Semantle: Can you think like a computer?

*have had

Semantle author here! forkclovetofu invited me in. First, thanks to everyone who is playing! It always gives me a smile to see something I make get used. I plan to continue to make updates to Semantle over the course of at least the next few weeks. I am especially happy to make minor usability and legibility fixes, which don't require as much thought.

New puzzles appear at midnight UTC, which early evening here in NYC.

A similarity hint for today's word is (or should be) near the top of the page. I've just added some bold to make it a little easier to see.

A 60% similarity is fairly typical for the closest word. If you think about it, this makes sense: each word is represented as 300 numbers, and with 300 numbers, there are a lot of ways to be different. Someone in the previous thread complained about the low similarity of "elevator" and "lift". I understand that complaint -- but there is really only one context in which they are similar: up-going boxes. By contrast, "lift" can also refer to "bro, do you even lift?" -- a totally different part of speech. "Mood lift" is some sort of scam drug; "mood elevator" is some sort of scam book. "elevator music" seems to be much more common a phrase than "lift music". And so forth.

I hope to add one more wacky hinting mechanism, but it's kind of complicated and it's not clear that it will be at all useful. Probably later this week, or over the coming weekend.

Hey David, welcome...
Sounds like an interesting project, I have been unable to play on my Android phone or my iPad, but I enjoyed reading the description... :)

Works on Android (Pixel 5, Firefox and Chrome) for me, and iPhone for my wife. If you give me more info about what sort of breakage you're seeing, I might be able to fix it.

well fuck, that's my first time getting the word. 46.

Measuring the "similarity" of words is a very squishy concept that could encompass a large number of concepts of what similarity means. But software deals in algorithms, not concepts, and if a programmer tried to implement more than one similarity measurement they would need to apply a method for resolving the unavoidable conflicts that would arise from applying all of them to the same case while each delivering a different numeric 'similarity' weight. All that programming would get messy in a hurry.

This leads me to think the basis for judging "similarity" and attaching a numeric weight to it would be the product of probably just one measurement, at most three. Figure out what drives the algorithm and you should be able to reduce the solution to an optimal strategy you could counter-program in your own computer. Creating such a program would literally exemplify "thinking like a computer". My guess is that somebody will do just that and the game will be over.

Thanks David, I just type a word in, hit guess, and then nothing happens... no getting close meter, no rank, no guess count... (it's just the Samsung browser on my phone, and Safari on the iPad...)
Anyway, didn't mean to turn this into tech support, please don't sweat it...

his leads me to think the basis for judging "similarity" and attaching a numeric weight to it would be the product of probably just one measurement, at most three. Figure out what drives the algorithm and you should be able to reduce the solution to an optimal strategy you could counter-program in your own computer.

the 'algorithm' here is a neural net, they are famously not amenable to reverse engineering and as david said each word is represented as 300 numbers.

Started to maybe get the hang of it today, my first not cold guess was only 303/1000, it could have led me to the answer quickly but I ended up going a completely different route and got it in 122.

121. Holy shit, I’m gonna waste a lot of time with this.

I am giving up pretty early until I kind of understand what is going on.

The interface is definitely much easier to use on the computer than the (Android) phone. The enter box is tiny on the phone and if I zoom in enough to see it, the word list and leading text is gigantic. But this has promise as one of those quick "brain break" games for me - so much less autopilot than Two Dots or any of the Candy Crush variants.

I love the idea of this game, but I'm just not getting it. I've given up in frustration both days so far, and playing around with it further after knowing the word only confuses me more about what's considered "close".

I am feeling that too, only I somehow managed not to get frustrated...yet.

what factors are considered in the comparison? Or is the fun lying in not knowing? I gave up too fwiw

Got it in 21 today! Much better.

I don't really have any advice for those who are struggling with it. I tend to start with a few guess words that are as semantically different as I can think of at the time and see what's closest, and then progress from there.

here's some info on Word2Vec, which is the underlying algorithm
https://radimrehurek.com/gensim/auto_examples/tutorials/run_word2vec.html
https://medium.com/@zafaralibagh6/a-simple-word2vec-tutorial-61e64e38a6a1

m0stly clean (Slowsquatch), what word were you trying? Someone just pointed out a bug with that symptom with "slice" (and about a hundred other misc words), which I just fixed.

If anyone wants to know what "different" (as opposed to "similar") looks like on a related dataset, try https://www.datcreativity.com/task (not mine -- I'm just a fan).

It doesn't know the word aeroplane

My first time playing it today and can't seem to get closer than 824 (32.73% similarity).

I tried a bunch of different words, I'm pretty sure it's just the doinky Samsung phone browser...

If you can, switch to Firefox. Not for Semantle. But because with Firefox, you can install an adblocker like ublock origin, making everything much more pleasant.

I gave up but I like the game, I'll try again tomorrow...

It doesn't know the word aeroplane

Now that it's the next day, I can say that, for instance, the day the answer was CANDLE, I got my highest hits from (iirc, I've actually forgotten already) ALTAR and, like, words relating to textiles you might find in a religious setting. I think one of them was WIMPLE, in fact. Those prompts made me fixate on textiles and wardrobe items, when I apparently should have been expanding to other things found in a ritual setting? Did anyone else get CANDLE and if so, what was your approach?

the 'algorithm' here is a neural net, they are famously not amenable to reverse engineering and as david said each word is represented as 300 numbers.

Then I can't help but think that the dynamic between the player and the game will be similar to other situations where the reward is entirely erratic and unpredictable. If the reward is powerful enough and is dispensed often enough to reinforce, the player will continue to seek it and may even become convinced they discovered the secret method for winning. If the reward is too weak or too infrequent to be reinforcing the player abandons the game.

i bailed before i got there as it was driving me nuts but i got highest with FIREPLACE, PYRE, FLAME. Couldn't find my way to where I was supposed to go. i have a funnier story about todays, but I will wait.

it should be noted that, having won once, i am particularly interested in seeing if i have figured out the internal logic somehow or if that was a fluke.
given that there are people (on this thread even) who are maintaining some consistency, I don't think it's an erratic or unpredictable puzzle so much as it requires access to a specific perspective, one that i'm having issues attaining.

My close guesses were LAMP and LANTERN, which you would *think* would give me a good basis for getting candle. I instead spiralled into guessing SEARCHLIGHT, SPOTLIGHT, LAMPSTAND, HEADLAMP and other stupid things, before taking a break and finally having CANDLE come into my mind.

I did also get a glitch where one of my closer guesses didn't come up with the green indicator bar, but instead said "cold" - I can't remember what word it was so this isn't a good bug report, I guess, but I figured I'd mention it seeing as we have the creator here.

gonna start c+ping my answers to post here post game

Can we just put the answer into one of those other sites and see the map of the close words or do we just keep guessing at the semantle site?

if you're seeing a way to do that, i'd be curious!

I don't think it's an erratic or unpredictable puzzle so much as it requires access to a specific perspective, one that i'm having issues attaining.

This is where I am, I'm by no means faulting the game. Actually some of the posts made since I mentioned my frustration earlier have helped me already shift my perspective for next time.

Yeah I think I'll give it a proper go tomorrow, as my initial confusion has been somewhat dispelled.

emil,y: Thanks for the report. The word you found wasn't a "normal" word (meaning, a word in the 370k-word all-lowercase wordlist that I am using to generate the top 1000). A friend of mine found this a few days ago with "winningest". It's weird to have two word lists, but I really didn't like some of the garbage in the full 3 million words of the word2vec data set I'm using (in one case, one of the top 200 words was some reporter's email address!). To reproduce: capitalize the first letter of today's 999/1000 word. I just changed the code to handle this case -- now it prints "????" with an explanation on hover.

It would be nice to display a map of the close words -- but "mapping" a 300-dimensional space is tricky. I guess I could run another round of principle component analysis on the top 1000, and take the top 2 components.... But I don't think that's going to happen this week.

er, today's 996/1000 word.

61 today, and I cheated by using a thesaurus at the very end.

I don't know why I have zero interest in posting my wordle results but I do feel like reporting on this one, I think because it's less predictable how hard it's going to be.

Wordle doesn’t really need a strategy beyond doing the obvious whereas this most definitely seems to require one.

It's a bit like Scattergories in the way it tends to lock down your brain at a certain point, so that you know there are other words out there but your brain fixates on the ones you've already come up with and can't get past them.

Yeah same

Yep

Cool. I just got it in 6, but that feels like beginner's luck!

― jaymc, Monday, February 7, 2022 4:09 PM (yesterday) bookmarkflaglink

Yesterday I had the fortune of randomly guessing "lantern" for the second word (after "frog"), which led me to "candle" soon after.

Today I got to the closest match in about 100, but then got stuck. Finally decided to cheat by using a thesaurus, which got me to the finish line.

63 guesses for me today - guesses included the 994, 997, 998 and 999th closest words - took a while to land on 1000, even when I could clearly see the pattern

when does the new word drop? Midnight? LIke the...like that other puzzle?

It would be really interesting to know the optimal, say, six words to start with to narrow down the possibilities.

107 words ... AND I used a thesaurus. The thing that broke it open for me was actually just free-associating. Not sure how I got from WHOOPSIE to ERROR but the answer came relatively quickly after that.

*************SPOILERS IN POST ABOVE SORRY*****************

will try to get a mod to fix

I got in 33 this morning, first one I've cracked in ages because I keep giving up after 40. I opened with circular which was low but a decent enough start.

I rarely get any green before 50 or even 100 these days and yep I usually give up before then. got today's in 65 though - first green at 60.

I solved Semantle #155 in 16 guesses. My first guess had a similarity of 14.54. My first word in the top 1000 was at guess #2. My penultimate guess had a similarity of 18.19.

that was the easiest one I've ever done, it helps getting 952/1000 on your 2nd guess though.

I solved Semantle #155 in 47 guesses. My first guess had a similarity of 6.79. My first word in the top 1000 was at guess #5. My penultimate guess had a similarity of 75.83 (999/1000).

I still play daily but I do feel like I've been on a streak of playing poorly, not getting stuff until the multiple hundreds. Today was 81, kind of a middling show. I figured out the type of word it wanted quite quickly, just took me ages to hit on the right one.

I solved Semantle #157 in 23 guesses. My first guess had a similarity of 3.74. My first word in the top 1000 was at guess #22. My penultimate guess had a similarity of 60.26 (991/1000).

I solved Semantle #161 in 39 guesses. My first guess had a similarity of 10.01. My first word in the top 1000 was at guess #12. My penultimate guess had a similarity of 32.10 (886/1000).

I solved Semantle #164 in 135 guesses. My first guess had a similarity of 9.17. My first word in the top 1000 was at guess #54. My penultimate guess had a similarity of 48.54 (955/1000).

Got on the wrong track early on so it took me a while, but it helps if you were paying attention to the news today

I solved Semantle #164 in 88 guesses. My first guess had a similarity of 13.52. My first word in the top 1000 was at guess #15. My penultimate guess had a similarity of 44.69 (912/1000).

That's a bit spoiler-y Roz. I was going to say I wonder if the choice of word was influenced by the news.

What the fuck?
I saw this thread on Site New Answers. Thought 'hey, I haven't done semantle in a while, let's give it a go.

I GOT IT IN ONE

I mean, kind of obvious why I guessed what I did, but....!

haha I'm kicking myself for not guessing it sooner, but I got sidetracked by chasing down a rabbithole of sailing related terms

Got the Wordle in one a few days ago too, when it was FIELD
<pulls tshirt over head, runs laps of the living room>

I solved Semantle #165 in 33 guesses. My first guess had a similarity of 20.41 (74/1000). My penultimate guess had a similarity of 65.49 (999/1000).

I got it in 1 tho! forks, you can tell Dave to send my prize now!

The site appears to be down.
Or has it moved?

yeah, semantle.com

Cheers, ledge!
Got it eventually...

I solved Semantle #165 in 212 guesses. My first guess had a similarity of 10.43. My first word in the top 1000 was at guess #9. My penultimate guess had a similarity of 2.40.

Recently I have got a lot better at cryptic crosswords, yet, seemingly, I am no better at this :-/

I solved Semantle #166 in 118 guesses. My first guess had a similarity of 7.21. My first word in the top 1000 was at guess #32. My penultimate guess had a similarity of -0.65.

I solved Semantle #166 in 60 guesses. My first guess had a similarity of 5.28. My first word in the top 1000 was at guess #7. My penultimate guess had a similarity of 18.26.

I solved Semantle #169 in 134 guesses. My first guess had a similarity of 30.58. My first word in the top 1000 was at guess #6. My penultimate guess had a similarity of 11.40.

today's was food-related, with all the frustration that entails. It wasn't actually *a* food though, which I thought it was for ages.

I solved Semantle #169 in 54 guesses. My first guess had a similarity of 18.82. My first word in the top 1000 was at guess #11. My penultimate guess had a similarity of 43.83.

I got 998 on guess 15, and then of course I guessed everything else BUT the target word :/

Best one in a *long* while. if it wasn't for the heat I wouldn't be awake now to trawl through it!

I solved Semantle #171 in 34 guesses. My first guess had a similarity of 20.90. My first word in the top 1000 was at guess #3. My penultimate guess had a similarity of 16.70.

Semantle fans, thanks for all of your support. Thanks to you, I was able to sell Semantle and try out a new career as an indie game developer. My next game is out now!

It's called Surfwords. Like Semantle, it's a very difficult word game. Unlike Semantle, it's played in realtime, and you'll be lucky to last thirty seconds.

I think you'll find it less frustrating (or at least differently frustrating) than Semantle: the word lists are carefully curated, and the game tries really hard to give you words that you will know -- but is willing to accept more obscure words if you enter them.

It's available now on iOS, Android, Windows, Mac, and Linux: https://surfwords.com/.

Great to hear buddy

will try this out later

I've beta tested Surfwords a lot and David's not kidding about it being challenging.

I daresay anyone with solid scrabble dictionary tastes is going to find this to be their cuppa tea. Try it a bit slowed down to start and see if you can get up to speed. I never could!

A key thought going into playing Surfwords is that Dave was inspired by Super Hexagon and wanted to keep the same twitchy gameplay but have it activate your vocabulary centers.

Separate conversation that I'm curious about the board's opinion on: Dave is selling Surfwords for $5. Considering that money goes to the one-man programming team, does that seem like a reasonable price? Who here still buys games a la carte on the Apple/Android store?

Anyone still playing?

I solved Semantle #199 in 40 guesses. My first guess had a similarity of -3.91. My first word in the top 1000 was at guess #31. My penultimate guess had a similarity of 28.37 (714/1000).

Better than yesterday's, when I had political in my first ten guesses but didn't think to try party for another 200 guesses or so.

Tried Surfwords, didn't like it. Sorry.

Thanks for the reminder! Haven't played in ages, having a full-time job now is probably the reason,but seem to have hit the ground running....

I solved Semantle #199 in 15 guesses. My first guess had a similarity of -1.24. My first word in the top 1000 was at guess #9. My penultimate guess had a similarity of 49.99 (999/1000).

The $5 price tag, peanuts as it is, and totally needed to support the developer, does put me off trying Surfwords, esp as it appears not to be ti many people's taste so far.

Bit more of a struggle today. One of those where I had early greens, then drifted away from it:
I solved Semantle #200 in 117 guesses. My first guess had a similarity of 22.41. My first word in the top 1000 was at guess #8. My penultimate guess had a similarity of 21.00.

chicken not being anywhere in the top 1000 for today’s feels like a mistake lol

I got it in 20 today which is good for me. Lately I've been resorting to hints once I pass 200 guesses which happens frustratingly often. Thought yesterday's was really hard.

I solved Semantle #255 in 171 guesses. My first guess had a similarity of 3.88. My first word in the top 1000 was at guess #13. My penultimate guess had a similarity of 55.68 (941/1000).

really weird one today

I solved Semantle #259 in 26 guesses. My first guess had a similarity of 19.38. My first word in the top 1000 was at guess #26. My penultimate guess had a similarity of 17.27. first one in a while. don’t think I ever guessed the word without having a green first. went from “flower” to “fresh” and “clean” (both with 23. similarity) to bingo

better than me, took me 116. rarely takes me less than 200 guesses these days. yesterday I had "convince" early on but didn't try "convinced" for another 200 guesses.

I solved Semantle #260 in 8 guesses. My first guess had a similarity of 14.06. My first word in the top 1000 was at guess #6. My penultimate guess had a similarity of 51.46 (984/1000).

I also played yesterday but didn't get a word higher than 40.00 in maybe 70 words. Didn't realize it would change before you finish

After Semantle, I quit my job to make games full-ish time. My latest is called Deco Deck, and it's a logic puzzle like Two Not Touch. If you like that sort of thing, you'll probably like this one. Give it a try -- it's available for iOS, Android, Windows, Linux, and Mac.

I will give that a try, my kakuro addiction is boring me to death and I need something new!

I tried it, pretty neat - looks good and it's certainly challenging, the mini was easy but I needed hints for the large after playing for nearly an hour!

one thing I'm curious about for games like this with a single unique solution - how do you generate the games? I presume there's an algorithm but I wouldn't know how to begin to design one.