⚙️ A Reminder: Do Things That Don't Scale

Doing things that don’t scale will never go out of style

In July of 2013, Paul Graham published his famous blog post titled “Do Things That Don’t Scale.” It is full of examples from massive, successful companies about how they got their start, many of their unscalable processes and how they ultimately grew into major success stories.

Well, now it is 2023 and all of these lessons are still true. Funny enough, they were also true 50 and 100 years ago. Walmart started as a general store in 1950 and most definitely did not have its supply chain or operations buttoned up to the level it is today. In 1883, Leggett & Platt was founded by two men in a plow factory but grew into a manufacturer that touches literally all of our lives, whether you know it or not (assuming you’ve slept in a bed or sat in a chair at least once). I have to imagine that they hadn’t yet figured out how to produce bedsprings at scale in the 1880s…because one of their innovations includes building manufacturing plants along the railroad lines to receive scrap metal from auto manufacturers… 

In 2023, our attention spans are short and our willingness to experiment cheaply to prove a concept seems limited. However, that is something every builder, fundamentally, needs to be comfortable with. When considering a large initiative, it is important to ask “how could we validate our hypothesis quickly and in a way that doesn’t cost a lot of time or money.” I am not suggesting that systems design should be a quick, cheap exercise that is unmindful of the future. I am suggesting that testing ideas in order to determine what is worth your time is incredibly important and will continue to be…even in this impatient world. 

Examples of doing things that don’t scale…or scaling them anyway

Have you ever used Google Flights? Fun fact: Google’s travel products work through a combination of direct API integrations with service providers as well as mechanical turk, meaning humans actually validate that the results are accurate and the product functions as expected. I’m sure that overtime, the need for human intervention has decreased substantially but this concept of mechanical turk has been foundational to the building, launch and continued support of that product…and we’re talking about Google! 

The cool thing? Amazon literally has a mechanical turk as a service offering called mturk. It has never been easier to validate concepts with the need for an auto-manual experience (read: humans in the loop). There are armies of people involved in training OpenAI and Google’s LLM thanks to companies like Invisible Technologies, Scale.AI and others.

My point here is that things just aren’t as automated as they seem. They never are and they won’t be for a long time. As builders, we should always maintain that perspective and know when perfect may be the enemy of good. Doing the things that don’t scale may lead to your ultimate success.