The question is not what you look at, but what you see.

  • Thoreau, Henry David. Journal of 5 August 1851 [Andy: The legendary Thoreau. Definitely worth reading beyond this quote.]

We don’t draw a face as it is; rather, we draw it as our models say it is.

It Began With A Sketch#

I recently took up the habit of pencil sketching, originally to add illustrations to my paper-and-pen journaling, and then as simply a pursuit in itself.

I chose to do this NOT because I’m a natural artist (in fact, I’ve firmly believed that my artistic skills ceased development after the age of 12), but because it was hard for me, as opposed to the way that words naturally flow through me, whether that’s on paper or stored in flash memory.

The start of the process was simple: get a few monochrome artist pencils, a sharpener, and some good sketching paper.

Then: start to draw.

Next step: Get frustrated with my results, which uncomfortably reminded me of my drawings in school.

Next, of course: Look up “how to pencil sketch” on YouTube. Watch instructional videos, and try out their suggestions in real-time.

It was at this stage that I discovered the key instruction from one of these mentors: “The first core skill for sketching is learning how to see.”

My Journey to Seeing Clearly#

And so, I started learning how to see: How to break down complex shapes into their component pieces, see the similarities that connect all forms together, and use their differences to complete the picture. How to hint at three-dimensional perspective in a very much two-dimensional pencil drawing. How to compress the rich detail of the world into a simplified pencil sketch, while successfully communicating the moment that I’m trying to capture.

Sketching can feel like writing, but letting your pencil flow without the reductionist step of using words. It’s pattern matching in life and sharing the summary yielded by the pattern matching.

Whether or not you ever pick up a sketchpad or dream of being an animator, I hope you understand how it is possible, with practice, to teach your brain to observe something clearly without letting your preconceptions kick in. It is a fact of life, though a confounding one, that focusing on something can make it more difficult to see. The goal is to learn to suspend, if only temporarily, the habits and impulses that obscure your vision.

Learning to see is the first step toward making things better.”

There’s a reason that Pixar believed that this principle was so important that every Pixar employee (the majority of whom were not artists) was taught by a legendary artist in Pixar-run workshops so they could learn “how to see”.

Since this practice of learining how to see has transformed not just my sketching skills, but given me a new framework for how I look at new technologies, I wanted to share an example with you.

Recently I’ve been thinking about the ecosystem of companies who support the AI giants, and how to understand which companies may benefit the most from the rapid growth of AI.

To put “learning to see” into effect, I began by looking at the core infrastructure that AI requires to operate and grow: Power, GPU, Storage, Bandwidth, often located in a Datacenter. While GPUs, Power, and Datacenters have been discussed repeatedly, I hadn’t heard nearly as much about discussion about Storage.

The Inevitable Growth of AI’s Storage Needs#

Given the need for ever-increasing pools of data to train new and better AI models, and that this data will likely be held locally by each AI company developing LLMs, this implies the need for monumental amounts of storage to hold this data.

In addition, this data will need to be increasingly rich and high-resolution. Since most of the world’s books have already been digitized and used for training (whether the authors & publishers liked it or not), movies and TV shows encoded, the web has been scraped, the content from user-created discussion forums are being licensed, there will be the need for even richer content: imagery of every surface of the earth (with fidelity exceeding that of the human eye), data on micro-weather patterns down to a single breeze, the texture of leaves on a single plant, the sound of every human voice in conversation, and so much more, all of which will need to be constantly updated and added to a growing pile of rich data.

While today it may seem laughable that people would want to make some of the privacy tradeoffs required for this data to get gathered, we’ve seen repeatedly throughout the history of technology that the majority of people will happily make a privacy tradeoff in return for receiving some amount of value in a “free” product.

So as this massive pool of data continues to grow, both in the amount being gathered each second, and the sum total of data gathered, there will need to be ever more additional storage to hold this data. (And given the competitive nature of AI, it’s very likely that each major AI company will maintain its own separate data set on their local storage, even if that data is wholly or partially duplicative of what their competitors hold on their own local storage.)

Leveraging AI for Deeper Research#

Since storage systems are not my area of expertise, I then setup a query to an LLM (in this case, Anthropic’s Claude) to analyze the storage market and generate a report which would specifically call out which companies would prosper in this scenario.

Claude obliged with a 3-page summary report on AI & storage companies. While my thinking goes beyond this report (my belief that an exponentially larger amount of data that will be gathered for training AI models), this was a helpful outline of the storage market covering companies that I wouldn’t have otherwise seen.

Once I had this, I then had a chance to research their technology offerings to better understand how the storage space will be disrupted and/or expanded by AI’s unquenchable need for data, to evaluate these companies for possible new investments, and to simply enjoy diving into a new rabbithole of information to seek knowledge.

Conclusion#

Learning to See:

If I’d started my process with “just look at the AI landscape today”, without taking the time to break down that complex landscape into component pieces, I’d run the risk of getting flooded by a wave of coverage and products without reaching coherent insights, and missing the chance to speculate on where the largest growth opportunities may lie.

If you try applying the “learning to see” method to your own research, I’d love to hear about it! (Until my commenting system goes live, message me on Substack)