until we map our brain completely, the closest thing we will achieve to understanding ourselves is using LLMs to map our belief systems, values, interests, strengths, and unique patterns of thought. Personal software is a win for the individual in an over-structured world.
LLMs Understand Us Deeply
What is powerful about LLMs is they understand the world in a deep, high-dimensional way. So you can imagine that there is a scale of a single dimension like curiosity, where one person is the least curious person in the world and another is the most curious person in the world. Similarly, you can imagine another dimension along with extraversion and introversion and many others.
LLMs understand these dimensions and can place you at any point along that scale depending on what it learns from you. What is even more important is they can do this across millions of dimensions. This has opened an unprecedented way to understand the complex black box of the human mind in a precise way.
Old personality tests were built in a way that attempted to bucket people into general categories that psychologists observed between people, such as across the above curious-uncurious and extraversion-introversion dimensions. AI doesn’t need these broad strokes. It simply learns every possible permutation of human thought, emotions, beliefs, etc and knows where you uniquely fall in latent space.
Thought Experiment
The thought experiment I like is asking someone to imagine you had a document (context file) that contained everything you have ever thought, written, or said. Think of it like filling a room with pages that contain all of this information.
LLMs and the attention mechanism work by sifting through all of this data and weighting it based on the relevance to the prompt. An example would be like asking “what does Alexander think is the meaning of life?”
If this information has ever been explicitly defined in the document, it would be pinpointed, weighted heavily, and returned. If it has never been explicitly said or thought, the attention mechanism would focus on related thoughts, such as what you think of life in general, religious beliefs, thoughts on morality, and your line of work.
Based on the related findings, it searches through an internal database of vector programs and finds people who have held similar beliefs and their likely belief on the meaning of life. It triangulates understanding with imperfect information, because it has a general understanding of the world. Old predictive ML and personality tests would simply bucket you in a personality category based on your responses or suggest you believe X because people who bought similar books as you believe X. This narrow understanding of the world was the best we could do at the time.
In General Person Embeddings, I laid this all out for the most part. However, I imperfectly understood the mechanics of LLMs and thought that you could embed the entire context file into a single or multi-layer embedding that would capture general characteristics about you. That would not work. What would work, however, is asking a billion narrow questions about you as a person and capturing both the output of the LLM and the zero-shot embedding and mapping it in latent space for later retrieval. This cache could be later recalled and similar data points in high-dimensional space averaged or concatenated could return who you are. The big learning here is you want the context file to be as exhaustive as possible but you want the query to be as narrow and precise as possible so you don’t dilute your information pull.
Or you could simply just always have access to the context file. One problem with this method is that context is limited to around 128,000 tokens today, so you couldn’t search through the entire room of documents in practice. Fine-tuning your own model is similarly impractical. The most effective way to run this would be the following:
Still, you require all of your context to be retrievable. I will break down the different locations we can find and aggregate our context from.
You could run a search over all of the data to find the context that is most relevant, before using it as context. This could be done through an intelligent form of RAG after breaking up the data into small sections or smaller LLMs that sift through and distill the data, but cut out all the noise.
Ideally after the first 2 steps, you have returned a context file that is correctly sized and formatted so that it can run through the LLM and return the closest approximation of correct user understanding.
Applications
This will enable applications like Delta and Mariko, where the software takes into account a deep understanding of the user. I believe deeply in “software for the individual” and that software will continuously understand us better so we can more directly navigate the complexity of the world on our own human terms.
However, the more important point is that this understanding will enable hundreds to thousands of applications to become deeply personalized and each will require similar solutions.
Jean’s vision is to power software for the individual. That doesn’t need to be through a narrow application. In fact, I think the more important company to build is the one that powers all software and builds the foundation for the next stage of companies.
—
The rest of this document has been redacted.