Відео

Awesome video ! I usually don’t comment but this time I find it necessary that you see appreciation for what you do

Love how the Nascar guy used a real life wallbang to good effect.

train it to play CSGO

This is not AI. You claim that you believe this is closer to RTA than it is TAS, and that is categorically wrong, on every level. This is brute force TAS clip building with rules. The only thing that sets this apart from a TAS, is you seem to eliminate things like track-length nose-slides and the like, which makes it appear closer to RTA, but is still tool assisted. This is not a 'learning' AI. There is no AI on the planet that 'learns'. All we have is iterative brute force computing bases that mimic language and input to create an artificial response within a set of restrictive rules and guidelines. If you believe your AI truly 'learns', then remove the neutral network rewards. That which learns is self-determinate enough to set it's own goals.

bro heard of Roko's basilisk and said 'hold my beer'

you should make a video of how are you doing this. Lot of things I do not get. Like when AI is controling the care after it learned - does he react live or he has key press combination already before the game? Or actually is it controlling with virtual wheel or keyboard? How do you even make it control the game? Game does not have API for computer program to control it or does it? And lot of those technical things are interesting. How the AI is trained?

but slacilng down resolution and colors - colors are clearly enough info for AI . Now resolution - does it matter? and I did not even think it sees it like human. I would imagine it see some coordinate points of the track and lines joining those coordinate points. Hmm, but actually if AI can beat humans visually then this could be used in war. Try making world of tanks AI player or call of duty :)

if it is hard to discover it for AI, why not help him ?

hitting the wall to keep speed and not breaking a car is like cheating. The game physics suck if that is how it works. Maybe AI did not assume he needs to win by 'cheating'

Please make your AI do the Deep Dip track.

This is very interesting. There should be an annual international TM AI competition where universities, private entities and whoever wants to compete can race each other with their AI for fastest lap like there is with Robocup football (soccer).

Not an AI, take your clickbait elsewhere. Too bad you can't use middle-school bully level tactics with the "pin of shame" anymore.

Interesting.. Great content. 👍🏻

No! You're kidding! An air got good at something it was trained on?!? If you guys spent 1% of your time trying to safeguard Ali's THAT would be an accomplishment, but you don't. Have fun playing with your ai.

Hey! Amazing video. I find it, at the same time, a very entertaining and interesting introduction to machine learning/AI.

Close to yrs 3

The fundamental issue is that the AI has no idea what a car is. What a road is. Or even what it is. It has no concept of the fact that it is being asked to play a digital game based on real work items and mechanics. Until we can start teaching AI what the non digital world actually is, it will never become general intelligence. So far the closest we have been able to come is to teach AI to make digital representations of the vast majority of the world and the things in it, but that still doesn't breach that gap of understanding. Its like trying to teach someone who has been blind from birth what a color is by converting colors from a camera into sound. Sure. They might start to comprehend on some level the variations, and combinations. But they will never really understand what color is, let alone sight.

This should be "humanity" (whatever you think that means) VS "humanity", because AI is a human tool, not an independent entity with self-consciousness. Period.

Anything that an AI does has no value for society if it's not for automating human tasks that nobody wants to do.

this is actually terrifying.

I’ve got an idea. don’t try and be as fast, be as OPTIMIZED and EFFICIENT as possible. Try to have behaviors like a TAS run. And maybe as fast as possible because it’s cool seeing the ai blitz maps

For an AI to find shortcuts, you have to add "Try something new" function. Something that will trigger the luck. ??

Im starting to get really curious.. what if the ai plays "TM 2020" im curious on how this advanced network would handle the boost tiles and hovering on water. Especially the random turbo pad and the new car modes with the desert snow and rally cars

Please keep up making epic videos about your ai, even if it's just demonstrations, not improvements!

would love to see this on a07

.

Love your content. Small complaint. The music is sooo loud and your voice is super quiet. Hear from you (hopefully more clearly) in the next one

Can you make an TAS creating AI?

THIS is what AI was MEANT for, NOT dumb low effort tiktok's

23:10 so you're saying a fruit fly could play track mania :)

while watching this I was thinking about new Nvidia tools they released that use simulations and then sends the results to a real world robot... I wonder if the racing A.I. could be used to steer a remote control car on a racetrack or to participate in an annual competition where mini-cars race inside a maze, they use vacuum suction to stay on the track on tight turns

Now all we need is an AI that can edit videos for you.

Solo le podemos ganar con shortcuts

Thanks for the vid! For future vids, the audio may be better balanced with louder voice / quieter music.

How did he simulate independent input for each window of the game. 5:15 each window opens and the machine plays the game while he is still able to interact with the rest of the computer.

A wall bang works but in a real track be literally suicide

I don't really know much about it, but you talked about the rewards when the Ai is close to the line. Why not just a reward when it goes further along the line, not just close to the line but only when it goes forward along the line

He just gave all track mania play a gut punch. This Ai brain no bigger than a fruit fly!

If you have access to all these human runs, can you not train the AI directly on their data first before letting it then go on to learning through playing? Give the optimisation algorithm a warm-start basically rather than starting learning from nothing where it has to try everything before finding strategies that descend the gradient curve

that one yellow orange and black car at 20:39:

This video is so well made. Insane cropping, narrating and showing the AI do work at the same time is super entertaining. I don't even play Trackmania or have a bigger interest in AI. Well done Sir.

Max Verstappen spotted 7:42

I had seen your video before the first one and I was so intrigued to know more about that but I forgot to subscribe to your channel.I m happy I get you in feed, nice to see the progress. Keep working and I just have a small question about the specs of your pc as I too want to start something like this.

train it on deep dip 2

AI using Naruto kagebunshin training technique 😂

Make it learn how to bug the game

Train the ai with TAS runs and see what it breaks to win.

I want to correct something. The Neural Network doesn't "understand" a single thing it's doing during the game. Yes it might seem so, but it isn't. If you're curious why, this would probably be a good point to start: en.m.wikipedia.org/wiki/Chinese_room You can imagine it like an cracked version of word predictions. It looks at what has happened before and now, and then based on how it's trained gives out, what it thinks "should" come next. It is just educated guessing * 100.

Eh Ai was fun at the start but it got lame over time. It's try and fail until you succeed. It's useless outside of that one part. Skill is not transferred to anything else other then that one path. Every record will fail if you give it enough time. Ai is just lame and that's about it.

COME BACK MASTER❤