Okay, the GPT-3 hype seems pretty reasonable
This morning TechCrunch covered an interesting round for Copy.ai, a startup that employs GPT-3 to help other companies with their writing projects. GPT-3, or Generative Pre-trained Transformer 3, is a piece of AI from the OpenAI group that takes text from the user, and writes a lot more for them.
As part of the process of covering the Copy.ai round, I got caught up in the idea of AI-powered writing. I've long been more curious than afraid of automated writing. So when the Copy team described their very positive impressions of the GPT-3 AI writing tool to TechCrunch during an interview, I was intrigued.
To scratch this newly-formed itch, I doodled around this morning with a competitor of sorts to the Copy team , Headlime. And, freaking heck am I am impressed at what folks have managed to build around the GPT-3 technology.
Sure, GPT-3 can add words to a prompt. But the technology can do a lot more than that. The GPT-3-powered Headlime managed to not only write some medium-good stuff for me, but also managed bring in concepts concerning my reporting beat that were in my head but not in the prompts I provided.
I can't do better than just show you what I mean. So, here's what happened when I used Headlime for the first time, sans help.
Here's the first thing that Headlime showed me, a language selector and a request for a description of the post that I wanted to write. I decided to push the system a bit by just telling it about a piece I need to write in light of today's market action:
Ha ha, I thought, that will kick it in the teeth and I, a biped of intelligent meat wrapped around some calcium sticks, will feel grossly superior to the computer player. I hit go and then realized that I actually had to provide 500 characters of stuff, so I rambled for a bit to fill in required length:
Time for the next step! Hitting the button brought up a list of possible headlines for the post I was helping create, which were honestly not terrible:
Fair enough, yeah? At this point I was starting to become impressed.
I selected the first headline as it was my favorite and moved along. Next came the work to get an intro put together for the post, a process that involved the strenuous work of clicking a button:
Here are the options proffered:
Again, not bad.
What struck me about these are not merely minor variations on each other. They are structured differently, taking various angles on what I was halfway talking about in the 500 characters of bilge I had fed into the system. I was starting to wish that I had given GPT-3 a bit more to work with up top, as it was trying its best after I had clearly not.
Intro selected, I was brought into a CMS of sorts, where our selected bits were included, and your humble servant was asked to do a bit more writing.
I was happy to oblige, only for the system to stop me and offer to take over:
Having precisely no idea what a credit is, or what two of them cost as I was on a free trial of sorts, I hit the Write for me" button. This is what came out:
Look at how it finished that sentence I started, even after I used em-dashes! The software gets the next sentence backwards, but is right back on the horse afterwards talking about how higher interest rates make exotic investment classes like venture capital less attractive! I was gobsmacked.
I will keep playing with the tech and the various software wrappers that are being built to productize GPT-3. More notes to come. But I wanted to pause and share my initial delight. This is cool. I can't recall the last time that technology actually shocked me. But, well played GPT-3, you're amazing.