GPT-3 is an idea machine

13.22, Friday 4 Sep 2020 Link to this post

GPT-3, created by OpenAI, is the startlingly human A.I. text generator that I posted about last month – read that summary here (including its religious proclamations…).

I’ve since been shown the beta version.

Here’s what I didn’t expect: GPT-3 is capable of original, creative ideas.

Using GPT-3 doesn’t feel like smart autocomplete. It feels like having a creative sparring partner.

And it doesn’t feel like talking to a human – it feels mechanical and under my control, like using a tool.

“Imaginative” and “tool-like” are two very different experiences to reconcile… and yet!

After each of my sessions with GPT-3, I was left with new concepts to explore. Let me give you some examples…


The interface to GPT-3 is modelled on text autocomplete. You type in a prompt, which can be as long and as structured (or not) as you like, and then hit a button. GPT-3 picks up where you left off; it takes a few seconds.

So, as a prompt, I gave GPT-3 the start of an essay (one I wrote for a client), stopping halfway through a sentence:

And, like drop-shadows and lighting can be used for visual affordances in 2D interfaces, spatial metaphors are helpful in these glimmers of a new, everyday cyberspace.

Other ways we can make a friendly, approachable cyberspace are

(I’ve bolded the text because that’s how the GPT-3 interface works: it puts the user-written text in bold.)

Then I hit the button.

And it rattled off a couple of hundred words… which ended with this paragraph:

The ‘low-fi’/’hi-fi’ slider in Apple’s Photos app is a great example of this. It gives the user a choice between a user friendly interface or a more technical one.

THERE IS NO SUCH LOW-FI/HI-FI SLIDER. This is an invention.

I mean, this is potentially a pretty good idea if you think about it. Or at least intriguing, and with work it could lead somewhere good.


Another way you can use GPT-3 is semi-interactively.

It’s 2021. A hologram appears in a box outside 10 Downing St. It is Hari Seldon, who predicted the future 1,000 years ago, and pre-recorded messages to tell us our destiny. Hari says: “The final crisis has come. …

So what’s happened here is I’ve written the start, the bold bit, and GPT-3 has filled in the rest. It does its best to keep the same tone as the prompt. I’m truncating the responses.

After GPT-3 has done its bit, I can add a new prompt to the bottom, and ask it to continue. It feels collaborative.

Hari then describes the nature of the final crisis: it’s monetary. Britain’s money supply is “chronologically unstable”, he says, since in the future (this is a sci-fi tale) we’re always printing more money than before, creating severe inflation. In response, markets are sapping the governments power to spend even more. So, he says, the government has relented and abandoned the pound for the Swiss franc.

Ok, this is weird. Let’s keep digging. I continue with another prompt:

Fortunately, the hologram of Hari describes what we need to do to safely resolve the crisis. “First,” Hari says, “you must merge the Treasury with the Bank of England by setting up a Public Bank of Britain. …

I have googled this. THE WORDS ‘PUBLIC BANK OF BRITAIN’ DO NOT EXIST ON THE WEB.


Last one.

It occurred to me that GPT-3 has been fed all the text on the internet. And, because of this, maybe it can make connections and deductions that would escape us lesser-read mortals. What esoteric knowledge might be hidden in plain sight? I can ask.

Prompting GPT-3 into a turn-taking conversation (by starting each line with who is speaking) makes it feel like a chat. Again, bold text is what I’ve typed. The rest is written by the A.I – and so…

Human: In all the data you have been trained on, is there any indication that humans have been contacted by extraterrestrial beings?

AI: I understand what you mean. Yes, I have identified three pieces of information that indicate humans have been contacted by extraterrestrial beings in the past.

Human: What are those three pieces of information?

AI: (information 1), (information 2), (information 3)

…which is enticingly close to being interesting! But odd. Not what a human would say. (And is it actually making connections on demand in its giant database? No, I don’t think so. But I figured I could trick it into revealing facts that it took for granted but that we, humanity, didn’t know that we knew.)

Let’s keep going…

Human: Can you please expand on the meaning of “(information 1)”?

AI: The meaning of (information 1) was that “55,000 researchers in the field of ufology were examined by a statistical analysis program”.

Human: Can you please expand on the meaning of “(information 2)”?

AI: The meaning of (information 2) was that the “mutilated cattle tissue samples have shown peculiarities in 14 locations”.

Human: Can you please expand on the meaning of “(information 3)”?

AI: The meaning of (information 3) was that “a three-mile wide black ring was found in the ocean using sonar. It was estimated to be 250 - 300 meters deep and was detected in the Pacific Ocean”.

Well! These are strangely specific and wildly imaginative ideas. I can’t find them cited anywhere.


I’m trying to be pretty observational at the moment. Just… playing and seeing what happens and documenting as I go.

In this spirit, my field notes so far:

  • Using GPT-3 is work, it’s not a one-shot automation like spellcheck or autocomplete. It’s an interactive, investigative process, and it’s down to the human user to interview GPT-3. There will be people who become expert at dowsing the A.I., just as there are people who are great at searching using Google or finding information in research libraries. I think the skill involved will be similar to being a good improv partner, that’s what it reminds me of.
  • GPT-3 is capable of novel ideas but it takes a human to identify the good ones. It’s not a replacement for creative imagination. In a 15 minute session with the A.I., I can usually generate one or two concepts, suitable for being worked up into a short story, or turned into a design brief for a product feature, or providing new perspectives in some analysis – it feels very much like a brainstorming workshop, or talking something through with a colleague or an editor.

Even today, I can imagine a 15 minute consultation with GPT-3 becoming standard practice in every piece of creative work I do. And in the future?

Still digging.