becomingsoup

Ari's Art Blog

31 | they/them | eats too much soup and probably will turn into it somedaykind of (very!!) obsessed with robotsInsta @becomingsoup (I post art there more often)

960 posts

AIs Named By AIs

AIs named by AIs

Neural networks can be good at naming things, I’ve discovered. Recently I’ve been experimenting with a neural network called GPT-2, which OpenAI trained on a huge chunk of the internet. Thanks to a colab notebook implementation by Max Woolf, I’m able to fine-tune it on specific lists of data - cat names, for example. Drawing on its prior knowledge of how words tend to be used, GPT-2 can sometimes suggest new words and phrases that it thinks it’s seen in similar context to the words from my fine-tuning dataset. (It’ll also sometimes launch into Harry Potter fan fiction or conspiracy theories, since it saw a LOT of those online.)

One thing I’ve noticed GPT-2 doing is coming up with names that sound strangely like the names of self-aware AI spaceships in Iain M. Banks’s Culture novels. In the science fiction series, the ships choose their own names according to a sort of quirky sense of humor. The humans in the books may not appreciate the names, but there’s nothing they can do about them:

Hand Me The Gun And Ask Me Again Zero Credibility Fixed Grin Charming But Irrational So Much For Subtlety Experiencing A Significant Gravitas Shortfall

Now compare some of the effects pedals GPT-2 came up with:

Dangerous But Not Unbearably So Disastrously Varied Mental Model Dazzling So Beautiful Yet So Terrifying Am I really that Transhuman Love and Sex Are A Mercy Clause

And some of the cat names:

Give Me A Reason Thou Shalt Warning Signs Kill All Humans

Did GPT-2 somehow have a built-in tendency to produce names that sounded like self-aware spaceships? How would it do if it was actually trained specifically on Culture ships?

A reader named Kelly sent me a list of 236 of Iain M. Banks’s Culture ship names from Wikipedia, and I trained the 345 million-parameter version of GPT-2 on them. As it turns out, I had to stop the training after just a few seconds (6 iterations) because GPT-2 was already beginning to memorize the entire list (can’t blame it; as far as it was concerned, memorizing the entire list was a perfect solution to the task I was asking for).

And yes. The answer is yes, naming science fiction AIs is something this real-life AI can do astonishingly well. I’ve selected some of the best to show you. First, there are the names that are clearly warship AIs:

Not Disquieting At All Surprise Surprise And That’s That! New Arrangement I Told You So Spoiler Alert Bonus Points! Collateral Damage Friendly Head Crusher Scruffy And Determined Race To The Bottom

And there are the sassy AIs:

Absently Tilting To One Side ASS FEDERATION A Small Note Of Disrespect Third Letter of The Week Well Done and Thank You Just As Bad As Your Florist What Exactly Is It With You? Let Me Just Post This Protip: Don’t Ask Beyond Despair Way Too Personal Sobering Reality Check Charming (Except For The Dogs)

The names of these AIs are even more inscrutable than usual. To me, this makes them much scarier than the warships.

Hot Pie Lightly Curled Round The Wrist Color Gold Normally Comes With Silence 8 Angry Doughnut Feelings Mini Cactus Cake Fight Happy to Groom Any Animals You Want Stuffy Waffles With Egg On Top Pickles And Harpsichord Just As Likely To Still Be Intergalactic Jellyfish Someone Did Save Your Best Cookie By Post-Apocalyptic Means LGRPllvmkiqquubkhakqqtdfayyyjjmnkkgalagi'qvqvvvvvvvvvvvvvvvv

At least it does sound like some of these AIs will be appeased by snacks.

Bonus content: more AI names, including a few anachronisms (“Leonard Nimoy for President” for example)