Problems come in near infinite shapes in the real world, but for the sake of brevity, let's categorize problems into roughly two sets:
A repetitive problem might be something like tying your shoes, or automatically turning on the air conditioner when the thermostat gets too hot. Repetitive problems have a clear output, and can be habitualized or programmed.
A non-repetitive problem is writing a blog post (like this one!) or building a rocket ship for the first time. Any creative act with unexplored territory falls under this category. These problems can't be programmed due to their high variability, and obscure completion metrics.
AFAIK, there are two things in the known universe that can reliably attempt to solve both of these types of problems: the human brain and GPT-3.
In fact, this is what made GPT-3 so exciting to so many people. It seemed to have an intuitive understanding of concepts, relationships, and causality; just like our own brains!
Where GPT-3 was weak was its inability to "store" context in memory, and the ability to be able to do second order level "thinking".
GPT-3 in 2022 brings with it a whole new suite of capabilities. It has a larger "memory" window for topic retention. GPT-3 is great at following complex instructions in sequence, giving it more "if this, then that" type thinking abilities.
Let's take a look at a few examples of GPT-3 solving repetitive and non-repetitive problems.
Here, I'm using GPT-3 as a scheduler. This scheduler is complex because it has listed constraints (two meetings, one at noon and one at 2:30) and non listed constraints (I want to code for three hours, but it is up to the model's discretion as to when is the best time to do that).
This example has aspects of repetitive and non-repetitive problems. Putting in a meeting at noon and 230 is trivial. We can extract that data using a program easily.
The other aspect, finding three hours to code in a work day given the meetings, requires more creativity.
As you can see in the picture below, GPT-3 added exactly three blocks for coding and not more. GPT-3 understood that a meeting at 230 starts “during the hour of” 2pm.
Here, GPT-3 shows us its excellence at being able to interact with applications outside of the API itself.
Let's say that a large part of your daily workflow involves checking Stack Overflow for new answers to a question.
You could, of course, click on your web browser, open a new tab, type in "stack overflow" in your URL bar, and navigate to the page you're looking for.
Or; you can automate this task and put it on a hotkey or activate at a certain time in the day.
Here, we're telling GPT-3 to write the logic for this task, and then we can directly copy paste it into Mac Automator as an AppleScript.
This shows that GPT-3, much like the human brain, is malleable enough to solve problems in different, external contexts (much like how a student can excel in history class at 2pm and on the baseball field at 5; both require different problem solving techniques but the brain can do both).
These toy examples are merely scratching the surface of what is possible. Since most (I might argue all?) problems fall somewhere on the spectrum of repetitive and non-repetitive solutions, GPT-3 can serve as an extremely helpful companion in solving both types.
Be on the lookout for many more more intelligent, tiny, single-task, embedded robots near you.
A collection of ideas about retirement, and what happens after
Be wary of websites that are just slick designs, with no quality backing their offering
Creating that vignette effect by using box-shadow in a clever way
Each week, I send out something that looks like this. Sign up below if you want in!