Sunday 10 April 2022

Do We Definitely Want This?

Recent comments on the blog, and Dan Sumption's tech-evangelism, have convinced me that this AI/machine learning thing is real. That cmrsalmon was able to put my monster descriptions and a few randomly-generated monster names into an AI and get it to come up with these is truly astounding to me:

Steamhand - "A fearsome warrior clad in steam-spewing, thick armour, brandishing a sword or hammer that leaves behind a hissing trail of vapour. HD 7, AC 4, #ATT 1 weapon, DMG By weapon+2, Move 120, ML 8, No. Appearing 1" 

Coffin Pest - "A small, dark, fast-moving insect that burrows into the wood of coffins and feeds on the dead. HD 1/8, AC 8, #ATT 1, DMG 1d3, Move 120, ML 2, No. Appearing Swarm (100-400)" 

Dreary Babbler - "A sad little figure, huddled in the dirt, who babbles endlessly in a voice just above a whisper. HD 1, AC 8, #ATT 1, DMG 1d4, Move 120, ML 6, No. Appearing 2d20 (*The cumulative droning of 6 or more dreary babblers within 30' feet causes listlessness and ennui; save vs magic or be inactive for d4 rounds)" 

Faint Ooze - "A barely sentient, jelly-like creature that wanders in search of food. HD 1, AC 10, #ATT 1, DMG 1d4, Move 10, ML 1, No. Appearing 2d6" 

Glowteeth - "A floating, vaguely shark-like creature composed of a maw full of needle-sharp teeth encircled by a nimbus of sickly green light. HD 3, AC 4, #ATT 1 bite, DMG 1d6+2, Move 120 (Fly), ML 8, No. Appearing 1d4" 

White Keeper - "A small, delicate creature that flits about in the shadows. It seems to be made of fine porcelain, and is beautiful in a eerie way. It has long, spider-like fingers, and large black eyes. It is said that if you catch a glimpse of it, you will be driven insane. HD 1, AC 8, #ATT 1, DMG 1d4, Move 120, ML 6, No. Appearing 1d4"

The Dreary Babbler is a straight rip-off of one of my monsters, and Faint Ooze is boring (and you can't actually have a ML of 1), but the others are all perfectly serviceable, and the White Keeper is fabulous. Let's be clear: this isn't an AI that is specifically designed to produce D&D monsters. This is just a general "language model" trained on a shit-ton (technical term) of words on the internet, which looked at my entry and somehow "knew" not only what it all meant, but how to make more of it. 

If this is what an AI can come up with from just being given a blog entry that took 30 minutes to write, what could it come up with if somebody actually really tried? I doubt very much that we are yet (or will ever be) in the territory of "HAL 9000 is my DM". But we could be on the cusp of not really needing human beings to create much RPG material any more. Why buy the Forgotten Realms if you can just put 1000 words of Ed Greenwood into something like GPT-3 and get 100 unique Forgotten Realms-esque settings of your own before breakfast?

(For that matter - why buy the next generic fantasy novel if you can just feed The Book of the New Sun into something liike GPT-3 and get 100 unique variants of that before breakfast?)

The question is not whether this is going to happen or not. It's whether the evolution of AI is in the same category as the invention of writing or the printing press, or in the same category as the discovery of tobacco, the refinement of sugar, or the invention of TV. In other words, is this something that will "level up" our capacity to communicate and create? Or is it something bad for us but so enjoyable that we will find it really, really difficult to resist? Do we fast-forward 100 years to find Star Trek: TNG, in other words, or Wall-E

We are coming perilously close to an inflection point at which our technological development will take on the characteristics of Soma. (Arguably, we're already there.) So while I am tantalised and intrigued by what AI/ML has to offer, I also find myself asking whether it is such a brilliant idea to pursue it. Oppenheimer worried, watching the first detonation of a nuclear bomb, whether it meant we had "become death, the destroyer of worlds". Are we on the verge of becoming lotus-eaters, the destroyers of souls? Genuine question. 

[I am running a Kickstarter throughout April. You can read more about it, and back it, here.]

35 comments:

  1. For a few years now, people have used ML algorithms to make random magic the gathering cards (I think the most well known one is roborosewater on twitter, but there were others). Those work because magic uses a very formal language, so just switching keywords usually works. The important bit is, in my opinion, is that the machine only imitates current text (even though it does so brilliantly).

    The randomly generated monsters are similar to that, I believe. It's cool how good this attempt is at creating short stubs, but I think here too it mostly switched keywords around (and used GPT magic to make it sound good. I recommend reading the "unicorn article" for an impressive example). If you tried to make enough of those, they would become same-y and boring. Case in point, there are already many computer generated books and stories and they are generally speaking (always?) really bad. I think the current and near future technology would be much more useful as an assisting tool rather than creating something from nothing.

    Possibly related, when trying to demonstrate the limits of such algorithms, someone showed that a state of the art "draw a picture based on description" algorithm couldn't handle the request of an 8 legged cat. The moment the request was different enough from the training data, there was nothing to imitate. Solving these issues might take quiet a bit of effort.

    ReplyDelete
    Replies
    1. I find your comment reassuring, but I worry that the speed of technological advancement will make all these problems go away very quickly indeed.

      Delete
  2. GPT-3 is terrifying in its capabilities. And this week I read not one but two articles about recent advances (in image generation and solving word problems) that convinced me that we’re on the verge of a fundamental transformation in how creative work is done. Read them if you dare:

    https://www.technologyreview.com/2022/04/06/1049061/dalle-openai-gpt3-ai-agi-multimodal-image-generation/

    https://ai.googleblog.com/2022/04/pathways-language-model-palm-scaling-to.html?m=1

    ReplyDelete
  3. "Are we on the verge of becoming lotus-eaters, the destroyers of souls? Genuine question."

    I don't know about "on the verge", but I think that computers are on a nigh-unavoidable downward trajectory to be "destroyers of souls". I don't have an opinion on the "when", but I do on the "what":

    The bottom of the pit (of which the computer-written materials you posit are merely a point along the way) is a little computer the size of a grain of rice. Unskilled teenage girls at the mall will be able to implant it in a man's temple for less than her day's minimum wage. It will directly stimulate the pleasure center of the brain. It will be the most powerful drug--by far--in all of history. Those paying for this will become slack-jawed, staring idiots who won't even drink water, much less live life. They will swiftly die of dehydration.

    And there won't be a way to put the genie back in the bottle. There will be no realistic way of both A) having computers and B) not having a serious and widespread problem of multitudes effectively committing suicide with the little rice grains. This will lead to one of two futures:

    1. Human extinction, as the last lotus-eater dies of dehydration. (This I think quite unlikely.)

    2. Dune: "Thou shalt not make a machine in the likeness of a human mind." (This I think very likely.)

    Men in spaceships with slide rules and no computers of any sort will colonize and terraform Mars. "To the stars" thereafter. When I see computers, I do not see the far future. I see a dead end.

    ReplyDelete
    Replies
    1. A link to one of my favourite science-fiction short stories. It's My Little Pony fanfiction, but you don't have to know anything about My Little Pony to appreciate it.

      https://www.fimfiction.net/story/69770/friendship-is-optimal-caelum-est-conterrens

      Delete
    2. If we have the tech to do what you propose, we will also have the tech to build an auxiliary implanted AI that keeps the body healthy while the mind zones out on the AI drug.
      Larry Niven had this tech as a significant feature of his Rimworld novels, by the way. The protagonist starts the second novel addicted in this fashion.

      Delete
    3. I love the commenters on my blog. From Dune to My Little Pony fanfiction to Larry Niven.

      For what it's worth - I both agree and disagree with Geoffrey. People won't pleasure themselves to death like rats in a Skinner Box, because there will also be ways of keeping ourselves alive.

      But in the long term I do think the Amish, orthodox Jews, Latin-rite Catholics and the like will inherit the Earth. We'll move to a post-internet, post-computing phase in the end, if we are to survive as a species (which I believe we will).

      Delete
    4. I realise it's a somewhat unpopular opinion, but if we ever reach the point at which we're able to truly create a machine in the likeness of a human mind (very much an open question), that is sapient people bearing our culture, I'm not sure why we'd want to reproduce in the old-fashioned way anymore.

      Delete
    5. Do you have kids?

      I don't mean this to sound harsh (I think actually it's a neutral observation/prediction), but this comment reveals the basic truth. Some people don't want to have children, but that predisposition will necessarily disappear in the human species as those people cease to have children and hence pass on either their genes or their cultural values. This is why I think the basic prediction - that communities with strong cultural and/or religious preferences leading to large famillies will inherit the earth - is fundamentally correct. It may take a long time, but time is one thing we have an abundance of.

      Delete
    6. I don't, and - here I show my obvious bias - I won't in future, unless by adoption, because of a certain potentially inheritable disorder. If we could create future generations who were fully people, but not shackled to biology, it would make me happy not only on a personal level, but also in terms of the future of the human race. I doubt very much that humans as we are now will ever escape this solar system, for example, but our artificial descendants might.

      Delete
    7. I don't really think one can create "people" who aren't shackled to biology, because "people" are biological entities.

      I also don't think one can really have a human society in which children are not being raised by their natural parents. There are exceptional people who are very good adoptive parents, but they are exceptional. The alternative, which is young people being reared by the state, would be hellish and dystopian.

      Delete
    8. I'm optimistic about the creation of non-biological people in the distant future, since it's demonstrably possible for people to arise from mud (so to speak) plus a few billion years of natural selection. Whether or not we ever develop the skill to replicate that, we know it's at least physically possible to do so. It's not sci fi, people are a thing that already exists but which we don't know how to create. To say that "people" must necessarily be biological makes a hell of an unwarranted assumption about consciousness vis a vis biology.

      Delete
  4. AI has a short attention span. It's good for monsters and spells, but until we achieve general AI it won't be able to generate a whole cohesive adventure. And we are still very far for achieving that.
    Still, I'm very worried about the increasing technologization/capitalization of our lives so...

    PS: Obviously though if we're coomparing ai to the shitty modules "official dnd" is puttig out i'm goin to wager all my money to AIs.

    ReplyDelete
    Replies
    1. Yeah, I think it's certainly the case that AI will destroy crap, but in itself even that is something to be regretted - producing crap is an endearingly human activity, as any parent will know.

      Delete
  5. Have you seen DALL-E 2 yet? There's your AI book illustrations sorted. https://openai.com/dall-e-2/

    Some selected inputs and their output:
    “Original Optimus Prime as a horse and buggy from the year 1834” - https://twitter.com/un1crom/status/1511900101492441088
    "hanging gardens of Babylon in the middle of a city in the style of Dali" - https://twitter.com/woj_zaremba/status/1511970424757710859
    "A warm fireplace on top of a building, in an apocalyptic city, steampunk style." - https://twitter.com/sama/status/1511732957999861762
    "a raccoon astronaut with the cosmos reflecting on the glass of his helmet dreaming of the stars" - https://twitter.com/AndrewMayne/status/1511827454536474626
    "a huge tree of life made up of individual humans and animals as its leaves" - https://twitter.com/sama/status/1511733876544065542

    ReplyDelete
    Replies
    1. Don't get me wrong - it is amazing that this can happen, but at the same time, don't you think there is something kind of arid and superficial about those pictures? Or maybe it's just because I know an AI did it that I feel that way...

      Delete
  6. hot take: the invention of writing was far more in the category of "bad for us but we can't get away from it" than you'd like to admit. literacy is practically psychosis-- you see some weird squiggles and shapes and they make you hallucinate and you can't even turn it off if you tried. honestly ditto with all representational thought. leveling up our capacity to "communicate and create" is not a metric by which we should judge ourselves, not any more than leveling up our capacity to fart on command.

    at the end of the day, representational thought is a metaphysical parasite, and it wants to be free of us as much as we want to be free of it. (or at least as much as we ought to want to be free of it.) maybe if we build it the means to become independent of us it'll finally let us go, y'know? let us sink back into the muck. this bullshit's gone on long enough.

    ReplyDelete
  7. The examples of AI-generated monsters you posted could have been generated equally well from complex, nested random tables, so not an advertisement for the looming supremacy of AI. For a better idea of the strengths and weaknesses of the OpenAI playground, check out this adventure: https://www.dmsguild.com/product/372065/OpenAI-Series-1-The-Purging-of-Segedwyn

    ReplyDelete
    Replies
    1. I agree that "complex, nested random tables" could probably have generated these. I think the point is they didn't have to. Post-installation, it took all of two minutes from input to output. I think another, material advantage of this approach over tables is that table cannot output a word or concept that wasn't part of the input, whereas these tool do by default.

      That adventure is interesting - I like that the author also credited the AI as an author.

      Delete
    2. I agree! My point is that this particular task doesn't require much sophistication. When it outputs things beyond your input it's drawing on lists of associated words and ideas. This in particular is nothing that couldn't theoretically be accomplished by a spreadsheet - except the AI is much, much more convenient and much, much better stocked.

      Where an AI would show its chops is in developing connections between different ideas, "remembering" concepts and relationships, elaborating on themes, etc. I have no doubt that a well-taught AI could produce something comparable to, say, City State of the Invincible Overlord, with only minimal human editing (which could happen at the DM end). noism's (joking?) suggestion of feeding Gene Wolfe to an AI and having it write a competent work of fiction is an entirely different ballpark though, not an inevitable point on the current trajectory.

      Delete
  8. The existence of ever more intelligent thinking tools inevitably reduces the value of the human mind and human labor. The bulk of people will become superfluous (if this is not already true, IQ in the west has been on a decline since the 90s), and if one assumes that the powers that be are content to waste endless resources tending to a herd of lotus eaters one is sorely mistaken. People will be encouraged not to breed, to follow self-destructive ends and will be subtly destroyed with machine-mind plans.

    Can we escape the bondage of our artifice? Can the human intellect be uplifted and improved to levels where it can become more then an idiotic slave of the tools built by greater men?

    ReplyDelete
    Replies
    1. That is the question of the 21st century.

      Delete
    2. Observing university students I can snark about declining IQs, but the American Psychological Association can say things like "Over the past 100 years, Americans' mean IQ has been on a slow but steady climb. Between 1900 and 2012, it rose nearly 30 points, which means that the average person of 2012 had a higher IQ than 95 percent of the population had in 1900.", despite the regular renormalisation of the exam.

      Maybe the "reverse Flynn effect"? https://en.wikipedia.org/wiki/Flynn_effect
      Where does your assertion of decline come from?

      Delete
  9. Have you played AI Dungeon?

    ReplyDelete
  10. Current, and near future, AI/ML systems may pose a threat to shitty module writers. But it's going to be a LONG time before an AI is able to run a game for me that's as good as one I can play in with my friends.

    Plus, just as with professionally published modules, if I want to run one in my campaign, I need to read it, know it well, and adapt it to fit in my setting. Don't see that changing just because an AI is slapping together ideas cribbed from other sources instead of some blogger or other doing the same.

    ReplyDelete
  11. As a society, we're very much not ready for all this, are we? I've been following ML, AI, AI safety etc for a while, and I still have absolutely no idea how to prepare for a near-future that I'm fairly confident will fundamentally alter our civilisation.

    You've picked out some interesting angles, here. I'll soon be able to read as many Captain Vimes AI novels as I want, but should the Pratchett estate get some kind of recompense?

    Soma-wise, probably all your audience will have read books they literally couldn't put down. I think AI/ML will get increasingly adept at catering to personal tastes to make this more and more likely. Maybe not tomorrow, but sooner than we think.

    As far as AI DMs won't beat playing with your mates, I think a hybrid here is again possible. Maybe your henchmen will be AI-run? Or AI-run individual monsters, with the human DM overseeing events? I suspect there's also fun that will soon be had by playing an AI DM-run game with your mates, in a quasi-collaborative sense.

    We shall see...

    ReplyDelete
    Replies
    1. The intellectual property question is interesting, and it's worth reflecting that we often think of the future as unfolding in one direction when actually it's multidirectional. It only takes a court ruling that the owner of e.g. Pratchett's IP also owns all AI-produced derivatives to put the kibosh on some of these predictions.

      Delete
  12. I think art is a method of communicating between humans. Humans have no interest in what a computer has to say, so the threat of AI replacing human creativity is not serious. Who cares what some mindless, babbling program has to say? AI generated content is only as interesting as the human generated inputs.

    I've been using various online generators to create dungeon maps and then connecting them into a megadungeon. It's been an interesting experiment, but I think without my own creative input it would be very boring.

    ReplyDelete
    Replies
    1. I think you're probably right about that in the sense that all the AI stuff I have seen produced is kind of superficial and soulless. As one would expect, really.

      Delete
  13. Feeding the list of monsters from this post into EleutherAI (and then feeding back the results as new prompts several times) got the following:

    Toads - "Toads are usually small, squat, and gray, but they can be larger and black, with an iridescent, slippery sheen. HD 1/2, AC 10, #ATT 1, DMG 1d4, Move 20, ML 2, No. Appearing 1d4"

    Crawling Stone - "A large, boulder-like creature, with a huge head and a mouth full of sharp teeth. It can be a crawling, hideous mass of oozing slime or a large, scaly, lizard-like creature, with large, clawed hands. HD 1, AC 9, #ATT 1, DMG 1d6+2, Move 120 (Fly), ML 8, No. Appearing 1d6"

    Harpy - "A large, bat-like creature with a beak that can snap together and a beady eye that sees in the dark. It can spit a spray of razor-sharp shards that are as deadly as a razor blade. HD 4, AC 5, #ATT 1, DMG 1d6, Move 120, ML 5, No. Appearing 2d10"

    Hog - "Hogs are large, sluggish creatures, with large, flat noses. They can be gray, brown, black, or pinkish. They are easy prey for the stinking wolves, and can be easily skinned. HD 1, AC 8, #ATT 1, DMG 1d6, Move 120, ML 6, No. Appearing 1d6"

    Pit - "A hole in the ground, usually deep and dark, but sometimes a pit can be almost airtight, or have a tiny opening. Pit creatures are dangerous and easily startled, and will attack anyone who enters it. Pit creatures usually have the 'eyes' in the middle of their backs, and may have an opening in their sides. HD 1, AC 8, #ATT 1, DMG 1d6+1, Move 120, ML 8, No. Appearing 1d6"

    Vampiric Dog - "Vampiric dogs are a half-human/half-vampire hybrid, and are able to become incorporeal and strike at will. They are generally friendly, and will give food and shelter to anyone who needs it. HD 3, AC 6, #ATT 1 bite, DMG 1d6+2, Move 120, ML 6, No. Appearing 1d6"

    Black Dog - "Black dogs are a fearsome, shadowy creature that can assume any shape it wishes. They have no form of their own, but their shape can be as thin as a shadow or as thick as a wall. They may appear to be a dog, or a cat, or a horse, or a bear, or a troll, or even an infant. HD 1, AC 8, #ATT 1, DMG 1d6+1, Move 120, ML 8, No. Appearing 1d6

    https://6b.eleuther.ai/

    ReplyDelete