You are correct according to this tweet [0]. That may become inaccessible as Nitter dies, but the text is:

> Working on an endless crafting game with llama 2

along with a video of this game.

[0] https://nitter.cz/nealagarwal/status/1747284257582506102#m

Off topic, why is nitter dying? I've noticed the main instance's SSL cert is down for a bit now, and other instances are pretty rate limited. Did something happen/change?

All (non-paid) Twitter accounts are now rate-limited, which makes a shared Nitter instance untenable. It's probably still possible to host a personal instance for yourself, although they might ban your account for it.

https://github.com/zedeus/nitter/issues/983#issuecomment-191...

Ahh, neat application of it. Explains the somewhat dubious combinations I was seeing.

Forest + fire = smoke

Mud + water = swamp

Swamp + plant = Venus Flytrap

Okay, okay.

Venus Flytrap + Smoke = smoke detector

<confused smile meme>

Tree + Water = river was also pretty baffling.

Unfortunate. I can see the appeal of using an LLM for this but the results are pretty mediocre.

Tree + Water = River

This is a great result, a branching of water/wood.

low-key genius or high-key misunderstood

As this is powered by an LLM, you are exploring its latent space. That means there isn't one logic behind everything - any association is fair game. Here, probably the strongest one wins.

LLMs are a great way to prompt human ingenuity in mental gymnastics

Great point.

I got “edward” after linking “vampire” and “eclipse” so yeah.

Not sure if this is the point you're making, but that is probably due to "Edward Cullen" of Twilight.

oh my got same

That's weird, I did the same thing but I just got "Twilight".

And "Edward" + "Love" is "Bella".

how do i create love?

To make love you have to put together Prison+Venus= love

If you did that, you'd just get; "Twilight", you'd have to add vampire to Twilight to get Edward.

I also got Dandelion + Engine = Helicopter, along these same lines.

I got s'more + smoke detector = captain hook

Obviously the prompt to the LLM is just to create the most obvious association. It may not mention "crafting" at all. Maybe it does though. Is there something obvious to craft that uses a tree and water in the process?

To design a game like this you need to do a lot better than just creating the obvious association. It needs a mix of obvious recipes and clever recipes, so that there's challenge and a sense of achievement. Also, there's a starting point. What should the graph look like?

I'm sure Neal has done hard work in getting it right.

Weed + tree = bong!

how do you make weed lol

If you look at a river system from space I guess it looks tree-like with branches.

the tree of water is the global river system." very hydrological

Meh, what would be your great response to Tree + Water?

A human can only generate a small fraction of the combinations and would have a hard time coming up with most combinations which are already nonsensical.

What is your non-disappointing idea for, idk, Tears + Pottery (AI: Bowl) or Money + Salt Lick (AI: Cow) or Skull + Lake (AI: Loch Ness) or Dracula + Pirate (AI: Vampirate) or Curse + Money (AI: Debt)? Now do that thousands of more times.

The infinite aspect is the thing that keeps it interesting, I think. The fun is getting a new, weird result like "Dracula" and "Pot of Gold" and seeing if you can generate new weird results from the existing set.

Loch Ness? I skipped to the Loch Ness Monster. Got the Lake only later on. And then I got Nessie.

This is a Quagmire

Porkosaurus, Soup Nazi, Sphinxie, Sodium Chloride, Abdominable Snowman, Baconator and both Yeti and Godzilla. And Yogazilla which is a "First Discovery".

Found myself: Thomas the locomotive

Some of my favorite first discoveries:

- Cactus Missile

- Mammoth Marula

- Super Lizard Wizard

- Cheesethorn Bush

- Cheesnado Lizard Wizard

- Cheesnado 2: Cheesnado Strikes Back

- Cheesewolftrap

I got Gay Bachelorette Party, first discovery

Glitter ketchup, same

Got Pengwynwyrvenpheonix (first discovery) Bird + Dragon, Wyrven | Penguin+Wyrven, Pengwyrven | Pengwyrven + Penguin, Pengwyn | Pengwyn + Pengwyrven, Pengwynwyrven | add a pheonix and there you go

Update : Pengwynwyrvenpheonixtopus is out Sharknado+Multiverse = sharktopus | Pengwynwyrvenpheonix + sharktopus = pengwynwyrvenpheonixtopus

Help what is a Pengwynwyrvenpheonixtopusunicorn

Yeah Loch Ness Monster + Water gave me Nessie

Tree + Water is easy: Mangrove, Bald Cypress, Rhizophora, etc

I wouldn't consider those to be better. You're enhancing the tree side of the equation but gave no examples of the water side.

And why didn't you do the rest of my examples? ;)

Jungle

Forest

Fruit

Nuts

> Tears + Pottery (AI: Bowl)

Urn

Money + Salt Lick makes sense, if it’s a Cash Cow.

Tree + water = canoe

I somehow got to Mega Evolution which I got from megalodon and some pokemon, but don’t ask me how to get to Pokémon.

Anyway, once you get that kind of a modifier, anything’s game for megaification.

Did you figure out how to get math? I’ve been trying so hard!

As usual with language models, you have to put in the work yourself to have fun with them.

I laughed when "Vindaloo + Bubble" gave me "Burp", and "Burp + No Bacon" gave me "Sad"

Ham Solo + Darth Bacon = Pork Star

Ninja + Pig = Pork Chop

The rewilding guys would probably say 'that tracks'. Many of their efforts to get rivers to flow year round usually involve trees. Moss, bugs, rodents and grasses first then trees. Usually can help many areas to have year round streams again. As roots help water linger longer in an area. Which leads to streams.

Window + Dune = Sandwich

That one surprised me.

Sandwich + Sphinx = Sphinxwich

This one delighted me. The Sphinxwich doesn't combine well with other stuff though.

I got Atlantis + Spongebob = Atlantis Squarepantis :D

I found fish + fire = sushi amusing - it isn't necessarily wrong, but it also doesn't feel right either...

How did you get to fish? Because boy do I have a story for you.

Mine starts in Atlantis, then Poseidon gives me a fish. Then two fish turned into a shark and I ended up with a sharknado.

Then I found the titanic, we hit an iceberg, I found a treasure and then pirates chased me, but I got away, sold the treasure for money and became the richest man, then climbed Mt Everest, and later had a tea party.

Anyway, there has to be a better way to get fish than Unda da Sea.

Just add water like 5 times, and you get fish

Fish + fire = sushi

Did you know that Bill Gates is the richest samurai?

Megalopolis + Volcano = Pompeii

Sushi + Asia = Japan

Japan + Destruction = Godzilla

Godzilla + Megalodon = Cthulhu

<drops mic>

Catholic Church + Pride = Pope

I got (hurricane + crocodile = hurricodile) lol

Cthulhu + richer = Richthulhu

My favorite was Lotus Flower + Mud = Buddha

Oh, that's a good one! I'm trying to collect religions and countries.

Lotus+Flower=Buddha Buddha+Religion=Nirvana Nirvana+Music=Kurt Cobain

I got bob marley, reggae, ska, bob marley and the wailers, and flo rida the rapper

North America + Fire = Canada. Naturally.

And if you set Canada on fire?

Maple syrup.

I got swamp with plant + water.

Then swamp + fire = dragon

Bankrupt + Pirate = Captain Jack Sparrow

Bacon + Judaism = No Bacon

holy water + wine = catholic church

I've also got a few where it just mashes adjectives together; so far I've found Time Poseidon, Rainbow Steam Robocloud and Broken Unicorn, among other similar ones.

My best so far is Superninjaghostmansnowghostman

Nuclear power station + Tsunami = Fukushima

Mountain range + Pollen = Hay Fever

volcano + fish = sushi

Smoke detector + fire = Firefighter

how do you get an ICBM? (Inter-Continental-Ballistic-Missile)

Here’s all the combinations I’ve came up with so far: Swamp + Mud = Quagmire Divorce + God = Odd Sun + Hourglass = Time Glass + Hourglass = Time Ice + Oasis = Penguin Sand + Stone = Pyramid Mirage + Time = Illusion Dinosaur + Lightning = Godzilla Oasis + Water = Mirage Egg + Time = chicken Golem + tide = Titan Titan + time = Chronos Poseidon + lighting = Zeus Titan + Chronos = Cronus Time + Fire = Sun Sun * Titan = Apollo Ash + Mud = Clay Godzilla + Love = God ? + ? = Spongebob Unicorn+Gold=Alchemy Unicorn+Alchemy = Philosopher’s Stone Gold+Alchemy=Midas swamp+chicken=duck duck+roast=goose goose+goose=flock flock+wind=flight Narwhal+time=unicorn Lightning + Treasure = lots of stuff (Rich, idk Narwhal+unicorn=narwhalicorn Jonah+time=narwhal Whale+oasis=jonah Plant+seed=tree noah+ark=flood curse+jesus=cross bank+intrest=money dandelion+cactus=desert Back to the future+riddle=time travel back to the future+time=delorian Desert+indiana jones=tresure

Astronomer + Hangover = Astrologer

Nice. At some point I got " Crypto-gangbangasm + Cthulhu Lilith Porn".

My favourite was: "Chuck Norris" + "a-hole" = "dead a-hole"

i got the first discovery porn flower lol

I think it’s the first time AI has made me chuckle. I ended up with “Riddle”, so I combined that with “Tornado” and it gave me “Twister” which I thought was a great Christmas Cracker pun, and then when I combined Riddle with “Bottle” it gave me “Genie”.

just link to the real thing :) [0]

[0]: https://twitter.com/nealagarwal/status/1747284257582506102

Such that it's already dead?

Neat.

Campfire+sushi took about 10 seconds before it gave up and did not combine them.

There's tons of combinations that take forever and nothing ends up happening. That's how I got around to the comment thread (clean+satan is why I'm here): I'm waiting for the latest combination to time out

Similar. Snowmobile Farm and Sandbox Farm both appear to timeout with no response.

However, that said, the idea itself is a neat idea, and could quite easily be turned into game ideas somewhere.

They seem to be asynchronous, e.g. you can actually combine several pairs at the same time.

The request looks like "https://neal.fun/api/infinite-craft/pair?first=Phoenix&secon..." so it's probably typically caching the combination of phoenix+seeds but if there is no cache entry it would use llama to make up something. If there's a lot of attention on the site the llm service might be down or overloaded. And given the exponential/factoral (?) amount of combinations this may be reached surprisingly quickly. Just a guess.

As an aside, the game is technically interesting, being a really simple example of using llm generation for game mechanics. But it is not engaging at all and feels nonsensical to me, especially when compared to little alchemy https://littlealchemy2.com/.

I'm not trying to be negative and this isn't a dig on creativity of the wonderful Neal but more points to the immaturity of llms applied to games, maybe to my overexposure to chatgpt, and maybe a prediction that human touch will always be required to make something entertaining. I'm curious how llms will fit into an engaging game experience in the future.

>As an aside, the game is technically interesting, being a really simple example of using llm generation for game mechanics. But it is not engaging at all and feels nonsensical to me

You just gotta make a game out of it.

For example challenge yourself to try to craft "pizza".

Can even try to do it in as least number of crafts as possible.

Point is, just crafting random things to see what it spits out is OK, but trying to use your own logic to combine things to get to an arbitrary solution you come up with is much more engaging, at least to me.

That's a fairly big challenge since the game gets less coherent the longer it goes on. The early matches generally make sense, but after about 3 levels you start getting loops, and after 5 levels you start getting nonsense or outright failures from queries.

If you figure each of the things is an input parameter to a LLM this makes a lot of sense. They tend to have short memories and struggle with higher level introspection. Great for demos, but fraught with problems when using them to do real work.

Hmm, I’m not finding it to be too big of a challenge.

It’s a bit challenging yeah, but me and my friends are challenging each other to get to words and we can usually find a way to make it.

Things like “Godzilla”, “Universe”, “Vampire”, “Optimus Prime”, “Vodka”, etc are just some examples we did.

I don’t seem to be having problems going dozens of levels deep without loops and not running into many query failures. Results that are deep are still making some logical combinational sense to me at least.

Some words we haven’t been able to make, but that doesn’t mean it’s not possible. It just means we need to get more creative and sometimes think outside the box. There are so many ways you can approach getting to a certain result in my experience so far.

Doing this has been fun and challenging so far for me and my friends FWIW.

I tried your challenge to create pizza. My goal is to get some kind of food, but combining combinations of water, plants, fire, etc are way more likely to produce dragons and universes. I eventually got to chestnut which got to bread, but it was a lot easier to get to "Toast Toast Toast" or "Chestnutzilla" or "Treasure" + "Toast" = "Pirate". I finally got "Tostzilla" which has a pizza emoji, and then "lunch", and "breakfast", and "party"+"toast"="celebration" ?? but it feels random and illogical at some point I just gave up.

So to me it feels like playing against a soulless vector database rather than something engaging and well-crafted. I think what gives me this impression is that things are commonly related to each other using words rather than their meaning -- getting from "pirate" to "captain crunch" to "serial killer" is obviously following lines of language rather than the core concepts that relate objects. This is directly opposed to the actual act of crafting which is 100% rooted in the material world and has no relationship to language.

Maybe I'm losing my imagination, but doing it like you suggest, creating challenges, is makes it more fun. I think I'm just tired of thinking in language.

I'm also seeing a lot of my favorite game creators on twitter enjoying the toy and I'll trust their taste over mine :)

> But it is not engaging at all and feels nonsensical to me, especially when compared to little alchemy https://littlealchemy2.com/.

On the other hand, Little Alchemy doesn't have answers to the most basic combinations. Air + Earth = Dust, but Dust doesn't combine with Water. Earth + Water = Mud, but Mud doesn't combine with Air. Earth + Earth = Land, but Land doesn't combine with Fire.

It may be more sensical since it limits combinations to 0.01% of what's possible, but I don't think that makes it more interesting.

Campfire + Sushi = Fish

Might be worth using farside for the link, which I found out about today on HN.

https://farside.link/https://twitter.com/nealagarwal/status/...

I thought they were using some kind of vector space searches like embedding.. no idea if that's the case

Yeah I'm pretty sure you could do this just with the classic word embeddings (king =queen + man - woman). Maybe it doesn't work as well as with a full LLM.

Addition won't work for things that depend on the order of operations. If salt + water is ocean and water + fire is steam, what's salt + water + fire? Is it salt + steam or ocean + fire?

Associativity and commutivity in vector addition doesn't translate well to semantic meaning. Extrapolating your example, it'd also mean:

    v_king  - v_queen   ~= v_father  - v_mother
    v_king  - v_royalty ~= v_father  - v_parent
    v_king  - v_father  ~= v_royalty - v_parent ~= v_queen - v_mother

I don't see why those should all be true. Intuitively, trying to satisfy O(N^2) semantic pairings with vectors that are optimised for a very specific and different numerical operation (cosine similarity) feels like something that won't work. I'd imagine errors get amplified with 3+ operands.

Isn't the reason for lack of associativity/commutivity is that you're doing operations (addition/subtraction) that have them, and then snapping the result to the closest one of fixed number of points in your output dictionary? The addition is fine, loss of information is in the final conversion.

There's definitely some lossy compression when you snap it to the nearest known vector: enumerating every word ever written in human history wouldn't even come close to the 2^(16*D) representable points for a D-dimensional float16 embedding vector. In fact, even adding two float16 values is a form of lossy compression for most additions.

But I'd be surprised if either of those were the primary reason. The words "sea" and "ocean" are different vectors but they'll be very close to each other. salt + water = sea and salt + water = ocean both sound correct to me so the problem is more about whether the v_salt + v_water can even get to the vicinity of either v_sea or v_ocean.

If we constrain our selves to a pool of words of say Wikipedia entries, minutes names and maybe some other stuff, and use a "super node" like "addition" to kind of act as a math operation.. maybe this makes more sense?

This might explain how I got “Super volcano” and “Supervolcano”

One’s really big, and the other is really cool?

An LLM seems like overkill for a project like this. Why not word2vec?

I had the same thought.

Some of the surprising discoveries I made with word2vec embeddings:

human + robot ~= cyborg

silicon + electricity ~= solar cells

virtual reality + reality ~= augmented reality

As reported here: https://web.archive.org/web/20160806040004if_/http://blog.yh...

how do you get human and robot?

More dimensions?

The open vocabulary aspect seems important. Word2vec would never let me make "Pirate Pope Wars".

I have made a similar thing a few months back, word2vec is inferior because it can't really do compound words, it sometimes returns unintuitive results, and then also you can't prompt it.

I thought this would be very simple to do, so I tried it. https://silphendio.github.io/word2vec_alchemy/

That simple demo took me way too long and all that time is wasted because word2vec simply sucks for this use-case.

I love silly and pointless web pages, here another one: https://sharkle.com

First click brought me right back to neal.fun lol

It seems it doesn't work on my side

I knew this was powered by an LLM once I crafted both "seahorse" and "sea horse"

For me it was when one of my early combinations of Pegasus (might also have been unicorn or flying horse, the latter already being a duplicate of Pegasus) and water became hippocampus, but with the hippopotamus emoji ().

I'm sure it was fun for the creator, bit I'll stick with non - AI games for now.

i really hate it when you are crafting something and get the same thing as another thing in your side bar but it has one CAPITAL LETTER!!! like what that dosnt change anything you fuggly rat

Yes Jimmy I feel the same way say if you have fence and prison it becomes fence I

rag just because of it.

The dumbness of the LLM inference ruined this for me. Fossil + Fire = Dinosaur? Okay.

it also get's caught in stupid loops.

this game is fun for a bit, but shallow once you realize the gimmick.

you can go completely random bullshit go and get "super cyberwolfman zombie phoenixman" that's nonsense, of course it's "first discovery!"

Wouldnt it be worth caching the results? For the first couple of million combinations at least. I suppose that would take away some level of serendipity. But I imagine it would make this a lot cheaper, considering its popularity.