Tried it! I really like the idea, but I think the clue generation could use some work. Every clue ended in "in games", and honestly most of them were not really game related to start with. For example the clue "Place in games where characters go to rest and replenish health or mana" had the solution "bar"... which I wouldn't describe as right. Similarly "The name of a popular character who may need rescuing in some games" was "Emily".
I think it might be worth working on prompting to make sure the answer is a unique solution to the hint (or at least closer to unique). What model are you using here?
Thank you for trying.
You are absolutely right.
In my defense i released the very first 20 puzzles without proof reading them. I just wanted to see what the AI can deliver as a starting point and get an idea if crossword players would like or hate the general idea of it. I've just started seriously playing it myself and some clues are indeed strange. The clues were generated by gpt-4o. You can try later puzzles i did some prompt adjustments because i noticed the forced "in games" after 10 puzzles or so. It gets a bit better at later puzzles. Thank you for the feedback
Really goes to show how bad top-tier LLM's are at rather basic tasks like creating a clue out of popular media. This should be among the core competencies of major models with tons of available training data and simple summarization requests.
LLM Hallucinations are real. Admittedly for the prompt I provided just the word and some basic rules. It should be possible to increase the quality if e.g. I provide the sentence in which the word was used to the LLM. Nevertheless Hallucinations will always be a problem i think there is no way around a human quality gate in the process?
Nice work. I've also experimented with procedurally generated crossword puzzles, though I really wanted to constrain them to symmetric layouts like what you would find in the New York times which made it more difficult.
There's an outstanding issue and that is (from what I can tell) at least 75% of the answers correspond to relatively generic nouns or verbs.
Part of the deep satisfaction in solving a crossword puzzle is the specificity of the answer. It's far more gratifying to answer a question with something like "Hawking" then to answer with "scientist", or answering with "mandelbrot" versus "shape".
It might be worth going back and looking up a compendium of games released in the last couple decades, cross referencing them with their manuals, GameFaqs, etc. and peppering this information into the crossword.
Thank you for your input. Your conceptual point on what makes crosswords fun is gold! Your practical advice is not only great but also very doable: Remember I used a 3.7million Reddit-Comments dataset which is a sub-optimal choice to generate words out of. (But i liked the headline behind it). So changing the dataset to something you mentioned (game compendiums, manuals, faqs) will probably increase the answers significantly.
I love crossword puzzles, and this is very well done! A while back I made a crossword puzzle generator using constraint programming. It is definitely way slower than whatever heuristic you are using, but it is very good at generating dense crossword puzzles using small word lists. This would be useful if you were interested in making a crossword puzzle using words related to a single game, for instance. You can read more on it here if you are interested: https://pedtsr.ca/2023/generating-crossword-grids-using-cons...
Thank you for sharing. I will check it out. And no worries regarding the speed. The website is cheating the 20 puzzles are pre-compiled so to say, the actually work was done in C++, the web server is just indexing an array of perfect puzzles at that point. TY!
The 13th puzzle is probably the first decent one. I still can't quite put together how it's getting the clues in some cases (BMing = using pedals? noly = Just this, no other?) but it's certainly the first one that's within reach of being reasonable.
TYVM for this feedback. I feel the exactly same way. This is purely a NON-PROOF-READ starting point, a field test of what AI can put out unsupervised. The prompt was changed in the latter puzzles those should be a tad better. But as already discussed there can be done a lot more to improve the quality of words and clues significantly.
>5. Grid Generation: Implemented a heuristic crossword algorithm to create grids and place words efficiently.
I always think about doing something similar for a similar project. Are you able to do it completely automatically or do you have to help finesse the words to fit?
Haha, true. I was sure it is "Inn" from Hearthstone Innkeeper but that's a case of LLM Hallucinations. See my other comment on how this can be improved but not solved completely imo
> 1: Allows movement between different areas in a game
> Unportal
what?
>19: An online forum section where gamers come together to discuss specific topics
> ITT
I don't think that's right...
Most of the questions just seem like normal crossword questions, but with the term "in games" added to it.
I'm not gonna sugarcoat it: this sucks. The crossword grids often have totally isolated words. 1 across and 2 down start with the same letter. The questions are nonsensical. I'd hardly go so far as to call it an interesting proof of concept.
Thank you for trying! Please see my reply on the other comment it fits here as well. This is a non proof read version and i agree with you consider this a starting point. Also feel free the try out the later puzzles (10+) i changed the prompt at some point because i noticed the same thing.
Edit: At some point I was unsure myself, if i don't know certain things about gaming or if the AI is making things up, haha, so this feedback is extremely valuable for me thank you
TYVM for trying it out and feedback and it confirms what other users are saying. This is an unsupervised version with no human quality gate. Please consider this a starting point. That being said, I feel bad that you couldn't even enjoy it, that was not my goal. Sorry for that. I thought people would take a smile away regardless of some words/clues being bad. I have to admit i underestimated the appeal of quality. Some people say it gets at least a little bit better from puzzle #13, maybe you can get some enjoyment out of that? No worries if you are fed up at this point. Thanks for trying it out
Thank you! You right, no i18n yet, I can assure you that the URL https://capsloq.de/crosswords/joystick-jargon has the only playable crosswords at this moment. The other links in the Nav are not crossword related. Will definitely do i18n maybe even the crossword themselves in the future.
Unportal? What the fuck is unportal? I have played games for 30 years and I've never heard the term "unportal". Google gives no useful results. That clue made me angrier than any crossword clue I've ever seen.
One more thing: Besides the obvious that clues and answers should make sense can you provide some insights in what makes a good crossword to you? I'd love to hear that from someone playing them for 30 years!
TYVM for trying it out! Oh no, now I feel bad again. Please remember these are unsupervised generated clues by a Large Language Model (openai). More of a field test, however I now understand how serious you and others take crosswording. I vow to do better for the next crossword puzzles. It's not my intention to make people feel bad, sorry!
Tried it! I really like the idea, but I think the clue generation could use some work. Every clue ended in "in games", and honestly most of them were not really game related to start with. For example the clue "Place in games where characters go to rest and replenish health or mana" had the solution "bar"... which I wouldn't describe as right. Similarly "The name of a popular character who may need rescuing in some games" was "Emily".
I think it might be worth working on prompting to make sure the answer is a unique solution to the hint (or at least closer to unique). What model are you using here?
Thank you for trying. You are absolutely right. In my defense i released the very first 20 puzzles without proof reading them. I just wanted to see what the AI can deliver as a starting point and get an idea if crossword players would like or hate the general idea of it. I've just started seriously playing it myself and some clues are indeed strange. The clues were generated by gpt-4o. You can try later puzzles i did some prompt adjustments because i noticed the forced "in games" after 10 puzzles or so. It gets a bit better at later puzzles. Thank you for the feedback
Really goes to show how bad top-tier LLM's are at rather basic tasks like creating a clue out of popular media. This should be among the core competencies of major models with tons of available training data and simple summarization requests.
LLM Hallucinations are real. Admittedly for the prompt I provided just the word and some basic rules. It should be possible to increase the quality if e.g. I provide the sentence in which the word was used to the LLM. Nevertheless Hallucinations will always be a problem i think there is no way around a human quality gate in the process?
No it shows how important good prompting techniques are.
Haha, for those two you mentioned I assumed it was 'Inn' and 'Zelda'. I don't even know who Emily is.
And googling 'Emily video game character' didn't bring up any noticeably popular video game characters.
I put "hub" and "Peach"... There seems to be a lot of possible answers
Nice work. I've also experimented with procedurally generated crossword puzzles, though I really wanted to constrain them to symmetric layouts like what you would find in the New York times which made it more difficult.
There's an outstanding issue and that is (from what I can tell) at least 75% of the answers correspond to relatively generic nouns or verbs.
Part of the deep satisfaction in solving a crossword puzzle is the specificity of the answer. It's far more gratifying to answer a question with something like "Hawking" then to answer with "scientist", or answering with "mandelbrot" versus "shape".
It might be worth going back and looking up a compendium of games released in the last couple decades, cross referencing them with their manuals, GameFaqs, etc. and peppering this information into the crossword.
Thank you for your input. Your conceptual point on what makes crosswords fun is gold! Your practical advice is not only great but also very doable: Remember I used a 3.7million Reddit-Comments dataset which is a sub-optimal choice to generate words out of. (But i liked the headline behind it). So changing the dataset to something you mentioned (game compendiums, manuals, faqs) will probably increase the answers significantly.
I love crossword puzzles, and this is very well done! A while back I made a crossword puzzle generator using constraint programming. It is definitely way slower than whatever heuristic you are using, but it is very good at generating dense crossword puzzles using small word lists. This would be useful if you were interested in making a crossword puzzle using words related to a single game, for instance. You can read more on it here if you are interested: https://pedtsr.ca/2023/generating-crossword-grids-using-cons...
Thank you for sharing. I will check it out. And no worries regarding the speed. The website is cheating the 20 puzzles are pre-compiled so to say, the actually work was done in C++, the web server is just indexing an array of perfect puzzles at that point. TY!
The 13th puzzle is probably the first decent one. I still can't quite put together how it's getting the clues in some cases (BMing = using pedals? noly = Just this, no other?) but it's certainly the first one that's within reach of being reasonable.
TYVM for this feedback. I feel the exactly same way. This is purely a NON-PROOF-READ starting point, a field test of what AI can put out unsupervised. The prompt was changed in the latter puzzles those should be a tad better. But as already discussed there can be done a lot more to improve the quality of words and clues significantly.
Great effort!
Clickable link: https://capsloq.de/crosswords/joystick-jargon
thank you very much! And thanks again for the clickable link i didn't even notice
>5. Grid Generation: Implemented a heuristic crossword algorithm to create grids and place words efficiently.
I always think about doing something similar for a similar project. Are you able to do it completely automatically or do you have to help finesse the words to fit?
The algo does allow for fully automatic crossword generation, let me try to summarize the general flow:
1. Begins with an empty grid and starts placing words horizontally from the top-left corner
2. For each word placement, it verifies that valid words can be formed vertically at each intersection point
3. It maintains a list of possible letters for each cell to ensure all constraints are satisfied
4. The generator consults a dictionary to find valid words that fit the curent grid state, allowing for diverse solutions
5. If no valid word can be placed, it may decide to insert a black square, carefully checking that doesn't violate any crossword rules
6. When it reaches an dead end, the system backtracks and tries different options
7. It employs smart heuristics to guide word selection, such as favoring longer words in certain positions
8. Throughout the process it automatically adjusts parameters like word length andblack square placement to find a valid solution
There is no manual intervetion, however the quality depends heavily on the input dictionary and tunable parameters.
> Place in games where characters go to rest and replenish health or mana
> Bar
Very wrong, unless we’re simulating what someone unfamiliar with the gaming jargon would think - in which case very accurate.
Haha, true. I was sure it is "Inn" from Hearthstone Innkeeper but that's a case of LLM Hallucinations. See my other comment on how this can be improved but not solved completely imo
> 1: Allows movement between different areas in a game
> Unportal
what?
>19: An online forum section where gamers come together to discuss specific topics
> ITT
I don't think that's right...
Most of the questions just seem like normal crossword questions, but with the term "in games" added to it.
I'm not gonna sugarcoat it: this sucks. The crossword grids often have totally isolated words. 1 across and 2 down start with the same letter. The questions are nonsensical. I'd hardly go so far as to call it an interesting proof of concept.
Thank you for trying! Please see my reply on the other comment it fits here as well. This is a non proof read version and i agree with you consider this a starting point. Also feel free the try out the later puzzles (10+) i changed the prompt at some point because i noticed the same thing.
Edit: At some point I was unsure myself, if i don't know certain things about gaming or if the AI is making things up, haha, so this feedback is extremely valuable for me thank you
Tried it and think your approach is really cool.
Sadly, the clues and the words relating to them feel off, making the whole game rather unenjoyable.
TYVM for trying it out and feedback and it confirms what other users are saying. This is an unsupervised version with no human quality gate. Please consider this a starting point. That being said, I feel bad that you couldn't even enjoy it, that was not my goal. Sorry for that. I thought people would take a smile away regardless of some words/clues being bad. I have to admit i underestimated the appeal of quality. Some people say it gets at least a little bit better from puzzle #13, maybe you can get some enjoyment out of that? No worries if you are fed up at this point. Thanks for trying it out
Why is there no 1 across on the puzzle? There are clues for 1 across but not on the "board".
which puzzle number are you refering to?
I think i just saw it Puzzle #1 Clue #1. That's a bug in the numbering algorithm thanks for point it out.
It happens on puzzles 1, 2, 3, 6, 7, 11, 12, 13, 14, 15, 16, 17, 18, 19 and 20.
You right. Thank you for debugging that. And it's always the numbering 1. Definitely a bug in the algo. Will look into it
Awesome job! Also, might I suggest a bit of i18n? I can't read German..
Thank you! You right, no i18n yet, I can assure you that the URL https://capsloq.de/crosswords/joystick-jargon has the only playable crosswords at this moment. The other links in the Nav are not crossword related. Will definitely do i18n maybe even the crossword themselves in the future.
Unportal? What the fuck is unportal? I have played games for 30 years and I've never heard the term "unportal". Google gives no useful results. That clue made me angrier than any crossword clue I've ever seen.
One more thing: Besides the obvious that clues and answers should make sense can you provide some insights in what makes a good crossword to you? I'd love to hear that from someone playing them for 30 years!
TYVM for trying it out! Oh no, now I feel bad again. Please remember these are unsupervised generated clues by a Large Language Model (openai). More of a field test, however I now understand how serious you and others take crosswording. I vow to do better for the next crossword puzzles. It's not my intention to make people feel bad, sorry!
AI hallucination