The Improved Trackmania AI is Terrifying...

[@MasterHigure]

4:30 "On tech you will never gain two seconds in a single turn" You will if you race against me.

[@PrinceBaconYT]

Man I said the exact same thing when Wirtual pulled that line out of himself. x)

[@beerdrinker_6930]

That's genius lmao

[@T_ISEC_G]

he would even gain two seconds without any turn

[@alexanderf8451]

The inability to wallbang is a specific kind of problem for AIs and other optimization methods. It has a found a sort of local minima, in this case a strategy where almost all small changes result in worse performance. Because its so rare for hitting a wall to be beneficial getting the AI to do it is hard. The biggest obstacle is that a wallbang is beneficial depending on a bunch of information about what comes after it, information the AI has no access to so generalizing the skill might require a completely different method of training the system.

[@kartirnium]

This is like a chess sacrifice for positional advantage. Very interesting problem to solve where intangible metrics are really hard to assess, like the loss of material, or hitting the wall.

[@derickd6150]

Yeah I am trying to get a reinforcement learning algorithm working for my masters and anyway, in my reading I came across model based RL where a neural network is used to model the environment and its response. So for example, it tries to come up with a partial differential equation to describe the motion that results from whatever initial conditions it is given. This type of RL then decides what to do in each moment by asking the network, if I do this, where will I end up? And after that ? And after that? You get the idea. It lets it “imagine” a future without needing to execute it. Thus if it could describe the results of hitting the wall correctly, having done so a few times, then when it imagines it’s future paths at any moment, it may correctly identify that a wall bang is the fastest option here

[@joshuamaldonado1721]

Even in training this AI the creater had to force it to slide for it to learn how to.

[@hiimkyle1894]

There's a different video with millions of views about a different ai trained to play trackmania, and in it's early stages it was wallbanging all the time, because it got a short term advantage off of it, but it was updated to recognize that as not worth it

[@JaceITG]

Regarding the fact that the benefit of wallbanging usually does not materialize until later actions are taken; I think that a type of model that's not restricted to analyzing sequences in one direction (present => future) could maybe have an application here? In my thesis pursuing AI for determining difficulty of patterns in rhythm games I came across LSTMs, a form of Recurrent Neural Network which retains information about a current state for use later in a sequence. By doing so, I believe it can propagate adjustments based on changes they may have further along the temporal sequence. Doing so might have the potential of allowing the AI to learn techniques which pose an immediate negative effective upon the loss metric, but place the car in an objectively better state for future gains, which itself is an intrinsic skill for playing Trackmania.

[@ballom29]

"AI lack creativity" flashback from scattershot wich found at 2 occasions mind blowing strats in SM64, namely : - just walking : yeah, there was moments where it found than simply walking was faster than doing some fancy movements, it became a meme with that AI afterward - not swimming : it managed to get quite a good lead by not swimmign right off the bat but instead just changing mario's angle, this actually shorten the distance mario needed to travel. Mind you, scattershot wasn't competing against human RTA....but TASes, so it wasn't challenging their real time skill, but their creativity at founding better strats.

[@GunAinmNoAodann]

I've really come to appreciate your consistency in defending and supporting the content creator community. It reflects that same "no negativity here" mentality I tend to get from the TM community.

[@Xeractyll]

Hmm. Honestly curious what a no-cut TAS could do here. Anyone wanna make one? It'll give us something to compare to.

[@TheNuclearBolton]

What would be the benefit of this? How could it assist or positively affect the evolution of the A.I.?

[@PandaMane]

@@TheNuclearBolton This AI isn't trained on other players' runs, it learned from scratch with info about the map.

[@ItzHammer]

@@TheNuclearBolton You could give the same argument for TAS's usefulness with human records. It's a theoretical perfect run, that shows how far we are from fully optimizing a track. This would give something the Humans/AI developers to shoot for, and can also bring to light new strategies for a faster time. It would also show how humans optimize tracks in comparison to AI.

[@Phriedah]

You could start the TAS from Loso's run and just let it bruteforce for a while, see what it cooks. OFC I won't do it myself, I'm just an armchair director in the YT comment section.

[@Noodlyk18]

I feel like TAS will always be faster, with human assistance you can just theorycraft the absolute perfect run and just do that, even capitalize on what the AI is doing and do it better.

[@alexdacat7052]

5 years later "but then, The AI got this run."

[@karelpgbr]

I can already tell that in 8 to 16 years, Wirtual is going to make a video on “How humanity beat the robot uprising.” “Trackmania Nations Forever is a game released in 2008, the goal is to finish a map as fast as possible…”

[@TEQTONYK]

"And on the 3rd of October, 2035, nearly 4 years after 256³ Deep Fear Final Edit was released, Hefest got this run..."

[@luishernandez8121]

@@TEQTONYK "Hefest came back from his retirement only to beat the TrackmaniaNator in a final battle of humanity against AI"

[@profluffy7985]

Love to see the main channel get consistent uploads!

[@Der_Yoloist]

Isn't the main channel just "Wirtual"?

[@subpar3462]

@@Der_Yoloist i think he's joking about how he uploads so much here that it's basically turned into the main channel

[@Der_Yoloist]

@@subpar3462 ah I see ^^

[@stespin]

Mmh. 0:17 Those inputs are the closest thing to a Riolu WR I've seen so far... He was the ultimate innovator and we didn't even realize.

[@SapkaliAkif]

Yeah, kinda puts the legitimacy of the contest into question.

[@DreadKyller]

The AI hasn't been programmed to not hit the walls, the AI hasn't been told any rules specifically, the AI figured out it's own rules based on what actions rewarded and punished the AI's score. If the AI had happened to hit the wall like that and ended up faster from it, it would have gained reward and it's possible the AI would have internalized that logic and tried to optimize it, but in training the AI never encountered a time where hitting the wall resulted in more speed or faster time, and thus it never learned to attempt it. There was no manual input from the creator saying to avoid the walls, the AI learned that on it's own due to every time it did so ending up with less speed. This type of lacking in ML AI models is a well known problem with no easy solution, further training will likely not help learn this as these types of tricks are rather precise and the AI can easily fail to repeat it enough times to figure out when it's effective and when it's not. The AI has limited vision, it doesn't know the track much ahead, so it's not great at figuring out setups multiple steps ahead, it's working in a relatively small area of the track at a time, whereas we can internalize the entire route more or less. It's possible that the AI could learn such advanced techniques via a more advanced training system that makes use of Demonstration-guided reinforcement learning, where a human would be able to take over mid-run from the AI and perform inputs before switching back to the AI. The learning algorithm then compares the actions it would have chosen and the actions actually performed by the demonstrator, and over repeated demonstrations tries to get it's output to match the provided input without degrading the rest of it's performance. Even so some of these tricks wouldn't be reasonable to generalize without further information about the track to know whether it'll setup well for later down the line. Another major limiter of the AI in this case is the speed of training, currently the training algorithm does a single run at a time, or multiple runs in multiple instances, but the runs are played in real-time, meaning it takes a long time to get thousands of runs of data, especially if for each generation you're running a hundred+ different models. For many other AI projects, the AI can play the game or tasks at hundreds of times normal speed and train in a few hours what would take days or weeks to train on Trackmania. If someone could create a software that perfectly recreates the physics of the game that can be run standalone and can report out the information the AI needs per 1/10'th a second in this case (this limit currently is imposed by the way dealing with OpenPlanet works) then the training could be done much faster.

[@michaelbuckers]

The AI would naturally wallbang all the time, it's just what happens when you drive full throttle into a tight corner. And the way its dev programmed reward function made AI think it was better to hold forward and wallbang instead of using brakes, and his solution to that was giving AI penalty for wallbanging instead of addressing the error in his original reward function. Well it did fix the issue, now AI uses brakes, but now it refuses to wallbang for any reason. AI is like evil genie - 100% of what you did NOT specify, will get abused to the fullest extent. When AI is not given reward functions beyond their main goal, they work out the perfect strategy eventually. Emphasis on eventually. When people want results fast, they start messing with it. And that results in very suboptimal outcomes, though they do come quickly.

[@solsystem1342]

​@@michaelbuckers It's rewarded for speed and neoslides not for avoiding walls

[@michaelbuckers]

@@solsystem1342 Originally it was rewarded for current frame speeds. That resulted in silly amount of wallbanging. Then it was penalized for wallbanging. That resulted in not wallbanging. Then it was rewarded for going sideways, because rapidly slowing down in a neoslide conflicts with reward #1. Then it was penalized for going sideways at slow speeds since neosliding produced more reward than going fast, so it would tokyo-drift in place. Now he needs to allow wallbanging but such that reward #1 doesn't undo everything else. It's a huge mess because the author thought the appropriate solution to producing good AI behavior is slapping more and more hand-crafted rules on top of each other, not just giving it freedom to do whatever it wants with very small auxiliary rewards to get it going from a standstill.

[@galoomba5559]

@@michaelbuckers If i remember correctly, the author removed those extra rewards after the AI learned to neoslide.

[@yikes6758]

@@michaelbuckers the problem with not reinforcing any behavior is that it will fall into a behavioral pattern that would be suboptimal, but it doesn't have the broad re-strategizing capability to optimize that much better once it's already determined an effective route. you also can't penalize or reward behaviors at specific parts of the track, it's all or nothing. so if it determines that wallbanging gives speed early on in its training, it'll carry that forward and the dataset will bias towards wallbanging as it keeps wallbanging for speed; a feedback loop. it's like how black neighborhoods have higher crime rates, but that's because they're over-policed, and so crimes are found more often by the heightened police presence. then that statistically higher crime rate is used as justification to keep policing black neighborhoods. police will keep going to black neighborhoods to look for crime, thereby creating the statistic.

[@kachelstacktus]

6:13 Could you upload the interview with Mudda? I have watched that run probably half a dozen times by now and would LOVE to see his explanation of how he approached the track.

[@alexanderf8451]

I thought this was about the AI that managed to finish Hockalicious.

[@kyloben4848]

not just finish, dominate

[@DEADB33F]

Super impressive, but I'll save my true amazement for when AI can enter a random TOTD competition, learn the track in real-time and regularly be winning tournaments.

[@zonanimz9990]

Posted right before I had to go to bed 😭😭

[@Humphertdinckle]

Same but I gotta watch it

[@Pajzsika-bpl3x]

Same but I gotta watch it too

[@julien1978-2]

Yes posted right before I slip under your bed and watch you sleep so annoying 😢

[@ElZedLoL]

​@@julien1978-2MonkaS

[@cosmicwonderer6269]

Good night

[@potatosordfighter666]

"This just in, software built for the specific purpose of doing a thing well does a thing well. Next, the weather"

[@clowndamian4354]

I been seeing trackmania videos for 3 days now but i can just tell your a fantastic youtuber

[@ErnieZee]

Isn't he? So happy to welcome you in the community :) I don't know if you know that but that is Wirtuals second channel. The main one is called just Wirtual. Uploads are pretty rare over there but the videos are super high quality. Streams on twitch are worth checking out as well :)

[@clowndamian4354]

@@ErnieZee no i didnt thx

[@cubicinfinity2]

It's actually interesting now. I didn't really care when it was in its infancy. I look forward to the introduction of new track pieces and elevation.

[@stevenhaeding4955]

The AI inputs look very similar to trackmania legend riolu doing a slowed motion speed run sped up. All those minor adjustments do look suspicious 😂

[@triplespace007]

Who's the other KZbinr he mentions around 7:46? I've looked up what I thought I heard and nothing showed.

[@entityeoe2285]

I was waiting for this after I watched the Hockolicious ai video or at least a similar related one

[@Napert]

the thing with ai is that while you go training for 70 hours per week, you only gain these 70 hours, while an ai not only can do full 168 hours per week, but also can just multiply the training time it does so the ai can end up doing 4 hous per hour

[@lucascarvajalkossnar8630]

I fucking died when I saw it purposefully get a ramstein bug.

[@bjorn9875]

I think that the question of AI in a game like this is also a question of if it should be "generic" (read screen, but not "remember" the track), or more specific, (know the track, and optimize for that track).

[@Kyrelel]

If it memorised the track, it wouldn't be considered AI

[@QueentasticTDP]

The day AI learns wallbangs its over for humans

[@DaftFader]

The AI is starting a new Track Mainia Clan. It's called "Intended Way Only", and all 40 current members are the same single AI. 😅🙏🤣🤣🤣

[@jasonorjoshlee7607]

We'll be seeing AI beating plasma in no time.

[@NeesyPlaysGuitar]

It was over before you even noticed it had started lol

[@cheezo1372]

I don't even play trackmania. I randomly came across the game from one of those deep reinforcement learning AI videos like a week ago and became a trackmania viewer lmfao. I think that one jazzy song you play in the videos helped lure me in too

[@gxzmic]

At this rate calling someone a bot will become a compliment 💀

[@JakGruen]

I just recently watched a video about a far more advanced TrackMania AI. I'm surprised it's not part of this video.

[@Clover298]

The Trackmania community is learning the joys of beating the Tas

[@MFSkyline]

What if you taught the ai to shortcut? (On a different map)

[@frkcybtw]

basically just TAS at that point

[@Pablo360able]

The problem is you can't really do that with reinforcement learning. At least, not without already knowing what the shortcut is, how to execute it, and what incentives to give the oversight program to encourage the AI to improve execution on that shortcut.

[@catdisc5304]

​@@Pablo360ableyeah, the model would basically have to be trained for each track individually with specific rewards/goals in place for not driving on the road and figuring out a shorter route than the intended route. Sounds pretty difficult imo

[@xavaxkenka5802]

I would be so interested to see the interview with mudda about deep fear, I would listen to that so attentively.

[@ChuckSploder]

I thought you were going to beat Yosh's AI

[@cjay880]

just wait until the AI starts looking like a tas run.

[@lilthreadd]

i dont think thats ever gonna happen tbh, maybe on some very simple maps

[@baba0rum]

​@@lilthreaddnot ready, wait until the ai find nosebug

[@ragerancher]

I think it will be interesting if the AI can be developed to learn a very general set of skills as opposed to learning a track, eg learning how to wallbang, neo slide and other tricks and then retain that on a completely new track if it seems similar conditions to where it was beneficial before. It would also be good if there was some way to allow an AI to memorise the track as a whole rather than in small sections (like maze robots do) so it could attempt to optimise a route based on inputs it may not have got from taking snapshots of the track. Imagine then if, rather than putting the AI around tracks, you could instead put it into dedicated training sessions where it learned how to perform certain skills and could then apply those to any track, in the same way a human would.

[@johncasey9544]

I'd bet that's entirely possible, but it would require a decent amount of training data and compute, which may not be something anyone wants to bother putting time into. Seems like with current ml generalizing is a problem of data volume.

[@maxsurman7838]

You'd simply have to randomise the track it trains on each run, from a large selection of tracks.

[@darthnovii]

There are two things you should never challenge a computer on: fast arithmetic, and consistency.

[@chickenmaster3879]

It's really annoying how people say that the AI is terrifying when they're using something like this as an example. TMNF has four inputs, five in niche situations. This simplicity is aboslutely perfect for mahchine learning, since progress speed is determined by not just how well the base code is, but by the complexity of inputs (screenshots in this case) and outputs (only 4), as well as the cleverness of how the AI is integrated. So the screenshots are almost certainly monochrome and lo-res so the AI doesn't get "overwhelmed". Of course an algorithm that slowly optimizes and does best with simple inputs and outputs would do well in a 5-output game, while only using one type of road and having the track be flat. This AI initially trained for about a half a day, slowly optimizing a pretty good run that was crushed by top Trackmania players in only a couple of hours. These AIs aren't terrifying, they're just slowly reiterating over and over. I get that the title was just light clickbait, but this only fuels the AI hysteria going on, and it isn't helping.

[@APaleDot]

To be fair, they also uploaded a video of this AI getting within a few tenths of the world record on Hockolicious.

[@yourface2464]

I don't think that's the intention behind the word terrifying here, but I get where you're coming from. I wonder if the entire AI panic is even a thing overseas, much like how alien panic isn't.

[@henry51488]

If you want to argue about how it is "terrifying", talking about what it does now has nothing to do with it. "Terrifying" implies that what the AI will do next or in the future have negative consequences. You can argue that it is physically impossible for the AI's future to have negative consequences, but I think it is a rather pessimistic look at a "technology" and also much harder thing to prove. It's much easier to look at other "games" that AI has "solved" to give us some possibilities. We can take chess for example and ask "will the game be changed if an AI can give real time feedback during a run or if an AI can exhaust all possibility of shortcuts or routes?" I'd argue yes since these are some of the major aspects people enjoy in this game. Imagine if finding shortcuts, fastest ways to go through a map, etc. are all possible to "solve" as soon as a map is created. The community loses the suspense of not knowing what is the best if the best route and best ways to make each turn can be calculated or exhausted. This would remove a major part of the fun the game offers.

[@SilentTM]

and still this algorithm sucks, if you put wirtual against an AI and give each '5 runs to optimize the track', Wirtual would crush

[@SimonRika-o5t]

How many millions of trackmania players crushed this AI? Oh, only a few of the very best human players managed it? And you don't see the correlation with the "real world" problem of AI's performing far better than the average human and thus making the human redundant? Also, how much effort and money was put into developing this Trackmania playing AI, vs the kinds of AI's Microsoft and Google are spending billions to develop? Seriously, this is basically someone's home project outperforming most humans after only a few hours of training and you seem to think it is indicative of the true state of the commercial / governmental AI's being developed?

[@mateopanza9350]

you should watch yosh's video, where he makes an AI until he can't beat it

[@mathieuaurousseau100]

Do we have any idea whether that AI would be good on any other map or if you'd have to restart training from 0?

[@uniquename6925]

I'd be willing to bet it only works on that map.

[@ninjablade2]

More than likely it'd start again from 0; though the programmers may have improved on their training methods since the start of this one, so it could be technically starting from "more than 0"

[@alexanderf8451]

There was a recent TMAI that used similar techniques and was able to complete multiple tracks, though it only performed at a top human level on the one it was trained on.

[@quentind1924]

I watched another video of an AI driving a track, and at some point in the video, the AI had to play the same map but in reverse (the map was flat with no jumps), iirc it was good, but not as good than on the normal map. And honestly it’s pretty locical and it’s the same thing for us : When we play a new map, we didn’t had any time to train on it so we are not as good than on a map we know

[@flaming_ace]

Would be terrible on a new track, but would learn it much quicker than a fresh LLM would

[@andreasthiemke9520]

I wonder what the use cases of AI in TM is? Maybe if you could train it in a way that allowed for shortcuts and reroutes, could you then use it to find new shortcuts/reroutes, kinda like what JAV( I think?) did on E02.

[@timma_j]

Is there any chance we could get the full uncut mudda interview?

[@psychoedge]

Would be nice to include a link to the video Wirtual reacted to

[@jomo2x]

Sooo is the rammstein bug intentional near the end of the video from ai ? because this is kind of similar to wallbangs in a sense

[@houstonisfake]

It looks like the ML model was trained on this track which disqualifies this as an example of how good this model is. Let’s see how it performs on an unseen map

[@SpaceSwimmer69]

People farm wr on the maps. So why ml model can't?

[@rowcla]

@@SpaceSwimmer69 It being non-general makes it less 'ai that can play trackmania', and closer to 'ai that can slowly build a TAS for a track'. It's still somewhat interesting, but a very far cry from something that could apply itself effectively to a variety of maps

[@SioxerNikita]

​@@rowclaBut... It isn't a TAS, even in the slightest... In the traditional sense. It is not just "memorizing" the map

[@goffe2282]

So what would be interesting to you? This is an accomplishment even if it's only one map.

[@ryanh7167]

​@@SioxerNikita it literally is given access to frame perfect execution of actions every single frame. It literally is a TAS in the action cycle of the perception-decision-action process.

[@cadthunkin]

This is really fun to watch as there are so many possible solutions. The physics behind the game seem real enough to capture real driving. Everyone is calling their program AI these days though. Really this kind of thing is "generative design", because it does try things and play survival of the fittest, but its not real time. A real AI would drive it the first time and convince us it was as good as a human. Then get better and so on. It does not get a million tries, and choose the best one to convince you, see the difference? Either way, I'm impressed by the game model and the fact that it can do generative design. The calling everything AI is clickbait IMO, it got me :)

[@krsp420]

interesting if the ai can do it with 'human like' inputs. whatever that would mean. But to me, its basically a generative tas until it can be within sorta human inputs per second.

[@mateusfer7668]

ohh poor boy you dont now .;z

[@leaf821]

Day 2: Wirtual should try out Zeepkist, basically a mix between Trackmania and soapbox racing

[@Der_Yoloist]

100%

[@zivecef]

absolutely

[@areks.]

❤❤❤

[@kingacrisius]

I would say 95% soap box racing, 5% trackmania. They are barely similar and basically none of the skills transfer.

[@leaf821]

@@kingacrisius check out some zeepkist author medal runs. Few skills may overlap from Trackmania, but it is still an extremely technical game

[@CRJCrombo]

Wirtual on a mission to defeat the AI 😂

[@PepekBezlepek]

our boy Loso saves humanity once again ♥

[@neptune_main4049]

It would be very appreciated if you would link the channel or video this entire concept is based off of in the description

[@nickbaker5296]

on the riolu cheat video you had a tap rate threshold used to find other cheaters too. that should be the button press limit for the ai if we are really trying to fairly evaluate the AI/Human viable speeds/limits

[@Overlycomplicatedswede]

Can’t wait till this AI can nosebug

[@MatrixCL]

"The AI can do great things, but it lacks creativity." That sums up the difference between AI and human in all fields. Just saying in case anyone is worried that the AI will make us all redundant at some point. ;) (speaking as a software developer here)

[@aimo4thewin619]

You can see is hours played, he is so insanly improving, on his on record, respect

[@scottwesner9362]

Need a shower but this comes first.

[@-Phoenix_22]

With just 12 hours in the game I finally got all the gold medals on fall 2023!! Time to grind for authors!

[@cola98765]

Fortunately for real players kacky maps might be still out of reach of AI like this one. AI tends to reach local minima, this is why there was requirement to do no cuts, as even if there is consistent setup, AI like this will never go "let's cook with this" after going over wall.

[@StuermischeTage]

I'm unconvinced. Have you heard of the "low-input strat". The idea being to solve Kacky maps with only a few button presses. Quite a few Kacky maps have been bruteforce-solved, where they only require 2 to 3 inputs and result in a reliable finish.

[@cola98765]

​@@StuermischeTage In such cases AI would need to "know" that it's a low input map, which is equally as hard is the setup is not intuitive... Like that fall in the last Kacky tournament, where you needed to hug a wall, hold full steering, but only to a specific point. AI in this style is TERRIBLE in doing such actions, and would sooner arrive at end with many input pseudo-bruteforce solution.

[@BenGyaaattttt]

Is the AI better than humans at first timing maps? Or does it need to train on every map it attempts?

[@beninator5533]

Ai is basically just TAS. making a TAS is reinforcement learning done by humans. Now the computer can do it better but it takes longer.

[@snowe..]

Pretty sure yosh’s ai is the better of the two. It’s less overfit, and will therefore perform better on more tracks.

[@levgorlov3566]

Bro posts when im justing getting comfy in my bed

[@FerousFolly]

I'd love to see this AI thrown at the 20 minute speedtech map in trackmania 2020 map

[@Theonicle]

trackmania drivers stand no chance if they manage to make ai be able to do this on deep dip like maps

[@JohnBLZ]

I love how literally every single piece of content about AI is: "This new AI is terrifying. It's so much better than us. The human race is doomed and will face extinction LOL"

[@SnowTerebi]

You need to watch better AI content. Two Minute Paper is a great channel.

[@JohnBLZ]

@@SnowTerebi I'm not questioning the content. I"m just saying it's funny how most AI creators present AI as pretty much the end of mankind and then go down the rabbit hole with a smile

[@Scorpionmkganja]

I think we have to call pablogd for this one

[@mlittlej101]

Can't see anyone else mentioning it, but there's some low-frequency, intermittently patterned artefact in the audio in the "follow-up" section whenever Wirtual's microphone audio is being played. Sounds a bit like some transformer buzz, or some sort of other electrical interference being picked up by the microphone. It's really obvious in the last 5 seconds or so when there's an audio break in the outro between "see you in the next video" and "have a good one". Something to watch out for in future maybe? It's fairly distracting/irritating.

[@pixelmalfunction1772]

there is a overwall cut on the map skipping the neos i got a 2 min flat on my yt even tho my driving was vry bad

[@gergelypaless5042]

I like how Wirtual doesn't have any clue how AI works, for him it is just a black box with dark magic in it. Luckily I can already see useful comments on the topic

[@theracer_256]

AI can be very optimized, but it will never slam his head into a wall

[@the1barbarian781]

Can you link the ai video in the description?

[@nameanteater4772]

How does he know to wallbang? Is there a way of knowing or is it just trial and error and a bit of luck?

[@taliesine.8343]

experience If you see many wallbangs on different maps you can sometimes get a feel that there might be a wallbang. Then you just try it and if it doesn't work you move on

[@n45a_]

i would like to see a TAS on this map

[@givikap120]

TAS is like humans + AI at the same time, so no chance for both

[@abraxas2658]

would appreciate a link to the original video 🙏

[@destro6424]

Could Ai driven runs be classified as tas?

[@JPK314]

It's not tool assisted, it's just tool driven

[@marlewcal8736]

Id love to see this ai with human limitations implemented loke it cant click faster then a human etc it can only drive using imputs that are humanly possible and see how good it does

[@gradualdecay1040]

You're pretty good at this. 👍👍🇬🇧🇬🇧

[@Aria_uwu]

it gets to the point where it wont be beatable because its just becoming a TAS

[@CubeStarMaster]

Who knows maybe it will be better then a TAS and take less time to.

[@Aria_uwu]

for this thing to come close to any tas times on any maps it would take thousands or tens of thousands of hours of learning because its supposed to be driving not flying around the screen it would only peform tas methods if directly told to do so @@CubeStarMaster

[@Male_Parent]

I saw the twitch clip. Bonkers record.

[@zerorusher]

Im interested is seeing trackmania AI being trained with some constrains: limited taps for second for example. This constrains could lead to new human usable strats

[@georgesnow_]

Yes, i was waiting for this!🎉🎉🎉🎉😊😊

[@pittypatterputzzler5311]

This is nuts, dud has skills

[@hopeisenough1317]

Is this AI human repeatable? or are the inputs too fast?

[@youtubeSuckssNow]

The inputs are not human repeatable

[@pasu-_-404]

Wirtual when you are gonna do a face reveal?

[@thesupertight5684]

R/wooooosh!!!!!@@halloweentreehouse

[@Hq123Hq]

You guys think Wirtual is going to be a male or female? Can't wait to see the face reveal.

[@philmazzie3674]

His bi face is already shown whether he "reveals" it or not

[@doggo7078]

AIs will not be capable of creating new ideas or routing new strats. They reinforce their skill on 1 map and cannot think 'hey, what if I bonk here and skip this zone', that's why they cannot beat humans

[@parveenkusar1986]

Can you make a advanced guide on trackmania 2020

[@COIN_LuigiPres1]

used a cut lol 1:59 something (i suck)

[@tamimyousefi]

This is the most ironic way to win against the AI, as slamming against walls is litteraly the first 'trick' it learns simply because it didn't figure out how to steer. Even after learning steering it likes to slam against the walls because it's faster than bad steering. People then artificially change the AI mechanism such as to avoid slams to incurage better steering/sliding skills. Then the AI is stuck with that behaviour for a time that is much greater than its training time.

[@mihalydozsa2254]

I wonder if it would be possible to beat it by slowing down the game, since it also can react faster than any human... Also there is an other AI that is getting better, and it can drive on multiple maps.

[@maxsurman7838]

It's already limited in that it can only make a decision every 0.1s.

[@taliesine.8343]

I would find it very interesting, if some people tried to create an AI that can only use as many Inputs as humans can average per time unit. It would be interesting what lines they would take if they could only use as many inputs as we do.

[@lloydnoz7093]

I wanna see this AI on A01 race

[@Daragonal]

WHO IS AL AND WHY AM I SEEING HIM EVERYWHERE?

[@Stand0t]

tas?

[@Veptis]

They confuse test and training. So where is the generalization across the whole maps. Vision based makes much more sense than those low amount of crafted features like distance to road.

[@phoenixking339]

What about yosh's video

[@bautizhg7217]

Illegal wall bang

[@edsknife]

But can AI have personality?

[@SimonRika-o5t]

If it is trained to have one, yes, it can.

[@besher532]

I'm surprised by the amount of both armchair AI experts and actual people who are knowledgeable in AI in the comments

[@Godwinsname]

Is it weird that I absolutely do not care to see a computer program play a computer program? I never cared about TAS or beating AI-trained driving or w/e.