Back to Stories

Comments (17)

Syzygies•22 days ago
The Kodak Research Labs (like Bell Labs) let their researchers play. In the 1960's my father (who later devised the Bayer filter for digital cameras) coded this algorithm for "Jotto" the 5 letter word version of Mastermind.

Computers were so slow that one couldn't consider every word in the dictionary as a potential guess. He decided empirically on a sample size that played well enough.

I became a mathematician. From this childhood exposure, entropy was the first mathematical "concept" beyond arithmetic that I understood.

dang•22 days ago
This comment goes in the highlights list: https://news.ycombinator.com/highlights!

(I mention this so people will know the list exists, and hopefully email us more nominations when they see an unusually great and interesting comment, like this one.)

spondyl•20 days ago
Any plans for the HN API to expose highlights? It's a neat high signal collection that I have some custom bit of code that translates it into an RSS feed but an official endpoint would be neat :)
dang•20 days ago
That's a great idea, but probably will have to wait for the next version of HN's API.
irchans•22 days ago
> I became a mathematician. From this childhood exposure, entropy was the first mathematical "concept" beyond arithmetic that I understood.

Very cool.

entropicdrifter•22 days ago
I too became a fan of this game from an early age, note the username
UltraSane•21 days ago
What is it like being cited by Satoshi Nakamoto?
mqtx•22 days ago
That reminds me, I wrote it for iPhone when it was released in 2007 - back then there was no App Store, so apps could only be written for browsers. I think I implemented it whenever I learned a new language. By the way, I noticed that Claude, ChatGPT, and DeepSeek - none of those LLMs can solve Mastermind. They get lost after a few iterations, no matter how good the prompt instructions are. Source: https://github.com/muquit/iphonemm - there is a link to play on that page.
fuzzfactor•21 days ago
I'm in an older generation, but like some other commenters started young with logic games as a preteen. I got into a school that was recognized for its academic games program but that was years before Mastermind came out.

By then I was in college with some of my high school buddies when one of them got the original board game. These were not actually that common even among those who liked thinking games since they were about $10, for $1 worth of plastic. And the box was way small compared to the full-size popular board games which were about $5.

It was quite interesting that there were diverse international people pictured on the cover and it didn't take long to realize this was a game where no players need to speak the same language at all.

By the time I got my Atari computer, this was the only game I ever coded since it's absolutely perfect to play having the computer in the role of "codemaker". I had never seen or heard of a computer mouse but I had my trackball for playing Centipedes and Missle Command so used it to point & click, then before I was done I was dragging & dropping the colors from their repository to the hot spots. I was simulating picking & placing like the board game.

Eventually I used two columns so two players could play against the Atari simultaneously, each with their own trackball or joystick. Since it was always basically a "one-player" game doing the solving, then two people could solve at once. I was going to suggest trying it that way next time but I don't know if it's as straightforward to use two mice and associate one with each player.

conception•21 days ago
You can ask them to write a script to solve it however and it is a trivial task for an LLM.
mqtx•21 days ago
I'm talking about interactive play with an LLM - for example, explaining the rules and scoring in detail first. Then I tell it to pick colors, and I provide the score. When the score indicate 2 colors matched, it's supposed to keep those 2 colors constant in the next move. But after I pointed out that it forgot again after the next move and continued making the same mistake, it was Claude 3.5. So at the end I said, 'It looks like you can never solve this puzzle. Now that you know all the rules, can you write the game in JavaScript?' Sure enough, it did, and it worked on the first try - and was much prettier than my version. Same for ChatGPT, DeepSeek.
juxhindb•21 days ago
One fun assignment at Oxford's Comp. Sci. department was solving Mastermind in Haskell using minmax. Knuth has a particularly fun paper on it in fact - https://www.cs.uni.edu/~wallingf/teaching/cs3530/resources/k...
mrbluecoat•22 days ago
I grew up with this game. It hurt my brain but in a good way. I think a lot of my problem solving and interest in coding stemmed from it.
cortesoft•21 days ago
Had the same experience. I loved playing with my mom, and I remember spending a lot of time thinking about how to optimize my guesses. It fit all my hobbies around logic.
bethekidyouwant•22 days ago
My grade school had a game club at lunch hour. Mr. Newton took time out of his day, every day for us. Still think about him and the many games he brought from his had personal collection. Not sure what the point of this message is other than that he will be remembered until the last time I think about mastermind, Pacific typhoon, star fleet battles, axis and allies and many others..
SchwKatze•21 days ago
I had a friend in my high school that was just like that, any day with not class was a excuse to bring some boardgame and spend hours playing in the empty classroom. Xcom, catan, zombicide, ticket ride, and so on
Saigonautica•21 days ago
Oh, I loved working on this problem years back.

My strategy was simulate my possible next guesses against all possible codes, then pick the option that had the highest number of possible outcomes (sometimes this strategy is called MaxParts). It looks like the author's approach works for similar underlying reasons.

Besides this, I applied some optimizations for the starting move, and some further optimizations on considering 'irrational' guesses -- e.g. choosing a code that had already been eliminated as a possibility, because it returned more information (this was rare, but possible).

I ran my code against all possible games of 4,6 mastermind (I win in an average of 4.2778 guesses), and found that some starting guesses were more optimal than others! The pattern "AABC" (e.g. red-red-yellow-green) was the best performer. Perhaps this is a way that the author can improve their algorithm just a tiny bit.

rokkamokka•22 days ago
I love stuff like this. It always tickles my brain to try and find the optimal way (or, as optimal a way as I can) of solving puzzles. Sometimes it's easy, sometimes it's really hard. Oftentimes you can get something decent with not too much effort, and the dopamine hit is great when you see it working
psychoslave•21 days ago
I didn't back it with any formal reflection, but I played the game a bit with my daughter last year.

We had some fun and nice exchange around it, but closing all parties in just 4 rounds made it quickly boring.

helph67•19 days ago
Back in 2013 a friend created this Dos batch form using Ansi graphics https://www.dropbox.com/scl/fi/bxymba1fooefhw1kevihb/BATGAME...
kshahkshah•22 days ago
I play Mastermind with my kids. It hasn’t clicked quite yet with them but I’ve shown them my strategy of eliminating colors by making an entire row a single color successively. Either it’s not present or now you know how many of a particular color. Then you just need to figure out ordering. Again you can use a known “bad” color to avoid ambiguity of multiple white pegs
apaprocki•21 days ago
Similarly, I explained to someone how to systematically solve it like this, but then when I suggested they now apply it in a new game, I chose all the same color for the code and watched them second guess themselves until all guesses were exhausted. They just couldn’t accept the possibility that the code could simply be all the same color. Was a good opportunity to quote Sherlock Holmes.
jmount•22 days ago
Really great article. Internalizing that a good realizable Mastermind strategy is to always eliminate the same number of possible solutions is a great way to internalize the value of information theory thinking. Getting hung up if it is exactly "plan k steps forward optimal" (i.e. fine details of the remaining possible cases) can be counter-productive.
nairoz•22 days ago
Really liked it. I'm curious about games for which the guesses are not always a valid solution but can contain special operators. For instance, add a state which is true for 2 different values in Secret Code : 1or2. The code won't contain it but it can be useful for getting more information at each step.
NooneAtAll3•21 days ago
fascinating website bug: pop-up to select color in the last couple guesses is displayed below "6^4 = 1296" formula in the text below - so the pop-up closes when I hover over the numbers
m-hodges•22 days ago
This was my favorite game to play with my parents when I was growing up.
cwmoore•22 days ago
Cool. Tiny tip, “Worlde” is obviously a typo for the popular puzzle.
ghaff•22 days ago
I've never been much for word games but have been a Wordle regular. But a friend of mine made the very astute observation that Wordle is more like Mastermind (which I played and even wrote a Windows version of as an exercise once upon a time) than any traditional word game. OK. You need a solid English 5 letter vocabulary. But it's really pretty different from crosswords and a lot of other word games including those on the NYT site.
madcaptenor•22 days ago
I think you can distinguish between word games where it helps to know what the words mean (like crosswords) and word games where the meaning of the words is irrelevant, only that they are or are not words (like Wordle).

Weirdly, Scrabble resembles crosswords graphically but fits in the second group.

yen223•22 days ago
Sudoku is kind of the same, in that the numerical values of the numbers doesn't come into play at all. You can replace the digits with fruits and the game plays the same.
cwmoore•21 days ago
I enjoyed this comment. Fruit emojis don’t sum well, so by your classification separate out the similar Kakuro or KUBOK16 from Sodoku, and with that, I have to drop a plug for my https://www.kakurokokoro.com/ in which “clues” are the column and row sums.
madcaptenor•21 days ago
Except it’s harder to iterate over the numbers when they’re fruits. (Unless you have a list of nine fruits in your head you can rattle off.)
ghaff•22 days ago
Although I must say I do like Connections where even subtle meaning very much factors in.
NooneAtAll3•21 days ago
tho typo so obvious that it became a thing https://worldle.teuteuf.fr/
CyberDildonics•21 days ago
I don't know if you need "information theory" to do something that can be solved with grep.
SchwKatze•21 days ago
Sorry, could you elaborate more about using grep to solve this? I can't imagine how
internet_points•21 days ago

    $ printf '%s\n' {a..f}{a..f}{a..f}{a..f}|head 
    aaaa
    aaab
    aaac
    aaad
    aaae
    aaaf
    aaba
    aabb
    aabc
    aabd

    $ printf '%s\n' {a..f}{a..f}{a..f}{a..f}|wc -l 
    1296
So say you've guessed acab and gotten two black pins, then you know nothing is in the right position, so you can `grep -v acab` but also that two of your pins should be used somewhere `grep -E 'a.*[cab]|c.*[ab]|b.*[ca]'` (hope I got that right?)

    $ printf '%s\n' {a..f}{a..f}{a..f}{a..f}|grep -v acab | grep -E 'a.*[cab]|c.*[ab]|b.*[ca]'|wc -l
    756
That filters your options down.

However, I don't understand how/whether grepping can give you the optimal next guesses – some of these will give you less information than others. (Of course, you could run some simulations with grep, argmaxing over all the possible responses like the article says, but then you're not just doing simple grepping.)

CyberDildonics•21 days ago
Each constraint given to you by the game can be done in grep and used to filter a master list of all the answers by piping the filtered list through each constraint.

https://github.com/codebox/wordle.sh

CaptainNegative•21 days ago
I don't understand the point of this article, as it doesn't define an objective function. It just states a strategy that is only practically implementable for small board sizes (given the cited NP-completeness result) and then calls it good sans theorem or even conjecture.

I believe it is provably not the optimal algorithm for solving the problem under the minimax objective, and I have a hunch that (due to rounding issues) it is also not optimal for minimizing the expected number of guesses under even a uniform prior. So what does this actually accomplish?

nsomani•21 days ago
I agree with you. I agree with OP in the following sentences:

>We have now landed on our final strategy: start by figuring out the number of possible secret codes n. For each guess, calculate the number n_i' of codes that will still be viable if the Code Master gives response i in return. Do this for all possible responses.

But then I don't agree with:

>Finally, calculate the entropy of each guess; pick the one with the highest.

Why wouldn't we just pick argmin_{guess} sum{i in possible responses}{Pr[i] * n'_i} = sum{i in possible responses}{n'_i/n * n'_i} = sum{i in possible responses}{n'_i^2}? This is the guess that minimizes the expected size of the resulting solution space.