Very cool! I am a good Tetris player (in the top 10% of players) and wanted to give brick yeeting against an LLM a spin.
Some feedback:
- Knowing the scoring system is helpful when going 1v1 high score
- Use a different randomization system, I kept getting starved for pieces like I. True random is fine, throwing a copy of every piece into a bag and then drawing them one by one is better (7 bag), nearly random with some lookbehind to prevent getting a string of ZSZS is solid, too (TGM randomizer)
I actually grew up playing the Spectrum HoloByte version of Tetris for PC, which only lets you rotate in one direction. As a result, I ended up playing NES Tetris for years as a kid before realizing it lets you rotate clockwise / counterclockwise!
cool to hear about more randomizers. I am not a great tetris player but I absolutely love the game. I put NES, 7-bag and my own take on randomizer (crap) in my online tetris-like experiment https://www.susmel.com/stacky/ (you can press c for more controls or h to see shortcuts)
One of my dream goals was to make a licensed low lag competitive game kind of like TGM, but I heard licensing is extremely cost-prohibitive so I kind of gave up on that goal. I remember I said to someone I was ready to pony up few tens of thousands for a license + cut, but reportedly it starts at an order of the magnitude higher.
The worst thing is that the delayed auto shift is slightly off and it messes my finesse. (I used to play competitive tetris as well, but between getting older -> worse reflexes and vision problems I can't really play anymore. Weirdly, finesse muscle memory is still working.)
I don't think the goal is to make a PvP simulator, it would be too easy to cheese or do weird strategies. It's mostly for LLMs to play.
On the topic of reflexes decaying (I'm getting there, in my late 30s): Have you played Stackflow? It's a number go up roguelite disguised as an arcade brick stacking game, but the gravity is low enough that it is effectively turn based. More about 'deck' building, less about chaining PCs and C-Spins.
Thanks for all the questions! More details on how this works:
- Each model starts with an initial optimization function for evaluating Tetris moves.
- As the game progresses, the model sees the current board state and updates its algorithm—adapting its strategy based on how the game is evolving.
- The model continuously refines its optimizer. It decides when it needs to re-evaluate and when it should implement the next optimization function
- The model generates updated code, executes it to score all placements, and picks the best move.
- The reason I reframed this problem to a coding problem is Tetris is an optimization game in nature. At first I did try asking LLMs where to place each piece at every turn but models are just terrible at visual reasoning. What LLMs great at though is coding.
Interesting but frustratingly vague on details. How exactly are the models playing? Is it using some kind of PGN equivalent in Tetris that represents a on-going game, passing an ASCII representation, encoding as a JSON structure, or just directly sending screenshots of the game to the various LLMs?
It has to be turn-based. Even with Flash's speed, the inference latency would kill you in a real-time loop. They're likely pausing the game state after every tick to wait for the API response before resuming.
answered this in a comment above! It's not turn or visual layout based since LLMs are not trained that way. The representation is a JSON structure, but LLMs plug in algorithms and keeps optimizing it as the game state evolves
Thanks for the clarification! Kind of reminds me of the Brian Moore's AI clocks which uses several LLMs to continuously generate HTML/CSS to create an analog clock for comparisons.
Gemini 3 Flash is at a very nice point along the price-performance curve. A good workhorse model, while supplementing it with Opus 4.5 / Gemini 3 Pro for more complex tasks.
LLMs playing Tetris feels like testing a calculator's ability to write poetry. Interesting as a curiosity, but the results don't transfer to the tasks where these models actually excel.
Curious what the latency looks like per move. That seems like the actual bottleneck here.
I mean, if you let the LLM build a testris bot, it would be 1000x better than what the LLMs are doing. So yes, it is fun to win against an AI, but to be fair against such processing power, you should not be able to win. It is only possible because LLMs are not built for such tasks.
my unvalidated theory is that this comes down to the coding model’s training objective: Tetris is fundamentally an optimization problem with delayed rewards. Some models seem to aggressively over-optimize toward near term wins (clearing lines quickly), which looks good early but leads to brittle states and catastrophic failures later. Others appear to learn more stable heuristics like board smoothness, height control, long-term survivability even if that sacrifices short-term score
That difference in objective bias shows up very clearly in Tetris, but is much harder to notice in typical coding benchmarks. Just a theory though based on reviewing game results and logs
It would be more interesting to make it build a chess engine and compare it against Stockfish. The chess engine should be a standalone no-dependencies C/C++ program that fits in NNN lines of code.
My back-of-the-envelope guess would be that 99% of LLMs given the task to build a chess engine would probably just end up implementing a flavor of negamax and calling it a day.
Comparing against stockfish isn't fair. That's comparing against enormous amounts of compute spent experimenting with strategies, training neutral nets, etc.
It will lose so badly there will be no point in the comparison.
Besides you could compare models (and harnesses) directly against eachother.
It's not. Maybe if you used old versions of stockfish that predate the neural net methods used by current versions, because otherwise you'd be comparing the hand-rolled (by an LLM) position evaluation functions against an NNUE and the results of that are a forgone conclusion; stockfish will stomp it every time.
Maybe that's the result you want for some sort of rhetorical reason, but it would nonetheless not be an informative test.
Some feedback: - Knowing the scoring system is helpful when going 1v1 high score
- Use a different randomization system, I kept getting starved for pieces like I. True random is fine, throwing a copy of every piece into a bag and then drawing them one by one is better (7 bag), nearly random with some lookbehind to prevent getting a string of ZSZS is solid, too (TGM randomizer)
- Piece rotation feels left-biased, and keeps making me mis-drop, like the T pieces shift to the left if you spin 4 times. Check out https://tetris.wiki/images/thumb/3/3d/SRS-pieces.png/300px-S... or https://tetris.wiki/images/b/b5/Tgm_basic_ars_description.pn... for examples of how other games are doing it.
- Clockwise and counter-clockwise rotation is important for human players, we can only hit so many keys per second
- re-mappable keys are also appreciated
Nice work, I'm going to keep watching.
https://en.wikipedia.org/wiki/Tetris_(Spectrum_HoloByte)
One of my dream goals was to make a licensed low lag competitive game kind of like TGM, but I heard licensing is extremely cost-prohibitive so I kind of gave up on that goal. I remember I said to someone I was ready to pony up few tens of thousands for a license + cut, but reportedly it starts at an order of the magnitude higher.
I don't think the goal is to make a PvP simulator, it would be too easy to cheese or do weird strategies. It's mostly for LLMs to play.
On the topic of reflexes decaying (I'm getting there, in my late 30s): Have you played Stackflow? It's a number go up roguelite disguised as an arcade brick stacking game, but the gravity is low enough that it is effectively turn based. More about 'deck' building, less about chaining PCs and C-Spins.
By the way, kudos on your feedback. If I was OP, I would've been honored to get that type of fine-tuning comments.
- Each model starts with an initial optimization function for evaluating Tetris moves.
- As the game progresses, the model sees the current board state and updates its algorithm—adapting its strategy based on how the game is evolving.
- The model continuously refines its optimizer. It decides when it needs to re-evaluate and when it should implement the next optimization function
- The model generates updated code, executes it to score all placements, and picks the best move.
- The reason I reframed this problem to a coding problem is Tetris is an optimization game in nature. At first I did try asking LLMs where to place each piece at every turn but models are just terrible at visual reasoning. What LLMs great at though is coding.
Also, if the creator is reading this, you should know that Tetris Holdings is extremely aggressive with their trademark enforcement.
https://clocks.brianmoore.com
Curious what the latency looks like per move. That seems like the actual bottleneck here.
.....
l....
l....
l.ttt
l..t.
I mean, if you let the LLM build a testris bot, it would be 1000x better than what the LLMs are doing. So yes, it is fun to win against an AI, but to be fair against such processing power, you should not be able to win. It is only possible because LLMs are not built for such tasks.
Task: write and optimize a tetris bot
Task: write and safely online optimize a tetris bot with consideration for cost to converge
openai/baselines (7 years ago) was leading on RL and then AlphaZero and Self-Attention Transformer networks.
LLMs are trained with RL, but aren't general purpose game theoretic RL agents?
That difference in objective bias shows up very clearly in Tetris, but is much harder to notice in typical coding benchmarks. Just a theory though based on reviewing game results and logs
https://en.wikipedia.org/wiki/Negamax
It will lose so badly there will be no point in the comparison.
Besides you could compare models (and harnesses) directly against eachother.
Maybe that's the result you want for some sort of rhetorical reason, but it would nonetheless not be an informative test.