I find myself more intrigued by Infinite Craft's inner workings than the actual act of playing it. My first thought when I saw a screenshot of its mechanics via social media was "isn't that just that one mobile game from when those were new?" and sure enough after wracking my brain trying to remember the name I found that I was thinking of Doodle God, an iphone game from the first wave of popular mobile titles like angry birds, fruit ninja, etc
The concept is similar but Infinite Craft is much more an unstructured sandbox than DG. Its concept is simplicity itself, you combine words with other words to make new ones via some method of semantic association and formula for determining combinations through generative AI. The claim is that this collection of words and sentences is truly infinite and I believe it. As silly as the game is in concept, under the surface its essentially an application of natural language processing, which is an absolutely mindmelting field where computer science, linguistics and philosophy all meet to make the most impenetrable theoretic framework to describe how a system can be devised to make unicorn + death = dead unicorn.
I should preface the rest of this review by saying that I am very much not an expert by any means and that you should look into this yourself. In fact, NLP was the subject which made me drop out of my Computer Science Degree because I hated it so much (well, it also didn't help that this was during Covid so we had to learn this shit remotely), but I know a little bit more than most.
The main problem that NLP needs to "solve" is that a computer doesn't know natural languages (i.e english, mandarin, russian etc) but we do. Hence in order to analyse anything relating to natural language we need to make systems to make it parseable by an algorithm. Its one of those things where its such an integral and unconscious part of human cognition that you don't realize how complex of a task it is to understand words and phrases. One task we got early on was analysing a corpus of Amazon Reviews and determining which reviews were positive and which were negative. Even before any kind of analysis of the review as a whole can be performed you need to parse the whole thing and tokenise the individual words (easy in english admittedly, its just spaces) but then you also need to analyse the individual words and perform Word Sense Disambiguation i.e if the word smith appears in the review you need to determine using adjacent words if it means the name smith or the profession smith or the act of smithing. Even further you have stuff like entity linking, where certain words form part of a broader entity for e.g Duke of Parma refers to a single thing rather than 3 separate unrelated words. It was quite frankly, a nightmare which was nevertheless pretty interesting to learn about. Thankfully we had plenty of resources to tackle these issues (computer science is one of the most "standing on the shoulders of giants" discipline there is) and one resource I was reminded of when playing Infinite Craft was WordNet.
Wordnet is a lexical database, which is a fancy way of saying essentially a beefed up dictionary. Containing not only words and their meanings but also its relations in terms of hypernyms and hyponyms. I.e Amphibian is a Hypernym for Frog, and Oak is a Hyponym of Tree. You can also then make connections between words if they share Hypernyms (i.e we can tell the system that Crocodile, Alligator and Komodo dragon are all "coordinate terms" because they share the hypernym Reptile) and a whole bunch of other things that would take forever to fully explain. In essence, its a database that helps us determine semantic fields for words by essentially outsourcing the work to pre-made materials by humans. Another example would be SimLex-999 which asked people to rank word-pairs between 0-10 in terms of how related they were to each other and produced a dataset with these word pairs.
This is all to say, analysing language such as english is a herculean task, and I hope one day we get a peek behind the curtain at how the language model used by infinite craft works, because it really is quite interesting. The technology is still nevertheless quite rough at times; which can be forgiven mostly because the freeform sandbox nature of its mechanics makes its imprecision and unknowability kind of a non-issue. Hence, all the screenshots of funny combinations of pop culture properties and not the more banal or incoherent combinations that are swiftly forgotten as a new possibility occurs. If it were a more structured set of challenges like Doodle God you would swiftly get blocked by the eternal "point and click" problem of having to think about the one solitary solution and combination which made sense to the designer of combining glue + mummy to make papier maché. So what if most combinations are not particularly engaging or don't produce progress, the more words you make the more words you CAN make through combining previous words with older ones. Sometimes this runs into the issue that certain strings of words presumably associated by the game as part of the same semantic field seem to chain back into themselves. I.e fire + ash makes lava, lava + mud = volcano, volcano + fire makes ash etc. Another issue is, some combinations simply didnt make sense, but of course no system could ever produce infinitely many combinations which would seem reasonable to all people, especially when getting to more abstract concepts. Idk what a dead unicorn + candy would make if they were magically sealed together, but I guess "zombie unicorn" is as good an answer as any.
Some words cannot be combined at all, I guess we really are stuck with the "I don't know how to do that" problem of adventure games once more.
Infinite Craft is more of a novelty than a game one would feasibly dedicate themselves to, by a developer who seems quite married to the idea of short, communal browser based experiences. Their website is definitely worth a look beyond IC, its a lovely collection of experiments.
There is one aspect of the game I appreciated, despite the simplicity, Infinite Craft provides a sort of word based micro-storytelling, similar to other implementations of generative AI based on user input. I remember reading an interview given by Gareth Damian Martin, the Citizen Sleeper and In Other Waters dev where they said about their work on "procedural poetry" that even seemingly random or computer generated sequences form a broader sequence (like a story) by Humans' tendencies to fill in the blanks and assign patterns to those which seemingly have none. I think we've all extrapolated seemingly human charcteristics and motivations to NPCs based on certain intended and unintended behaviours. Or even simpler than that, ask a pet owner what their dog/cat's personality is, they will go wild anthropomorphising the animal's thought process to explain their every action.
Tree of Life + Tree of Death = Tree of Knowledge
Tree of Knowledge + Party = Adam and Eve
2 Equations to summarise one of the key episodes of Genesis in christian lore which would not be nearly as impressive if a human had manually come up with such a combination, nor if I hadn't spent most of my time getting much more pedestrian or nonsensical combinations like candy + fisherman = sugar daddy.
The concept is similar but Infinite Craft is much more an unstructured sandbox than DG. Its concept is simplicity itself, you combine words with other words to make new ones via some method of semantic association and formula for determining combinations through generative AI. The claim is that this collection of words and sentences is truly infinite and I believe it. As silly as the game is in concept, under the surface its essentially an application of natural language processing, which is an absolutely mindmelting field where computer science, linguistics and philosophy all meet to make the most impenetrable theoretic framework to describe how a system can be devised to make unicorn + death = dead unicorn.
I should preface the rest of this review by saying that I am very much not an expert by any means and that you should look into this yourself. In fact, NLP was the subject which made me drop out of my Computer Science Degree because I hated it so much (well, it also didn't help that this was during Covid so we had to learn this shit remotely), but I know a little bit more than most.
The main problem that NLP needs to "solve" is that a computer doesn't know natural languages (i.e english, mandarin, russian etc) but we do. Hence in order to analyse anything relating to natural language we need to make systems to make it parseable by an algorithm. Its one of those things where its such an integral and unconscious part of human cognition that you don't realize how complex of a task it is to understand words and phrases. One task we got early on was analysing a corpus of Amazon Reviews and determining which reviews were positive and which were negative. Even before any kind of analysis of the review as a whole can be performed you need to parse the whole thing and tokenise the individual words (easy in english admittedly, its just spaces) but then you also need to analyse the individual words and perform Word Sense Disambiguation i.e if the word smith appears in the review you need to determine using adjacent words if it means the name smith or the profession smith or the act of smithing. Even further you have stuff like entity linking, where certain words form part of a broader entity for e.g Duke of Parma refers to a single thing rather than 3 separate unrelated words. It was quite frankly, a nightmare which was nevertheless pretty interesting to learn about. Thankfully we had plenty of resources to tackle these issues (computer science is one of the most "standing on the shoulders of giants" discipline there is) and one resource I was reminded of when playing Infinite Craft was WordNet.
Wordnet is a lexical database, which is a fancy way of saying essentially a beefed up dictionary. Containing not only words and their meanings but also its relations in terms of hypernyms and hyponyms. I.e Amphibian is a Hypernym for Frog, and Oak is a Hyponym of Tree. You can also then make connections between words if they share Hypernyms (i.e we can tell the system that Crocodile, Alligator and Komodo dragon are all "coordinate terms" because they share the hypernym Reptile) and a whole bunch of other things that would take forever to fully explain. In essence, its a database that helps us determine semantic fields for words by essentially outsourcing the work to pre-made materials by humans. Another example would be SimLex-999 which asked people to rank word-pairs between 0-10 in terms of how related they were to each other and produced a dataset with these word pairs.
This is all to say, analysing language such as english is a herculean task, and I hope one day we get a peek behind the curtain at how the language model used by infinite craft works, because it really is quite interesting. The technology is still nevertheless quite rough at times; which can be forgiven mostly because the freeform sandbox nature of its mechanics makes its imprecision and unknowability kind of a non-issue. Hence, all the screenshots of funny combinations of pop culture properties and not the more banal or incoherent combinations that are swiftly forgotten as a new possibility occurs. If it were a more structured set of challenges like Doodle God you would swiftly get blocked by the eternal "point and click" problem of having to think about the one solitary solution and combination which made sense to the designer of combining glue + mummy to make papier maché. So what if most combinations are not particularly engaging or don't produce progress, the more words you make the more words you CAN make through combining previous words with older ones. Sometimes this runs into the issue that certain strings of words presumably associated by the game as part of the same semantic field seem to chain back into themselves. I.e fire + ash makes lava, lava + mud = volcano, volcano + fire makes ash etc. Another issue is, some combinations simply didnt make sense, but of course no system could ever produce infinitely many combinations which would seem reasonable to all people, especially when getting to more abstract concepts. Idk what a dead unicorn + candy would make if they were magically sealed together, but I guess "zombie unicorn" is as good an answer as any.
Some words cannot be combined at all, I guess we really are stuck with the "I don't know how to do that" problem of adventure games once more.
Infinite Craft is more of a novelty than a game one would feasibly dedicate themselves to, by a developer who seems quite married to the idea of short, communal browser based experiences. Their website is definitely worth a look beyond IC, its a lovely collection of experiments.
There is one aspect of the game I appreciated, despite the simplicity, Infinite Craft provides a sort of word based micro-storytelling, similar to other implementations of generative AI based on user input. I remember reading an interview given by Gareth Damian Martin, the Citizen Sleeper and In Other Waters dev where they said about their work on "procedural poetry" that even seemingly random or computer generated sequences form a broader sequence (like a story) by Humans' tendencies to fill in the blanks and assign patterns to those which seemingly have none. I think we've all extrapolated seemingly human charcteristics and motivations to NPCs based on certain intended and unintended behaviours. Or even simpler than that, ask a pet owner what their dog/cat's personality is, they will go wild anthropomorphising the animal's thought process to explain their every action.
Tree of Life + Tree of Death = Tree of Knowledge
Tree of Knowledge + Party = Adam and Eve
2 Equations to summarise one of the key episodes of Genesis in christian lore which would not be nearly as impressive if a human had manually come up with such a combination, nor if I hadn't spent most of my time getting much more pedestrian or nonsensical combinations like candy + fisherman = sugar daddy.
judgerussell
2 months ago