9 HN users recommended this demo10 months2015–2025score 11
Years with an HN recommendation, 2007–2026
Loved on HN as a hilarious, brilliant example of hacker-academic work and cited whenever game-playing AI or fuzzing comes up.
Tom7 explains and demos software that learns to play NES games by finding score patterns in memory and then searching for inputs that raise it. It is the video companion to his SIGBOVIK paper.
“Watch the whole thing, it's delightful.” (HN, 2021)
Thanks for pointing that out. I wouldn't have noticed who that was otherwise. His video on learnfun/playfun is both hilarious and amazing. [this video]
There was an AI created the absolutely brilliant Tom7 on on his YouTube channel 'SuckerPinch' on April 1, 2013 that paused a game of Tetris instead of losing. In my opinion, Tom doesn't get enough credit for his work. He released his AI a full 8+ months before DeepMind released their Atari playing AI. Watch the whole thing, it's delightful. The AI pauses Tetris right before it loses at around 16mins into the video. [this video] I also recommend his other videos. Brilliant guy. Edit: Here's the paper [link]
AI that paused Tetris before losing, predating DeepMind Atari.
Haha, I love it. There was a SIGBOVIK submission a couple years ago that looked at byte patterns within video game RAM (specifically NES) to automate the playing of. Pretty entertaining stuff. [this video]
Entertaining automated NES playing from RAM byte patterns.
Tom 7's work on machine learning and NES emulation is very impressive. This isn't the first time he's attacked that space; you can see previous experiences he had with an automatic game-playing AI he built called "LearnFun / PlayFun" on that YouTube channel also: [this video] [link] [link]
If you're interested in a "dumber" approach to machines "learning" to play classic video games (including Super Mario Bros), tom7's video [1] and paper [2] on Playfun are both fascinating. [1] [this video] [2] [link]
Dumber approach to machines learning to play games.
In the context of AI and gaming, I definitely recommend this series of three youtube videos: [this video] Some games are played better than human would.
You should look at some other things the author did. They fit in the same area of craziness, and are absolutely great. For example [link] , Portmantout: A portmanteau of every English word, or the first one, [this video] , Computer program that learns to play classic NES games. I also liked the humor in What, if anything, is epsilon?, [link] .
I'm not sure many people can pull this off. The author here was also great in his presentation about automated NES playing, [this video] . You don't see often academics & hacker work that is also combined with the right amount of humor. He even 3Dified the original Zelda!
Academic hacker presentation on automated NES playing with humor.
So as someone with only a lay understanding of what NP-Hard means - does anyone remember that neural network that someone trained to play the game[1]? So if you can train a neural network to solve an NP-Hard problem, does that mean by definition (since NP refers to a whole class of problems) that a neural network could eventually be trained to solve any given NPH problem given the time and compute resources to do so? [1]: [this video]
Why not use game emulators? With popular NES emulators you can advance the game frame by frame. You can read the raw memory addresses that correspond to the score. You can dump the memory at any time and reload the game to a specific game state. You can even manipulate the games in many fun ways by messing around with the game memory. Or give an AI algorithm access to memory addresses as additional information, instead of relying on pure machine vision, if you want to do that.. Here's an example of a guy who made a general game playing algorithm that brute forces it's way through any NES game: [this video] This isn't necessarily interesting from an AI perspective - the playing algorithm is just brute force. But it shows what can be done with the platform, easily reloading to previous states and exploring counterfactual futures (which is exactly the sort of thing RL algorithms do.) He also has a cool algorithm for finding the objective function of an arbitrary game, by watching a human play, and seeing what memory addresses increment. Which is a lot more easy to use than writing OCR code to read the score and game over states from the screen.
Brute-force general NES playing using memory access.
Here is an actual implementation of trying to implement a general AI to play all NES games. It's pretty interesting, but his approach does involve the algorithm having access to the raw NES memory map: [this video]
This reminds me of AI research using NES Games. The AI eventually became proficient at completing Mario levels, and along the way it discovered novel strategies for survival, obtaining points, and finishing levels. > Check out this timestamp to watch the machine "cheat": [this video] > Researcher's site about the project: [link] > The Paper: The First Level of Super Mario Bros. is Easy with Lexicographic Orderings and Time Travel...after that it gets a little tricky. : [link]
A lot of fuzzers use Mario or other simple games as an internal testcase. I'm aware of a hypervisor fuzzer from 2016 that did it, and I'm positive there are others (both before and since). Hell, tom7 has a fuzzer for exploring program states that uses Super Mario Bros as the example from 2013 ( [link] , plus a youtube video [this video] ), and he's definitely not the first either.
Early example of fuzzer exploring Super Mario states.
Tom7 implemented some of this with his Learnfun and Playfun ( [this video] ) - It looks for lexicographical changes to bytes in memory (i.e., one byte seems to almost always increase, and then when it suddenly decrements, a neighboring byte is incremented) as a way of calculating memory location for increasing values (score, x-position), and similarly detects bytes that generally seem to decrease (lives). It then tries to play the game, choosing inputs that increase the values that should be increasing (move right, gain lives), or tries not to decrease the values that are decreasing (don't die).
Explains how LearnFun finds score and lives bytes in memory.
A bit out of context, but it reminded me of this funny moment. The only winning move is not to play. [this video] (Background: Someone training an algorithm to win NES games based on memory state)
Video list and Hacker News comment labels from HackerTube (prompted by @karpathy), released under ODC-By 1.0, built on the open-index/hacker-news archive. Hacker News data is provided by Y Combinator; comments remain the work of their authors. Video titles, channel names and thumbnails belong to their creators and are shown via YouTube. HackerTube Watch is an unofficial edition. Data export of 2026-10-31.
19 Hacker News comments
From the HN archive; up to 20 kept per video, favoring different commenters. Recommendations first.
Citations, criticism and jokes
Read more