Sony’s racing AI destroyed its human competitors by being nice (and fast)
technologyreview.com
technologyreview.com
> GT Sophy doesn’t get the same view of the game that human players do. Instead of reading pixels off a screen, the program takes in updates about the position of its car on the track and the positions of the cars around it. It also gets sent information about the virtual physical forces affecting its vehicle. In response, GT Sophy tells the car to turn or brake. This back-and-forth between GT Sophy and the game happens 10 times a second, which Wurman and his colleagues claim matches the reaction time of human players.
The AI is getting precise 3D position data and physics data not only about its car but competitors cars - from a 3D physics simulation. The human competitors only get a 2D visual view based on what a race car driver can see from the cockpit. Maybe I'm misunderstanding but it seems like the AI is simply deriving the optimal parameters encoded in the 3D physics simulation through iterative testing based on precise 3D physics data supplied directly from the simulation. The humans aren't getting the same data. The human drivers must visually estimate the physics vectors and precise 3D positions from a limited yet rapidly changing 2D rendering. This is quite different from what I initially assumed based on the headline and opening paragraphs of the article.
Given the same precise 3D and vector data in searchable form along with sufficient compute power, it seems plausible a team of human statisticians could work out more optimal driving parameters than top human drivers limited to 2D visuals in nine 24 hour work periods (which is what the AI required to train).
AI models correlation(Measures(X), Measures(Y)).
Marketing sells it as using a theory of how X causes Y.
In this case, we want a machine to acquire a causal model of how its actions produce effects, by measuring X and Y itself under its own actions.
There is no AI that does this. What we have is a machine "with the answers already" (ie., no need to measure) that has a correlative model.
When it matters, the playing field will be stacked heavily against the human who will not have direct access to the informed encoded in the virtual world. Putting it differently, any AI in a simulation is in some sense indistinguishable from the simulation itself (or at least, you should assume so).
I believe that it would be naïve to believe the owner of the AI (whether a person, and organization, or another AI) would play by those rules. Better to assume an adversarial game board.
One approach would be to simply not compete against an AI even in our world. Another would be be to be in control of the framing or environment where you can outplay, outlast the AI. Do not assume friendship because its goals will be different from that of humans, unless you can be certain of mutual dependency.
The game theory around human/AI coexistence extends beyond these quick thought bytes, maybe someone is working one a treatise.
The AI should be given a game screen. I think it should also need to actually use a mouse and keyboard too if im being honest.
I would be way better at games if I only need to think about what I want to do in the game and have it done rather than use some method of physical controls to do so
I'd go one step further and say those mouse inputs should be "fuzzed" a bit as a function of "actions per minute" because humans lose accuracy (both in the time domain and spatial domain) when they try to make more actions faster and faster. Obviously APM needs to be capped as well, preferably actually slightly lower than human APM.
The main reason for this to me is that it will help guide the AI's towards generating strategies which humans can actually emulate and learn from.
If you want to have an AI-only tournament, then by all means make up your own rules. If you want to say your AI "beat a human" then you have to be playing the same damn game.
The benefit of keeping things in "their world" is that "our world" requires massive investment in sensors, while the exact same algorithms would be used. Letting them play in their world means you can focus on the algorithms. But, I agree that it should be fair. Putting a human in front of a screen with a steering wheel is in no way comparable to the sensory experience of being in a car. I'm fine around a race track, but racing simulators are nearly impossible for me.
If you imagine a real car, this is the sort of detailed sensor data you would get from it. The sensor data in the game is never wrong though, as it comes direct from memory.
The other thing which caught my eye was this:
> For example, GT Sophy often drops a wheel onto the grass at the edge of the track and then skids into turns.
This works consistently because (again) it's not the real world where the grass has subtle variations.
It would be very interesting indeed to see such an AI compete on a real track, but it may not be as clear cut a victory when the very finely tuned senses of a human racing driver are in play.
> This works consistently because (again) it's not the real world where the grass has subtle variations.
Well I mean maybe they just trained it on Initial D anime, Sony is a Japanese company after all
Roborace[0][1] was an attempt at an all-AI racing series that would be run parallel to Formula E (the all-electric single-seater racing series); however, while they did get to the point that they could run cars on track, they never got to the point where they could run those cars in competition. It seems to be defunct, having been renamed "Arrival R" after the core technologies were purchased by one of the teams that had intended to enter the series.[2]
1: https://en.wikipedia.org/wiki/Roborace
2: https://www.autocar.co.uk/car-news/business-tech%2C-developm...
What this AI was doing is something that only works in sim-racing. It's done by humans sometimes in iRacing. Drop a wheel on the grass to make use of the fact that the sim will cause a consistent rotation (yaw) from it. Which can help turn a car into a corner.
It can't be done in reality, because the effect on the car isn't predictable on a real race track.
Maybe a driver could work it out in training for a single race and in a certain corner but the cost of getting it wrong is not only not finishing the race but also having a race car in the barrier, so a damaged car.
On a race track, drivers usually know characteristics of every corner for selected weather conditions. For example: at the Nordschleife, exit of Foxhole into Adenauer Forst is safe to cut in the dry but a no no in the wet, unless you’re in a GT3 class car with enough downforce.
It is the same with kerbs, stay out of then when it is wet, use them in the dry. But it’s all touch and go when the conditions change and sometimes that leads to a crash.
I saw a video a few months ago of people running an endurance race in iRacing, and they'd consistently drop their left-side (IIRC, it could have been right) tires off the track on a specific straightaway. My understanding is that a) they didn't lose that much speed, b) because of the reasons stated previously, it's fairly safe in the sim, and c) it kept the front left tire temperatures closer to the optimum for the full duration of the lap. Thinking on it, one of the things I found cool was that, in a way, it's not entirely unlike what real drivers will do with wet-weather tires, where they sacrifice the best line to hunt for wetter stretches.
GT Sophy is receiving the position of its competitors (which is rhetorical statement because of course it needs to know their position), but where does it say that it's receiving the physics data of its competitors? It does mention that it does receive 'the information about the virtual physical forces affecting its vehicle', but only the forces affecting its vehicle and not the forces affecting the other vehicles (which, IMO, would be irrelevant and, quite possibly, impossible to process given the number of cars on track and the limits of its game loop processing time).
Ultimately it has no impact on professional sport at all. I watch motor racing entirely because of the humans. It is triumph and tragedy.
But as a nerd, I love seeing "AI" beat us humans. It is supposed to.
I have about 25 years in the games industry and have worked on a variety of racing games. The "AI" could literally be replaced by any technique we developed for solving this very problem in the last 30 years that has access to the underlying game data, e.g. the physics models and vehicle positions. Until the AI has the same information as the player, it is not an equivalent AI. It is simply a very clever aimbot. Any of the classical techniques would also pretty much destroy any human player too, those techniques are usually turned down from 11 to give each of them a personality or make it realistic for the human to win.
This is no different to an RTS "AI" bot destroying a human because it can mash a virtual key a thousand times a second and has complete information on the entire map and all units and all orders issued and can even see through the fog of war.
I congratulate the team for what they have achieved, but this is more marketing fluff than meat.
My bonafides as to why I hold this opinion on this subject: About 25 years in the games industry as a systems (graphics/engine/gameplay) programmer that has worked on many triple-A games and about 20 years in AI/robotics/machine learning with a MSc CompSci and an MSc in applied AI.
Also FTA: “You don’t want to do that because you’ll make a mistake. It’s like a controlled crash,” she says. “I could maybe do that one in a hundred times.”*
I don’t see how a computer racer that destroys human competition would make the game more entertaining, I guess there’s work to do there to not make the AI reliably pull of highly risky moves.
Well, if we left things to non-Human racing we'd soon realize we're severally out matched with enough data and an unlimited engineering budget: Adrian Newey's X2014 launched on GT 6 and was the first example of this [0]. I'm not sure if it's just about destroying the competition making people lose interest, or more like breaking the 9 minute mark at the Nordschleife and seeing that it can be done and raising the bar.
I guess I can see there being some novelty in that, but I'd get tired of it personally since I've suspected and always kind of knew this time would eventually come.
I'm convinced at this point we can build a Formula E car with enough on track data to break lap records that can never be achieved by Humans in any of the previous F1 iterations; it's the same when even the best fighter pilots cannot withstand certain g forces that allow drones/unmanned pilots to make aerial feats that would kill a Human.
Striving to be 'that guy' who got closest to the AI has it's appeal and I'm guessing Sony think so too? If the goal is to sell as many copies of the game and extra DLC, proprietary driving setups/steering wheels, online memberships, Sony TVs, PS5 etc... I think this strategy makes some sense to Sony. People literately spent billions on Fortnight upgrades in it's first 2 years [1], so it stands to reason if you can get a DLC for a specific car/setup that makes you competitive against the AI it could work.
0: https://www.wired.com/2013/12/red-bull-x2014/
1: https://www.theverge.com/2021/5/3/22417447/fortnite-revenue-...
“This back-and-forth between GT Sophy and the game happens 10 times a second, which Wurman and his colleagues claim matches the reaction time of human players” already may mean the AI is cheating.
A literal interpretation has the AI get live data every 100ms, and react as soon after that as it wants (given current CPUs, that could easily be within a ms), while humans act more as if they receive data from x ms ago, and react y ms later (with x and y in the tens of ms)
Well, I've played the endurance rounds with some friends on GT when we were in the off-season and rebuilding our cars so we pulled out our bucket seats from our cars and had the wheel and peddle setup to have a laugh: these are the kind of races that take 3+ hours with drive swaps in between and have gotten close to winning on completely worn tires only to then be clipped from behind from a lapped AI car and lose it all a couple of laps away from checkered flag.
It's frustrating, but we tried again that day and got the award and I think Sony knew that it's loyal customer base always would, which is why it's moving in this direction without even hiding what it's doing.
If you want pure racing simulators without all the fluff, perhaps iracing is exclusively where you'll have to go for that now.
I'm not exactly sure what to say about it now that I'm older with less time and patience (and perhaps jaded about the video game Industry as a whole) for such things or the fact that I study AI/ML but I see this more as a data-driven iterative step towards leveraging it's brand in order to maximize engagement Online via social media [0] and sales via video games new business model (DLC and other misc purchases) than it is about providing the 'The Real Driving Simulation.'
Which makes the intro of GT2 (still the best in my view) [1] all the more ominous with it's choice of music being a form of Machiavellian foreshadowing and crushing poignance.
Once it knows how to race you can restrict it. Have it be less aggressive, artificially lower its car’s power, tell it not to play to the edge of physics.
Gran Turismo 7 is a lot of fun, but the current AI players might as well be on a slot car track. They practically race single-file taking the same lines as every other car. It’s clearly not real.
Injecting virtual Sophies, tuned in various ways, may make the game feel more like a real race. It would also allow you to provide a much tougher challenge for the top end players.
Indeed that looks like the racing version of aimbots.
I remember a thread on HN before about AI in guys. A game developer said that at some point if you make enemies too smart, the game becomes more frustrating.
"AI beats players at game X" is less often a practical advance (e.g. classic chess engines already beat casual/semi-serious players, what's the need to utterly "destroy" them, really?) and more often a media stunt aimed at impressing potential customers for more profitable uses.
Maybe not a fun challenge to race against, but a benchmark to check your performance. Like modern chess engines.
https://www.youtube.com/watch?v=nZbA_CDoWrk&ab_channel=GRANT...
Don't let the hardcore simmers read this!
That’s only in phase 1.
Afterwards when humans are outnumbered by vast bot swarms, and all their content and capital is vanishingly tiny, these norms won’t matter. Also, human participation won’t matter, and will likely be similar to the paralympics or women’s tennis (but orders of magnitude suckier).
Entertainment for a tiny niche of the market (humans performing for humans) will not be an economically significant space, kind of like classical music today or better yet renaissance music.
EVERYTHING you are used to online — including these comments - can already be convincingly deepfaked by adversarial networks trained on throwing bullshit at a wall and seeing what sticks. And this bullshit can actually be a superstimulus that outperforms genuine human interaction. What is missing is decentralized botnets swarming at scale to gradually and unrelentingly amass karma points across the swarm, and use them to move sentiment, markets etc. but that’s coming. Afterwards it’s onto subverting democratic systems with sybil attacks, as well as judicial systems with parallel construction and deepfake “evidence”.
Bullshit. State of the art chatbots can barely convince one mystic minded guy who has no idea how they work and wants to believe that they are real. If you actually try to converse anything even slightly meaningful with one, it quickly becomes obvious that they have no idea what they are talking about.
Sure, they can produce some basic reddit memes and such, or plausibly sounding word salads that resemble what some write, but they are a long way away from actually being convincing chat partners. They do slightly better at imitating copypasta news reporting.
GPT-3 can bullshit many things convingly enough and at scale to make people reshare and vouch for it. Also, other chatbots can leave comments and reshare masquerading as real opinionated supporters in a “mob” and likewise barely answer you.
Remember… it’s not that you the one person who talks to them has to be convinced. It’s that your existing forums and chats and platforms can be gradually infiltrated with bots that convince OTHERS the vast majority of whom never did a turing test.
Think of how your opinion was formed on many political things, it likely happened non-interactively by just consuming articles and comments, and seemingly “defeated” arguments of OTHERS who could have all been bots.
I've seen how gnarly instagram bot comments already are, and wondered how many people are bots vs. people on there. It's not economical to turing test them all, so I've had to assume that they're all propaganda... and even then I've noticed the influence on deciding what's social.
How would you measure the success of this, if you were making them?
Any other thoughts on what the phases are?
Only humans are currently allowed to own capital and that's not going to change in the near future.
>judicial systems with parallel construction and deepfake “evidence” (Fake or real) pictures and video are in a lot of circumstances not admissible evidence in court, with deepfakes the set of inadmissible evidence will simply expand.
There is very real and very big concern about fake news and job loss and (in a slightly longer term) GAI going rogue but that's about it.
Sony regrouped, retrained its AI, and set up a rematch in October. This time GT Sophy won with ease. What made the difference? It’s true that Sony came back with a larger neural network, giving its program more capabilities to draw from on the fly. But ultimately, the difference came down to giving GT Sophy something that Peter Wurman, head of Sony AI America, calls “etiquette”: the ability to balance its aggression and timidity, picking the most appropriate behavior for the situation at hand.
It almost feels offensive, how it wasted my time without giving me equally valuable information in return.
There's certainly an audience out there that loves long-winded articles that meander around their topic so that you get the same amount of reading it would take for an in-depth look at a subject but still only come away with a handful of bullet points worth of useful information.
Having played GT since it's inception, and been involved in motorsports and used GT to learn a few tracks before GT Academy became a thing: on the track you have to have a certain degree of respect for your fellow riders/drivers, these are extremely dangerous machines capable of lots of harm from carelessness or even a simple mistake so being erratic makes you an extreme liability on track; in a game it's less so, and shunts without consequences don't require you to put forethought into an overtake on a banked/cambered turn or late breaking on an uneven/worn part of the track, or taking a certain line to prepare for another move 3 turns ahead etc... and I'm guess this AI has learned from this just like you would if you were about to make contact or potentially wreck on track which REALLY hurts.
The thing is, the crash may not even be the most damaging thing that happens because fights break out after a crash as well since so much adrenaline is pumping through your body and motorsports tends to attract lots of risk taking hot-heads--their are exceptions. So even if you think you are above learning the why, you will soon learn why it's practiced one way or another.
The etiquette component of on-track behavior is almost never present in online play because of that absence and it's why I've always had certain reservations with just assuming a person with sim only experience can be allowed on track with out some track experience before hand--sims are extremely useful and amazing in their feedback but it's not a full replacement. The Nismo GT Academy was cool because it combined both and made a few decent GT drivers in the process.
I think it's the old adage that 'slow is smooth, smooth is fast and speed is the economy of motion' which has allowed many smoother drivers/riders who use this to excel in their field and dominate: Alain Prost, Jorge Lorenzo come to mind.
iRacing has an MMR value called iRating which is used to match you against drivers of a similar skill level (like every online game) but it also has the unique concept called 'safety rating'. Safety rating is based on your 'incidents per corner' over some recent history of corners (in the 1000 to 10000 range). Safety rating controls which official racing series you're allowed to enter - higher level stuff like F1 cars requires a better safety rating than the entry level Mazda MX-5.
Having the need to maintain (or improve) this safety rating gives the 'risk' incentive to drivers which most online games are lacking. In real racing a desire not to put your car into a wall at high speed controls how dangerously people drive, but that risk doesn't exist online and so in a lot of games (e.g. the F1 games) people attempt stupid dive-bomb overtakes, take out their competitors and make it less fun to play. Safety rating puts that incentive back. It's not perfect of course, but it does help iRacing feel more 'real' and 'fair'.