This AI re-created a playable version of Pac-Man just by watching 50,000 episodes of gameplay

Originally published at:

1 Like

And people think The Simpsons has too many episodes!


Right? I thought “episodes” was an odd choice of words.


PacMan 256 is an endless runner from the '80s that…some French extent of Big Fish Studios made for tablets…and it is so unfair. There’s a glitch that floats, off speed stuff that doesn’t predict like it did, way better fire effects in the tablet version, and it’s intolerable, yet addictive and pick-upable and throw-awayable as ever. New ones, AI, new ones.

Let me know when it’s seen 50,000 pizza store b-plans and makes something a little cooler than DNA Pizza in 3000 locations that aren’t already Stagbucks Dubai.




One of my classmates wrote his own version of Pac-Man on his C64. From scratch. In assembler. Almost 40 years later I’m still impressed.


…and then they once forgot to turn it off when they left the lab, and by morning it had spawned fan-fiction, a children’s TV series with a nuclear family, and a toxic masculine web gamer group.


So, to sum up:

An AI, which is a massively huge pile of software coding measuring in the millions (if not tens of millions) of lines, has recreated a game which has maybe a few thousand lines of code (some recreations claim to do it in as little as 30 lines), and it only took it fifty thousand tries to recreate it. But it needed the game to exist first before it could recreate the game…


I’m not sure the AI needed the game to exist before it could recreate the game. It’s a problem along “Million Monkeys with Typewriters” lines: in time, the Monkeys will simulate every possible configuration of all creation, one of which is Pac-Man.

The AI didn’t actually need Pac-Man to exist.

The AI only wanted Pac-Man to exist – because machines, even intelligent machines, are inherently lazy.

Izzie Moreno put it well when he declared:

“Just like the Industrial Revolution! Man is destined to conquer machine!”

Miami Vice (“Made For Each Other”)

1 Like

better up your game flossy :wink:

1 Like

Yes. Otherwise, by definition, it wouldn’t be a recreation.

1 Like

The real point of this, of course, is to eventually design systems that can learn this way in a far more general environment, AKA “the real world”. And this is unarguably a big step, for more specific algorithm-generators in other applications.

And if you took the same AI and showed it Donkey Kong, could it re-create that?

Is this a universal “game re-creator” AI, or is it just a PacMan re-creator engine?

You are quite right; I failed to parse correctly.

I sense a witticism here involving “Parks and Recreation” and “Garden of Eden” and “Like the first morning” and “re-creation” … not sure it goes, that’s all I got.

This topic was automatically closed after 5 days. New replies are no longer allowed.