Sony’s racing AI destroyed its human competitors by being nice (and fast) • Hacker News Board

Re: Sony’s racing AI destroyed its human competitors by being nice (and fast)

#31

Post by trention » Thu, Jul 21, 2022, 6:31 PM UTC

EGreg wrote: post #13

“I don’t think we’ve learned general principles yet about how to deal with human norms that you have to respect,” 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). Ent…

> and all their content and capital is vanishingly tiny

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.

Re: Sony’s racing AI destroyed its human competitors by being nice (and fast)

#32

Post by lowbloodsugar » Thu, Jul 21, 2022, 6:43 PM UTC

RajT88 wrote: post #16

mrandish wrote: post #11

I think this is material info which needs to be noted: > 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.…

It caught my eye as well. 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…

Fundamentally, despite the claims, Gran Turismo isn't a very good driving physics simulator.

Re: Sony’s racing AI destroyed its human competitors by being nice (and fast)

#33

Post by lowbloodsugar » Thu, Jul 21, 2022, 6:50 PM UTC

mrandish wrote: post #11

I think this is material info which needs to be noted: > 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.…

Kind of a straw man though. Any real-world simulator is going to need perception and then decision systems. They are mostly two separate problems. Certainly, the decision system will need to be adapted to deal with momentarily incorrect or absent data, but for the most part will remain unchanged. Perception (telemetry really) on a closed-circuit race track is far more achievable than on city streets. So hats off to Sony for working on the decision system aspect. Would love to see it paired with perception system and put on the track.

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.

Re: Sony’s racing AI destroyed its human competitors by being nice (and fast)

#34

Post by t_mann » Thu, Jul 21, 2022, 6:54 PM UTC

Reminds me of a question I've always had about game AIs: suppose you have a really good one (like this one, or in chess eg), how do you make the easier difficulty levels? Just throwing random noise into the decisions would probably lead to non-enjoyable outcomes. Is there a structured approach to it?

Re: Sony’s racing AI destroyed its human competitors by being nice (and fast)

#35

Post by t_mann » Thu, Jul 21, 2022, 6:57 PM UTC

justinlloyd wrote: post #24

"This AI is amazing and destroys humans!" 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 A…

How do you do the turning down? Just curious, doesn't seem like there's an obvious way that's also guaranteed to be fun.

Re: Sony’s racing AI destroyed its human competitors by being nice (and fast)

#36

Post by batmanthehorse » Thu, Jul 21, 2022, 6:57 PM UTC

raz32dust wrote: post #20

andsoitis wrote: post #15

Earlier quoted context omitted.

That's why humans should insist on competing with AIs in our world not theirs (simulation). 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).

It's hard to draw the boundary. We could say that AI should only be given access to exactly the same information that humans have access to. But AI will have magnitudes more memory and compute resources. So for example, an AI could only have the 2D view but it could store, access and process the 2D view in ways a human cannot.

Yes, but more memory and processing power might mean the AI is actually better at performing the same task. Cleaner, pre-processed input data is cheating in a different way.

Re: Sony’s racing AI destroyed its human competitors by being nice (and fast)

#37

Post by ascagnel_ » Thu, Jul 21, 2022, 7:13 PM UTC

RajT88 wrote: post #16

mrandish wrote: post #11

I think this is material info which needs to be noted: > 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.…

It caught my eye as well. 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…

> 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.

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]

0: https://roborace.com

1: https://en.wikipedia.org/wiki/Roborace

2: https://www.autocar.co.uk/car-news/business-tech%2C-developm...

Re: Sony’s racing AI destroyed its human competitors by being nice (and fast)

#38

Post by cdkmoose » Thu, Jul 21, 2022, 7:43 PM UTC

RajT88 wrote: post #16

mrandish wrote: post #11

I think this is material info which needs to be noted: > 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.…

It caught my eye as well. 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…

F1 drivers do the same thing through a chicane or wide on a curve to maintain speed, leaving the smoothest road surface for that advantage (not the skidding).

Re: Sony’s racing AI destroyed its human competitors by being nice (and fast)

#39

Post by andsoitis » Thu, Jul 21, 2022, 7:53 PM UTC

raz32dust wrote: post #20

andsoitis wrote: post #15

Earlier quoted context omitted.

That's why humans should insist on competing with AIs in our world not theirs (simulation). 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).

It's hard to draw the boundary. We could say that AI should only be given access to exactly the same information that humans have access to. But AI will have magnitudes more memory and compute resources. So for example, an AI could only have the 2D view but it could store, access and process the 2D view in ways a human cannot.

> We could say that AI should only be given access to exactly the same information that humans have access to.

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.

Re: Sony’s racing AI destroyed its human competitors by being nice (and fast)

#40

Post by autoexec » Thu, Jul 21, 2022, 7:56 PM UTC

Double_a_92 wrote: post #4

isaacfrond wrote: post #2

tl;dr 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 aggre…

Thank you. I read the full article but they are always just soo unnecessarily long. All the interesting bits could easily be condensed into 5 sentences. It almost feels offensive, how it wasted my time without giving me equally valuable information in return.

I feel the same way about radiolab pocasts. Interesting topics, unnecessarily drawn out and filled with obnoxious amounts of repetition and pointless music/audio interludes to the point where I'd often just rage quit, but so many people love it and praise the production.

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.