This is an article about the effectiveness of large language models, which have learned to do tasks beyond regurgitating text despite only having been trained to regurgitate text.
As long as you are able to describe the task as a text stream and are ok with having a text stream as output, GPT-3 might actually be able to do the task with just a couple of examples as context. I saw a talk at Black Hat about using GPT-3 as a spam filter. Because GPT-3 has trained on so much spam, it knows what spam looks like. It just needs a couple of examples, and off it goes.
> The most complicated reasoning programs in the world can be defined as a textual I/O stream to a leviathan living on some technology company’s servers.
The references to the UNIX philosophy feel a bit tacked on, honestly. Probably just there to grab attention.