SteveB

I got a 1.9B model to watch a restaurant camera and tell staff which tables need cleaning 🍽️ It runs on my laptop. No cloud, and it generates zero tokens. Built over a weekend with @liquidai d1-2B

Faker is 30 with six world championships, and his screen flashes like a strobe light because he won't leave his camera in one spot for more than a split second. Every flash is him jumping over to a teammate to see what's going on in their lane, then snapping back to his own fight. He does it over and over for the entire game. He's using five keys. F1 puts the camera on himself, F2 through F5 put it on each of his four teammates, and he taps Tab in between to pull up the scoreboard with everybody's items. In League your screen only shows a small circle of the map around your own champion, so this is how he sees the rest of it. The minimap in the corner shows little icons for where everyone is, which helps right up until you need to know how much health the enemy laner has left or how many minions are about to hit your tower. And he can't just check once. An enemy who showed up in top lane four seconds ago could be standing in the river by now, so whatever he saw is already going out of date by the time he looks back at his own lane. The only way to keep the whole map current in his head is to keep cycling, all game, while also playing mid against the best players in the world. Google's DeepMind found out how much this matters when they built an AI to play StarCraft, which has the same camera problem. Their first version could see the entire map at once and beat two pro players 10-0. When they rebuilt it to look through a camera like a human has to, it lost its very first live game to a pro it had just swept 5-0. To be fair, that camera version had only trained for a week. DeepMind kept at it, and for the version they published in Nature they made the camera mandatory, only let it act on what was on screen, and capped how fast it could click after checking with pro players. It still made Grandmaster, top 0.2% of the ladder. They basically ended up treating the camera as part of the game itself. OpenAI's Dota bots skipped the problem completely. They read the game straight from the API and never looked at a screen. Most AI agents that use a computer today are stuck with the keyhole too. They take a screenshot, figure out what to do, click, and take another screenshot. Faker runs his version of that loop faster than you can follow on video. He plays BLG on Sunday for a shot at his fifth straight Worlds final. Try watching his monitor for one teamfight and you'll get dizzy long before he does.

This skill is mind-blowing 🤯

Once you experienced 1000 tokens per second it is hard to go back to anything else