Neural networks are a type of machine learning program that learns from examples they’re given, rather than relying on a human programmer to invent rules.
In an earlier experiment, I trained a neural network to write new names for Dungeons and Dragons spells based on a list of 365 examples. That’s a really small dataset for a neural network to work with, and I ended up struggling to find training parameters that would strike a balance between word-for-word mimicry of the original list of spells, versus a series of completely made-up words. By filtering extensively through the nonsense, I was able to come up with a short list of interesting new spells. (My favorites were Barking Sphere and Gland Growth).
However, blog reader Jo Scott was kind enough to collect the entire 4th edition list of spells - more than 1,300 spells in all. She explained that she’s playing a character who’s an artificer trying to create an autonomous spellcasting golem - essentially, a magical AI - and she’d like to have more weird spells for the golem to invent. (Her Dungeon Master okayed this and thus only has herself to blame when she has to deal with some of the spells listed below.)
Using the new dataset I was able to train a much better-performing neural network. It simply had many more examples of spells to work with; that is, more examples of the words and letter combinations that appear in D&D spells, and thus was able to deduce better rules about how to create them.
For comparison, here’s what the neural network trained on the original spell dataset was producing after it had looked through the spell list 30 times. This is raw, unfiltered output from the neural network.
Original dataset
Wome on frr Eser Wold Sereisk Lelent Warder Cleater Secfen Spiritul Plage Arawen Speak with Alanc Plonting Cloud Aurars Ensntalice Stige Dling Comenthon of Prost Monsen Scink Warrifg Resser RestractiGn Cloud Sreeat Glasp Blenss Bline Ons Dood to Stone
Aside from a couple of spells that just might work, most of the list is magicky-sounding nonsense, sometimes barely pronounceable.
By contrast, this is what the neural network was producing after it had been trained on the dataset that included all the 4th edition spells:
Full dataset
Curse Word Crackling claus Tidal treket Swirk with Wall of Storm Acter Lor distertion Glib ton Grasping Mane Tweel Strike Revitalizing Strike Truneming fortune Fall of the Wild Tunesrite Trickstrak empester Phantasmal assault Tidalt Atight Hadabol Leging Blade Bund Wind Dance of Sack and Prime Poxsare Dumination Mass Cure Fortion
They’re not ALL winners, but the difference is dramatic. This is why, although I can often have fun with small datasets, the really large ones (100,000+ metal bands, or 19,000 IPA beers) tend to produce the most consistently convincing results.
Even this more-sophisticated neural network is not without some oddities. For example, you’ll notice in the results below that it seems to have a particular fondness for bears. And it has invented the name “Dave” which is now, for some reason, its favorite.
I leave you with a selection of Dungeons and Dragons spells generated by the latest neural network.
Mister of Light Storm of the gifling Song of goom Forceful Boor Chorus of the dave Maine storm Frames of Death Song of the doom goom Death’s Death’s Proud Bear Wall of Distraction Date wards Plant of Peace Shield of Farts Song of the darn Ward of Snade the Pood Beast Ice shop Primal Rear Summon Storm Bear Divine Boom Soul of the bill Charm of the dave Spirit of the Spirit Fire shop Song of blord Song of distraction Forceful Force Spirit Boating Song of the ball Hail to the Dave Crusading Disk Summon ass Call to the Daring Treeking of Star Grasping Light Clinging blade Primal Prayer Bear War Cape Find Strike Song of the Unworthy Gate Sail Icon of Thorns Song of the door Star warper Stone of Death Chilled arrow Storm of the dave Fark Mate Charm of the cods Death of the Sun Greater flick Curse Clam Claming Blow Cursing wink Conjure Mare Remorse? Conjure Bark Darkworm Colt Daving fire Healing of Bat Mordenkainen’s lucubrabibiboricic angion
“However, Kosinski tells The Economist, the research is not intended to be used to profile or “out” homosexual men and women. Rather, it is designed to demonstrate—or even warn—that technological advances can be used for such means, and could pose a threat to our privacy as information is so much more easily accessible digitally.“
criticism of this study aside how are you gonna invent technology literally centered around identifying gays and then say you only did it as a warning to show people you can lmfao. like “hey guys, technology is advancing to the point that we can invasively out people and that’s bad. i know this because i built it myself”
