Automate the Simple Stuff With Stupid Tools

I have a friend who trained for two years at the best culinary school in the country. A five foot-three brunette with crackling energy, her life’s focus was on making the best damn risotto this side of the Mississippi. To graduate at the top of her class, she put in 80-hour weeks in the kitchen, all so she could ascend within the food world. 

Finally, after all of that blood, sweat, and tears, she got a job in the kitchen of a Michelin-starred restaurant in California. She had done it! Her job? French fries. For 12 hours a day, her life was dicing and frying. She did it for a year. Despite being able to flambé and roast and saute, she boiled Yukon Golds in duck fat day after day. Her reward was burns up and down her forearms and a $30,000 salary. The higher-skill-level, more complicated tasks were reserved for the experienced chefs. 

LLMs like GPT4 are the overqualified line chefs of the mind. Their task is to automate the French fries of thought. They do the simple, stupid stuff for you. IQ is a flawed metric, but it is a useful heuristic for when you’re applying AI. Think of each successively more powerful model as a person with more intellectual horsepower. 

This isn’t just a case of “this AI think gooder” for every subsequent additional version of the GPT paradigm. Just as every individual has certain natural talents, so, too, do AI models. There will be AI specialists and generalists, idiot savants and jack-of-all-trades. How that manifests today is with models focused on programming or those focused on general text generation. They’ll either be good at one specific thing or serviceable at a great many things. 

As a result, over the next few years, a decent percentage of thinking you do today will no longer be required (or compensated). 

This is a profound evolution in what tools can do. 

Old software only did two things: store data and utilize that data to automate workflows. The easiest way to think about this is a spreadsheet. You enter some information into each cell and then write a formula to analyze that data. Tools almost always do some version of both, but they’re specialized based on the market they serve. For instance, Salesforce holds customer data and does automations on top of it. Photoshop mostly focuses on photo editing workflows, but has some storage options. All software tools are some combination of these two core functions. This is an oversimplification, but you get the idea. 

Software was made like this: some task you did on your keyboard was annoying. So you wrote a program that did that annoying thing for you. But you still had to tell the computer program to “do the thing.” It was like you had all the potatoes in my friend's kitchen (data) and the world’s sharpest knife (workflows that use that data), but there still needed to be a hand that wielded the blade.