collect orthogonal experiences

After not speaking for years, I caught up with a friend of mine. As I gave him the rundown of my last 4 years, he made a comment that stuck with me: I tend to collect many orthogonal experiences.

Since we had last talked, I had gone from publishing economics research at NIST to doing recruiting for deep tech startups to an insurtech startup to quant to joining a VC fund to clinical research and volunteering at Children's National Hospital.

Within days of that conversation, I remember reading a paper attempting to reimagine the LLM embedding space. Then, I returned to the OG of the space: word2vec.

When I first learned about word2vec, the entire concept felt quite intuitive. Classically, within the embedding space: "King" - "man" + "woman" ≈ "queen"

And while this specific approach didn't stick around long-term, the notion remains -- our collection of knowledge is, in reality, very similar to that of a multidimensional space that can be illustrated through geometric relationships.

Taking the concept literally, our mind operates within a multidimensional space. The data points are individual experiences, ideas, facts. And the vectors between them are mental models -- relationships and trends we recognize to exist that can be transposed to link as many data points as possible.

However, these two are not nearly enough by themselves. There remain two other key factors in this picture:

• The dimensionality of the space itself
• The ability to increase the dimensionality of the space (and less importantly to acquire new mental models, and experience new data points)

The world contains effectively infinite dimensions that we struggle to process. To produce the highest fidelity view of each data point and the relationships between them, one must optimize for the greatest number of dimensions within their intelligence. The dimensions of our knowledge are revealed by the diversity of our data points and mental models -- orthogonal experiences force us to develop axes of understanding that wouldn't emerge from repetition within familiar domains. Expanding one's dimensionality requires recognizing that a new set of data points simply cannot fit within the existing dimensions -- that a higher-dimensional perspective is needed.

However, to do that, there also exists a set of factors which dictate one's ability to grasp new dimensions. This is best encapsulated by the concept of neuroplasticity, affected by factors from age to the strength of one's confirmation bias to the rigidity of one's worldview.

Why does this matter though?

Understanding the factors at play here allows one to optimize better for the experiences that are best suited for expanding one's knowledge.

In a long-term game like that of our life journey of learning, always prioritize the highest-order derivatives as they provide the most leverage. Just as adding to acceleration increases displacement more than velocity, and velocity more than starting displacement, we must recognize the same of knowledge.

Therefore we must optimize in the following order:

1. Our neuroplasticity
2. Our diversity of experiences
3. Creation of mental models
4. Collection of data points themselves

And through this formulation, what becomes abundantly obvious is that the greatest decision of all is to be as open-minded as possible, and to parlay this decision into the collection of orthogonal experiences, knowledge, and understanding.

As much as we are compelled to specialize and pigeonhole ourselves, there is no better strategy than to collect orthogonal experiences. Different problems, constraints, epistemologies.

This seemingly inefficient path is optimal for building representational capacity. More dimensions means higher fidelity perception of future experiences. Better pattern recognition across domains. Less projection error when reality presents something new.

If the world is richer than we can represent -- and it is -- then the binding constraint on learning isn't data volume or even mental models. It's whether your representational space can capture the orthogonal dimensions that actually exist.

Collect orthogonal experiences. Your mental model's fidelity depends on it.