The best Othello app I've found.
March 8, 2024 7:19 PM Subscribe
There are thirty Reversi (also known by the trademark name Othello) board game apps in the Google Play store. This is the best one and you can play it here in your browser. No ads, no trackers, no in-app purchases. It's a web app (the mobile app just opens the website.) A simple user interface with a minimalist approach to configurations. Completely free and open source. Written in Rust and TypeScript by Nate Stringham. Othello is a simple game but widely explored in computer science.
He was also on the team that produced a 4+ player social deduction game, Cat Me If You Can (which is also open source.)
There are other delightful free web app games linked on his GitHub page:
Hextris, a hexagon-themed falling block game by a team of developers (also free and open source.)
PWA (Progressive Web App) Memory Game by Narendra N Shetty (also free and open source).
Snake PWA by keeshii (also free and open source.)
Tic tac toe game by Thalles Maia. (Free and open source.)
These delightful, beautiful games remind me of a simpler time, in particular the old Orisinal Flash games which you can play without Flash.
He was also on the team that produced a 4+ player social deduction game, Cat Me If You Can (which is also open source.)
There are other delightful free web app games linked on his GitHub page:
Hextris, a hexagon-themed falling block game by a team of developers (also free and open source.)
PWA (Progressive Web App) Memory Game by Narendra N Shetty (also free and open source).
Snake PWA by keeshii (also free and open source.)
Tic tac toe game by Thalles Maia. (Free and open source.)
These delightful, beautiful games remind me of a simpler time, in particular the old Orisinal Flash games which you can play without Flash.
I was today years old when I learned that Othello and Go were not the same thing.
posted by The otter lady at 7:44 PM on March 8
posted by The otter lady at 7:44 PM on March 8
FYI, the default difficulty is rather easy to beat. The computer opponent doesn't seem to know even basic strategy, e.g. avoiding the 2x2 quadrants in the corners. I'll try it again with it set to hard (settings / gear icon in upper right corner).
posted by intermod at 8:29 PM on March 8
posted by intermod at 8:29 PM on March 8
When I was a kid I had an Othello set but not the rules, so I wrote my own. I wish I still had them.
posted by doctornemo at 9:29 PM on March 8 [2 favorites]
posted by doctornemo at 9:29 PM on March 8 [2 favorites]
Fantastic. I got a lot better at Othello playing that.
posted by constraint at 10:59 PM on March 8
posted by constraint at 10:59 PM on March 8
with it set to Very Hard, it crushed me
Just managed a 40:24 win against Very Hard after many, many crushings. Whew. I can stop now.
posted by flabdablet at 12:35 AM on March 9
Just managed a 40:24 win against Very Hard after many, many crushings. Whew. I can stop now.
posted by flabdablet at 12:35 AM on March 9
It has become clear that I am very bad at Othello.
posted by Going To Maine at 12:57 AM on March 9 [4 favorites]
posted by Going To Maine at 12:57 AM on March 9 [4 favorites]
I played a lot of Othello while writing my undergraduate thesis. I was on a Linux laptop that my significant other could not get to recognize its wifi card (it worked prefect for her in windows and the ethernet connection worked fine with both OSes). So to distract myself I played the few games the GUI came with. Which meant a lot of Othello. I'm just good enough to have looked into playing at a club level and gone "hell no. "
This system is good enough that when I make a mistake I recognize, I'm going to lose, but not so hard that I'm not winning the majority of games. Although it is embarrassing that I've had two losses where I was wiped out.
All in all a good game with a decent ai behind it.
posted by Hactar at 2:54 AM on March 9
This system is good enough that when I make a mistake I recognize, I'm going to lose, but not so hard that I'm not winning the majority of games. Although it is embarrassing that I've had two losses where I was wiped out.
All in all a good game with a decent ai behind it.
posted by Hactar at 2:54 AM on March 9
Training a GPT on Othello lead to the discovery that LLMs have an internal "world model" and aren't just generating plausible text based on superficial statistics.
posted by L.P. Hatecraft at 3:54 AM on March 9 [2 favorites]
posted by L.P. Hatecraft at 3:54 AM on March 9 [2 favorites]
The second oldest person ever, Kane Tanaka, played Othello to keep her mind agile.
posted by interglossa at 7:33 AM on March 9
posted by interglossa at 7:33 AM on March 9
Training a GPT on Othello lead to the discovery that LLMs have an internal "world model" and aren't just generating plausible text based on superficial statistics.
That's not really what I got from the actual explanation, to the extent I can untangle the bafflegab. It seems more like an LLM trained on playing Othello converges on the same move predictions as human programmed game engine: "two unrelated processes—(1) a human-understandable World Model and (2) a black-box neural network—reach highly consistent next-move predictions."
The claim is then that the LLM has generated an "internal" version of that "world model" (the game engine). Yet no one claims a chess computer is general AI.
We devised the intervention technique to change the internal activation so that it can represent a different world representation given by us. And we found this works concordantly with the higher layers of the language model—these layers can make next-move predictions solely based on the intervened internal activations without unwanted influence from the original input sequence. In this sense, we established a bidirectional mapping and opened the possibility of many applications, like the latent saliency map.
I wonder if they tried reversing the polarity of the retroencabulation matrix. Anyway, they conclude from that "two systems—a powerful yet black-box neural network and a human-understandable world model—not only predict consistently, but also share a unified mid-stage representation."
And note that "the world representation (board state) serves as a “sufficient statistic” of the input sequence for next-move prediction. Whereas for real LLMs, we are at our best only know a small fraction of the world model behind. How to control LLMs in a minimally invasive (maintaining other world representations) yet effective way remains an important question for future research."
So, yeah, this doesn't sound like it does anything to show LLMs are more than their statistics.
posted by snuffleupagus at 8:04 AM on March 9 [3 favorites]
That's not really what I got from the actual explanation, to the extent I can untangle the bafflegab. It seems more like an LLM trained on playing Othello converges on the same move predictions as human programmed game engine: "two unrelated processes—(1) a human-understandable World Model and (2) a black-box neural network—reach highly consistent next-move predictions."
The claim is then that the LLM has generated an "internal" version of that "world model" (the game engine). Yet no one claims a chess computer is general AI.
We devised the intervention technique to change the internal activation so that it can represent a different world representation given by us. And we found this works concordantly with the higher layers of the language model—these layers can make next-move predictions solely based on the intervened internal activations without unwanted influence from the original input sequence. In this sense, we established a bidirectional mapping and opened the possibility of many applications, like the latent saliency map.
I wonder if they tried reversing the polarity of the retroencabulation matrix. Anyway, they conclude from that "two systems—a powerful yet black-box neural network and a human-understandable world model—not only predict consistently, but also share a unified mid-stage representation."
And note that "the world representation (board state) serves as a “sufficient statistic” of the input sequence for next-move prediction. Whereas for real LLMs, we are at our best only know a small fraction of the world model behind. How to control LLMs in a minimally invasive (maintaining other world representations) yet effective way remains an important question for future research."
So, yeah, this doesn't sound like it does anything to show LLMs are more than their statistics.
posted by snuffleupagus at 8:04 AM on March 9 [3 favorites]
I still remember carrying the Windows 3.0 Reversi from upgrade to upgrade because it was the only one I understood
posted by scruss at 8:18 AM on March 9
posted by scruss at 8:18 AM on March 9
It does indicate that LLMs can learn a complex hidden structure (board state, relationship between spaces) when trained only on move sequences. This is analogous to Feynman's description of humans discovering physics: trying to figure out the rules for a game we only get to observe at a surface level.
The paragraph that you mocked for jargon says that they can change the LLMs internal board state (eg, adding or removing or changing a piece) and then watch it make appropriate moves based on the altered state. (The machinery for doing this is also quite simple.) This indicates that the internal representation of board state is important to the model, not some incidental statistical property in a more complex computation.
For example, suppose there were three variables x, y, z which you use for making a decision, and, incidentally, you can compute x+y+z, but you don't really use the sum to make your decision. Now suppose someone altered the variables so that they add up to 10 instead of 11, by adding 1/3 to each variable. This might not affect your decision process, despite the sum being different. (perhaps you only care which variable is largest, and adding a constant to each doesn't change that.)
What they're guarding against is the idea that they've found a way to observe the board state from the LLM internals, but not in a way which matters for actual move prediction. They show pretty conclusively that this isn't happening.
So, the Othello experiment demonstrates that the model can learn some complex combinatorial structure and use it for prediction, despite seeing only 'traces' of that structure in these move lists. If that's still 'just statistical manipulation' I would love to understand what /isn't/ statistical manipulation...
posted by kaibutsu at 8:37 AM on March 9 [1 favorite]
The paragraph that you mocked for jargon says that they can change the LLMs internal board state (eg, adding or removing or changing a piece) and then watch it make appropriate moves based on the altered state. (The machinery for doing this is also quite simple.) This indicates that the internal representation of board state is important to the model, not some incidental statistical property in a more complex computation.
For example, suppose there were three variables x, y, z which you use for making a decision, and, incidentally, you can compute x+y+z, but you don't really use the sum to make your decision. Now suppose someone altered the variables so that they add up to 10 instead of 11, by adding 1/3 to each variable. This might not affect your decision process, despite the sum being different. (perhaps you only care which variable is largest, and adding a constant to each doesn't change that.)
What they're guarding against is the idea that they've found a way to observe the board state from the LLM internals, but not in a way which matters for actual move prediction. They show pretty conclusively that this isn't happening.
So, the Othello experiment demonstrates that the model can learn some complex combinatorial structure and use it for prediction, despite seeing only 'traces' of that structure in these move lists. If that's still 'just statistical manipulation' I would love to understand what /isn't/ statistical manipulation...
posted by kaibutsu at 8:37 AM on March 9 [1 favorite]
Ideation. Non-deterministic forms of cognition. What we 'generally' call "thinking."
But maybe that's already admitted when discussing 'models,' given the comparison to a human-programmed game engine (which is a statistical machine itself); and so the claim isn't as extravagant as it sounds.
posted by snuffleupagus at 8:48 AM on March 9 [1 favorite]
But maybe that's already admitted when discussing 'models,' given the comparison to a human-programmed game engine (which is a statistical machine itself); and so the claim isn't as extravagant as it sounds.
posted by snuffleupagus at 8:48 AM on March 9 [1 favorite]
The fact that a system designed to efficiently return valid Othello moves winds up mimicking some of the underlying rules used to generate the inputs it was trained on is pretty cool. It says something interesting about what information is carried by the inputs and can be extracted from them. It would be interesting to see how long it would take a system to parse a more complex rules set, like chess.
The AI hype about "world models" and misrepresenting this as some form of cognition probably gets more clicks, but it obscures a pretty cool real result.
posted by The Manwich Horror at 9:14 AM on March 9 [2 favorites]
The AI hype about "world models" and misrepresenting this as some form of cognition probably gets more clicks, but it obscures a pretty cool real result.
posted by The Manwich Horror at 9:14 AM on March 9 [2 favorites]
Reversi is a wonderful game. I fondly remember Othello from my childhood, especially the materials. The green felt board, and the heavy plastic disks, and the sections that opened and closed. The pieces made excellent clicking sounds against each other, and a beautiful "snap" when flipping them over.
Apps are much better. They flip all the disks that need to be flipped, while the physical game, that's up to you. In the app, all you have to do is strategize. On the board, you have to stop strategizing and check and think and argue about what was flipped already and what was missed.
I bet that felt was because someone realized the sound of flipping on an unprotected board would drive non-players nuts. In Asia, they designed fancy goban to resonate with the sound of placing pieces. I wonder what other games have a high level of placement sounds, and how they handle it?
I love "simple" positional strategy games
posted by rebent at 9:45 AM on March 9 [2 favorites]
Apps are much better. They flip all the disks that need to be flipped, while the physical game, that's up to you. In the app, all you have to do is strategize. On the board, you have to stop strategizing and check and think and argue about what was flipped already and what was missed.
I bet that felt was because someone realized the sound of flipping on an unprotected board would drive non-players nuts. In Asia, they designed fancy goban to resonate with the sound of placing pieces. I wonder what other games have a high level of placement sounds, and how they handle it?
I love "simple" positional strategy games
posted by rebent at 9:45 AM on March 9 [2 favorites]
I am good at Othello. Very good. It's weird, because I'm bad at checkers, I hate playing chess, I don't really like one-on-one competitive games at all, but I love Othello, and I honestly can't remember the last time I lost a game -- until this morning, when the "very hard" setting beat me by 10 points. This is great! Thank you! None of my family members will play with me, so now I have a place to vent my Othello urges.
I disagree with rebent about apps being better, though. I still have that set with the covers that flip over, and I love the feel of the tiles.
posted by The corpse in the library at 11:16 AM on March 9 [1 favorite]
I disagree with rebent about apps being better, though. I still have that set with the covers that flip over, and I love the feel of the tiles.
posted by The corpse in the library at 11:16 AM on March 9 [1 favorite]
I wonder what other games have a high level of placement sounds, and how they handle it?
Mahjong, which we're learning. The heft and the clacking of tiles is very satisfying.
The default hardness of this Othello version is kicking my ass. Grrrr.
Viewing in a browser on my Android tablet, the board gets partially obscured with the "Install App" dialog, with every new game. Opening and closing the Settings thing clears it, but still, it's a pain.
posted by Artful Codger at 12:51 PM on March 9 [1 favorite]
Mahjong, which we're learning. The heft and the clacking of tiles is very satisfying.
The default hardness of this Othello version is kicking my ass. Grrrr.
Viewing in a browser on my Android tablet, the board gets partially obscured with the "Install App" dialog, with every new game. Opening and closing the Settings thing clears it, but still, it's a pain.
posted by Artful Codger at 12:51 PM on March 9 [1 favorite]
I've always been terrible at Othello. Can't imagine how a human could beat a computer at it. It just seems so cut and dried.
I would much rather play GIPF, DVONN (my favorite) or YINSH, if I want an abstract with tactile bits, (or Hive or even Go...)
posted by Windopaene at 4:21 PM on March 9 [2 favorites]
I would much rather play GIPF, DVONN (my favorite) or YINSH, if I want an abstract with tactile bits, (or Hive or even Go...)
posted by Windopaene at 4:21 PM on March 9 [2 favorites]
Yes, I'm starting to feel the difference between a computer opponent and a human one. I think my superpower is actually my impulsivity -- I can just throw down pieces and go, and my human opponent will get confused by the speed and not think ahead. But the computer can calculate the best move and do it.
posted by The corpse in the library at 2:18 PM on March 10
posted by The corpse in the library at 2:18 PM on March 10
« Older A better day after | Researchers discover Earth's earliest fossilised... Newer »
This thread has been archived and is closed to new comments
posted by AlSweigart at 7:21 PM on March 8 [2 favorites]