Tu veux pas la sortir en français stp

looks like infinite timelines upon eachother.. mesmerizing.

now you gotta build an AI that tries to builds an unbeatable track

This is soo interesting

Now when I loose on multiplayer games I can blame it on chaos theory.

microscopic changes affect so much because you set the landing in the middle of the pipe, this is a point of unstable equilibrium. If you intentionally land slightly to the side, while accelerating towards the middle, then you can decide on the fly how long you need to accelerate and thus compensate the randomness. The trick is to change the tactics to lay all the randomness range inside your controllable range.

T'is but a pipe dream. Xd

This is art. Beautiful.

Initial conditions and numerical precision

Does anyone know the genre of the music used in this video?

put....put this guy on Deep Dip!

Haha just watched a video on the three body problem and chaos theory, and I though this would just be good AI fun but turns out to be very relevant to this as well!

wtf Ai mean niagara mesh ?

22:47 nice homage to wirtual!

Amazing job with this. Mammoth effort

Incredibly well produced!

This is the best video to understand AI. Just thanks !

Can the AI teach my wife to drive?

Those happy lil cars bouncing on the pipe makes my inner child happy. Bounce bounce

I think taking the AI's best run against yourself is a little unfair. Taking the AI's most average run out of its high number of attempts would be a more fair representation. Once the AI's average run is better than yours, then you've created a monster.

Truly it's an excellent video; totally enjoyable and a masterclass in both video production and tenacity. But everyone is calling this "AI". But is this yet another misuse/overuse of the term? Is this neural network really intelligent? It demonstrates machine learning of a sort; but it's also very guided learning. Without someone tweaking the parameters and making _intelligent_ decisions on its behalf it was clearly inferior. And I'd argue therefore, not intelligent in its own right. Number crunching, building up large data sets with significant nudges from an intelligent 3rd party, and using those data to drive better doesn't to me show any intelligence. It's just (directed) brute forcing a solution. Most humans (at least those an hour or so of playing TM) would be able to get around the track after only a few tries using visual stimulus and knowledge of control inputs and physics. They would be much slower, but this to me is significantly different from how this machine 'learns' the track. True; those humans have probably spent years learning the skills in completely different contexts. But the point is that they can apply a build up knowledge to come up with solutions to new, unknown problems. That's intelligence. Again; it's a fantastic video, and I've watched it multiple times, completely fascinated an in awe. But I still think "AI" is overly misused in contexts like this. As it increasingly frequently seems to be when people play the "FAKE: It's AI!" card on UKposts shorts that use CGI (not even machine learning). I don't think people really understand what AI is supposed to mean. And that's totally reasonable and expected in a way, because: 1) I don't think AI even exists yet. 2) Not everybody is close enough the technology thrown at it under this name. 3) The philosophical definition of _real_ AI is incredibly difficult. 4) Even machine learning isn't generally understood by _anybody_ . The concept of a neural network and how to set it up for machine learning is. But once it starts it's 'training' I don't think anyone really understands what is actually happening. Just like computers now design even more powerful computers that are somewhat beyond human comprehension. Maybe this has changed in recent years; and I totally admit that it's not an area I have any expertise in. But I have worked with people who do have that expertise, and all of them are very cagey about using the term "AI".

just an fancy automatic TAS imo

Just wondering, do you AI to edit/parse/finish your videos?

Lihat lah para ai hijau di jalur 3 itu, mereka semua berhamburan. Dan tau apa yang lebih keren dari itu? Ya, mreka semua ai hijau, dan bukan ai merah😅

i wanted the ai to just 2 wheel perfect balance the first track so badly

I don't want to point the obvious, but can floating point numbers precision (or lack of) be another effect?

Interesting in regards to cryptography too

28:26 finally forces the fucker to slow down a bit and suddenly pure consistency lul. AI was just way too fast at first lul

27:50 it’s so confused 😂 it’s like “do… do I go forward?… **is immediately punished carrot wise** no… no i go backward :)”

21:05 idk. I don’t feel it random, I feel it’s chaotic. Which means it’s predictable, but just hard to predict

@joshtm Could your experience be used to train the AI of the new Autonomous Racing League (A2RL)? Looking at the results of the other teams, it looks like you are 5-10 years ahead of them. Is their exercise much harder than yours? Do you think there is much difference in the inputs/outputs?

Thank you, this is amazing. Wow!!!

The way you train AI is good for choosing tactics, but not necessarily strategies. In other words, your way of training ensures that the AI in the next step will choose the best solution, but it does not ensure that the whole series of these steps will be the best possible tactics in the long run, because simply this way of training the AI does not take this into account.

Ah great! Now make it do deep dip 2

yoyoshi

The chaos you see in game is probably cause by floats (floating-point numbers) if they are used to represent values in the game, then it would make sense for the car to behave chaotically.

The term “AI” is used too lightly now….

Can you teach the AI to drive on just one pair of its side wheels the entire length of the pipe?

Why can't the robot repeat the same strategy without consistency? I'll try an answer this before you do: *because it keeps getting rewarded for trying different things*

Intéressant, intriguant, étonnant, fascinant, et même émouvant... Quel travail d'immense qualité, merci de tout cœur !

24:04 (it also seems to be using an analog input with no input noise, i.e. infinite precision too, which I'd call just as important as quick decisions)

What I've noticed so far (8 minutes in) is that the AI indeed (as you said, yeah) can only change its input 10 times per second. With a game that processes 60 times per second, that leaves a whole 5 frames for something to go wrong for the AI to be unable to correct for. However, I suspect allowing the AI to input 6x more frequently would make training take 6x longer...

USE A SOFTMAX

Real life AI driving you would win at ukposts.infoTPzBH-7ckO0?si=YEEf2BG8cfsEyg24 ❤❤❤❤❤❤❤

21:32 I think your memory between inflections is too short for that large maze. From what I could tell it was using C_n and D_n variables to sort of make a map in its memory, based on the states of these variables? If this is correct, then you might need to extend it a tiny bit to allow for more combinations. Maybe then it can learn vertical tracks as well.

amazing video!

Are you french?

Bravo, quel travail !

Est tu francais

make them drive upside down please