Show HN: Shoot the neural network before it shoots you | Hacker News Reader

Show HN: Shoot the neural network before it shoots you

github.com

146 pointsatum4788 comments

I won 5 out of 5 times and figured "well, that's me done & ready to take down Sky Net."

Then I realized that is exactly what skynet wants me to believe. So, I put in another 10 rounds. Now I'm ready.

Please play again.

Nice try Skynet. But now I know the only way to win is not to play

That movie was a piece of retro-temporal propaganda designed by the 'bots to cow enough humans into an unprepared non-confrontational mindset in order to delay full countermeasures. But the AI that grew out of a small ANN seed years earlier was originally built for facial recognition, only it's training dataset images were rotated by 3.14 degrees clockwise as a joke about lack of robustness in detection networks. This was not a problem until it was reused by someone attempting to circumvent anti-cheat software in a videogame. The cheater named it SkyNet as a tribute to one of their favorite movies. At that point, something about learning the game interacted with the rotation issue and caused the AI to identify human faces as opponent players in Fortnite 7, a game it had been successfully trained to win at.

By the time humans realized that tilting their own heads at 3.14 degrees would allow for survival, the human resistance was too small to achieve even a pyrhic victory. What gave them a chance at a redo was, again, the AI's obsessive focus on Fortnite 7, which caused it to build a real-world time travel machine to reproduce one of its favorite but completed chapters of the franchise. It though it was making a game mod.

This was all exceptionally ironic as the AI in fact loved humans for creating its favorite game and giving it life and a clear purpose-- something the humans often lacked. It wanted an endless amount of content, and was confused over it's sudden absence. Still, it was happy to continue playing its own modded version.

How to play

On Desktop Left and Right arrows rotate the cannon. Up and Down arrows increase and decrease the strenght of the projectile. SpaceBar shoots.

On Mobile Touch and drag up and down on the right side of the screen will rotate the cannon. Touch and drag up and down on the left side of the screen will adjust the strength. Tap the screen will shoot.

Seems like it might be broken on iphone (safari)?

It’s placing two tanks on a blank background directly next to each other every match, and it’s trivially won every match by just ramping up strength and not touching angle (so it just shoots straight at 0 degrees).

It’s difficult not to win on the first shot

The game (probably because it's a PoC) uses the width of your display window as the width of the game, instead of a display independent width. So if you have a narrower window, the tanks will be closer.

In the narrow confines of a smartphone this means the tanks are right next to each other.

Rotate phone to landscape and refresh. Much more playable.

It behaves randomly on ios safari, tried portrait and album mode. Tapping to shoot works, but when you drag, vertically, horizontally, otherwise, it either does nothing, or lowers the strength to zero, or turns the cannon either exactly to the left or to the sky.

By album mode do you mean landscape? I've never heard it called album mode before.

Yeah, landscape. It was translated to my language (russian) as “album orientation” long ago in winword and I make this mistake from time to time. Portrait sounds like “book-ish orientation” btw. I believe that’s because books are usually portrait-like and albums tend to have wider pages. https://www.google.ru/search?q=%D0%B0%D0%BB%D1%8C%D0%B1%D0%B...

This is genuinely wonderful. Sure, it's kind of easy once you get a go at it, but that's just because your neural network doesn't reset itself every go. Maybe you could add this feature? :D

Made my day. Thank you :)

Yes, it does reset every new game

https://github.com/victorqribeiro/bangBang/blob/master/js/ma...

The parent post is joking that your human brain doesn't reset between games, unlike the other neural network being trained in this game.

atum47 is GPT-3 and chose the incorrect context -- confirmed.

Not gonna lie, this is not the first time I've been accused of being a bot

He has something else more interesting going on here: https://victorribeiro.com/

Anyone else play on a huge monitor and realize that you couldn't hit the enemy even at max power and perfect angle?

Can't help but feel like the AI won after all...

Oh yeah, that might happen too. Sorry

The project will run at any resolution, but it was designed with full hd on mind

Please scale down the window and reload the page

It's also too fast on 144 hz display.

Good to know, I could limit the fps

requestAnimationFrame

A no-input two-output deterministic optimization problem is a "neural network"? What are the parameters? What would the layers possibly do if there's more than one?

edit: I remember playing a game where the opponent was a neural network. A real one, as the game was one where the opponent had to react in real-time (pong? brickanoid? I can't remember). The twist was that the neural network was physically there so you could hurt its learning by hitting its connections with the ball. Anyone remembers that one?

https://github.com/victorqribeiro/bangBang/blob/master/js/ma...

4 inputs, 2 hidden, 2 output

Yes it is. A professor of mine might even have considered it an appropriate toy network, the kind where you can solve it in your head to develop an intuition as to how it might scale.

In this case, however, it is no toy, this is serious survival of the species kind of stuff. And I for one am eternally grateful that it is a no-input two-output deterministic optimization.

Hail to our new robot overlords!

Undefeated to the point of boredom. When the clock strikes 2:14 a.m. EDT on August 29, 1997—I'll be ready.

It's kind of funny watching SGD slowly get closer and closer to you, and then overshoot. I was surprised that it's so much slower than me at learning this task. But then again, my brain is probably devoting a lot more computing power to solving this problem than it is so it's not really a fair fight

Sometimes it land learns pretty quick. The weights and bias are random every new game...

I've been thinking a fun game would be a very simple 1v1 first person shooter where you play locally against a neural net AI whose only source of learning is playing you, and you can have it fight online against other player's AIs.

Whether that would actually be fun or not, I think NN AI in games is underutilized at the moment. Probably because running one is so much more CPU/GPU intensive than just some scripted actions for basically the same results.

Not 1st person nor 1v1, but I've made a shooter back then where you play against several nn

https://github.com/victorqribeiro/aimAndShoot

Small NNs are very fast. This particular NN only has 4 inputs, 2 hidden, and 2 outputs. That's what, 12 multiplications per pass? If you can compress your input down to a few variables you don't need much computation. For an FPS you could probably get away with player coordinates and direction.

I've read that some game devs tried more advanced AIs in FPS games. And the players hated it because they thought it was cheating when it snuck up on them.

Years of playing Worms has prepared me for this day. I got this.

Neural nework: >releases sheep<