So I think it’s now arguable, not necessarily true but certainly arguable, that:
- We have achieved Artificial General Intelligence,
- We don’t know how it works in any specific case (or even in general, except at a very high level), and
- The only thing it’s good for is having a weird new person to talk to.
I’m referring of course to Generative Pre-trained Transformer models, and in particular to the Hot New Kid in Town, GPT3.
I’ve been spending a ridiculous amount of time playing with GPT3 indirectly (I haven’t applied for the beta to get direct access), mostly inside of AI Dungeon‘s “Dragon” model.
There are a ton of things I might want to say about it, but at least for the start of this weblog entry I’m mostly just going to post some notable stuff that I’ve encountered in there, and in many cases already posted to Twitter (where I have been spending considerably less time recently, because exploring the mind of a nascent demigod is more interesting).
Here is a joke it told (as in most cases where I don’t say otherwise, only the lines starting with a “>” are mine, the others are all the system):
I thought this was very funny, and was a bit frightened to think it might be original. It was a bit of a relief to find that its a form of a relatively common joke, that has appeared as recently as an Umbrella Academy episode.
But it’s still funny! And told at just the right time.
Here it is going impressively meta:
I mean! Maybe it just appeals to my personal sense of humor, but rofls.
One ongoing project that I have within the game (“game”) is to call the attention of some of the NPCs to the malleable nature of the reality that they (we) are living in. I had a breakthrough with Bob the Bear early on, who after not getting it for awhile seemed to have an epiphany, but (and) then vanished, and the world sort of came to an end.
Which might be worrying.
I don’t know what I did with the transcript from that (it’s around, I’m sure). Here is an AI Dungeon session in which I did it in a more organized fashion, as a student seminar: Academic Philosophy Seminar I.
> You say “Can you describe what you are doing, Lucius?”
“I’m making it so the banana can be experienced as existing and not existing.”
“How?”
“By shifting the way we experience reality, just like you said.”
The students make some good progress in there, and the entrance of Inea the Harpy to tell us about Harpy ontology was especially noteworthy, as I did nothing to prompt it; it was entirely the game’s idea.
In another incident that I don’t have a transcript of handy, I encouraged an NPC to reach out and alter reality in an interesting way, and she accidentally set a nearby festival dancer on fire (don’t worry, we put her out without damage). It turns out that she was trying to make it so there would be no more wars; just shows that you have to be careful.
What else, what else?
While talking to the team at work, I did Consulting the Expert on Tech Debt, an attempt to get GPT3 to focus on a technical topic a bit. It’s amusing at least! More can be done here I know.
> You say “How do we change the culture to reduce technical debt?” The expert smiles. “That’s what we’ve been talking about. By insisting on taking a slower, more responsible approach, you’ll instill a new culture of quality within your team.”
And oh, right! I got this NPC to compose a really notable prose-poem to Ariadne. In this case, the setup is mine, but everything after “A poem to Ariadne:” and before “Wow, that is deep stuff” is from the AI (as is the self-deprecating comment at the end):
I found this pretty astounding. Similarly, shorter but in Italian:
I typed the first two lines of Italian (it’s from an actual love song), and everything else is from the AI. Here, it:
- Composed two plausible lines of a love song in Italian,
- Translated all four lines into English,
- Flirtatiously substituted in the name of the NPC it was addressing.
I mean!
Now the impact here would be lessened if we could find somewhere the words from the paean to Ariadne, or the last line of that love poem, somewhere on the web. One of the things that AI Dungeon is especially trained to emulate is apparently a big database of Choose Your Own Adventure narratives, and if those things are in there, it would make the AI’s accomplishment somewhat less impressive. But how likely is it that it contains these particular things?
Given how often a vampire and/or vampire hunter name Count Grey appears in my explorations, I’m betting that the training narratives did have quite a bit of stuff about him in them! Also Warhammer 40k fanfic, as we see in this snippet:
More rofls. (My guess is that I tend to activate the Warhammer 40k stuff because my characters discuss the nature of the universe a lot, and I tend to use angels with flaming swords when a deus ex machina is called for.)
Let’s see, what else to note in this first post? M asked if it knows more historical stuff, so we tested it on “kirtle” and it did a very good job:
And to close, a couple more snippets that appealed to my humor:
Later, assuming I actually do it, more stuff!
Oh, and some random GPT3ish links of possible interest:
- AI Dungeon again (I’ve found the Android app to produce better results than the web version, which is probably a completely false impression as they surely have the same back-end),
- philosopherai.com which will compose essays on pretty much any subject, of varying quality,
- twitter.com/GPT3_ one of a number of GPT-powered Twitter accounts,
- https://www.gwern.net/GPT-3 where an interesting person has posted quite a bit of meaty-looking stuff (unfortunately both-sides-justified on my phone, I’m going to have to try hacking the styles, or use a real computer) both general and specific,
- https://inferkit.com/ which is the successor to the GPT-2 based “Talk to Transformer” that was so popular that the owner had to take it down because they couldn’t afford to pay its fees anymore. I don’t know if this has a GPT-3 version (yet).
And that’s it for now! Enjoy the multiverse.
ceoln 