AlphaGo destorys Ke Jie with only one GPU
| ivory federal associate indirect expression | 05/28/17 | | marvelous turquoise athletic conference range | 05/28/17 | | smoky parlour pistol | 05/28/17 | | ivory federal associate indirect expression | 05/28/17 | | razzmatazz elastic band | 05/28/17 | | spruce free-loading abode | 05/28/17 | | razzmatazz elastic band | 05/28/17 | | ivory federal associate indirect expression | 05/28/17 | | razzmatazz elastic band | 05/28/17 | | ivory federal associate indirect expression | 05/28/17 | | smoky parlour pistol | 05/28/17 | | mischievous laughsome brunch death wish | 05/28/17 | | ivory federal associate indirect expression | 05/28/17 | | vermilion lettuce area | 05/28/17 | | razzmatazz elastic band | 05/28/17 | | vermilion lettuce area | 05/28/17 | | razzmatazz elastic band | 05/28/17 | | vermilion lettuce area | 05/29/17 | | ivory federal associate indirect expression | 05/29/17 | | Odious black den | 05/28/17 | | razzmatazz elastic band | 05/28/17 | | Odious black den | 05/28/17 | | Electric library | 05/28/17 | | smoky parlour pistol | 05/28/17 | | ivory federal associate indirect expression | 10/18/17 | | alcoholic legend gas station | 10/23/17 | | ivory federal associate indirect expression | 06/25/18 |
Poast new message in this thread
Date: May 28th, 2017 4:20 PM Author: ivory federal associate indirect expression
http://www.npr.org/sections/thetwo-way/2017/05/25/530016562/google-a-i-clinches-series-against-humanitys-last-best-hope-to-win-at-go
the newest version runs on one GPU and is undefeated in the 60+ games it has played against human players. it also beat a team of five expert Go players. it's quite possible an unaided human will never be able to beat it now.
the fact we are now seeing strong superhuman performance in games that are almost entirely based in pattern recognition should give pause to AI skeptics.
(http://www.autoadmit.com/thread.php?thread_id=3629037&forum_id=2#33415847) |
Date: May 28th, 2017 5:09 PM Author: razzmatazz elastic band
ke jie performed better than the team of 5
also fuck google for trying to create strong AI. world governments need to stop these ideological arrogant nerds before they terminator the rest of us.
also don't get how this is news tbh after lee sedol... ok, machine learning works
(http://www.autoadmit.com/thread.php?thread_id=3629037&forum_id=2#33416066) |
|
Date: May 28th, 2017 5:58 PM Author: ivory federal associate indirect expression
government regulations on the tech industry or university research groups wouldn't be helpful. the research would simply move to online anonymous research communities, which is a scarier scenario.
you would need strict controls on computer hardware. this is very unlikely to happen. good luck telling gamers they can't get a 7nm Nvidia GPU because it could potentially be used for AI research.
the best we can hope for now is that the tech companies don't fuck things up.
(http://www.autoadmit.com/thread.php?thread_id=3629037&forum_id=2#33416254) |
|
Date: May 28th, 2017 6:04 PM Author: razzmatazz elastic band
i dont know if this is right.
its not like "strong ai" is just a new fancy line of code. we're talking about something far more complex and integrated with other complex technologies.
major tech giants have the technoligcal tools to possibly bring together the various aspects. underground geeks do not even if they can create powerful machine learning programs
(http://www.autoadmit.com/thread.php?thread_id=3629037&forum_id=2#33416279) |
|
Date: May 28th, 2017 8:10 PM Author: ivory federal associate indirect expression
i don't think we are talking about a massively complex piece of software. the progress in deep learning has been largely through simple, related architectures and lots of computing power/data. this is consistent with our current understanding of the brain. the neocortex is highly regular and appears to be running one general learning algorithm. they can see this through rewiring experiments, where they reroute sensory data to different regions of the brain. eventually the brain reorganizes itself to perform sensory processing in the new area. this suggests there is probably a compact program that could replicate many complex functions in the brain.
Google or whoever probably won't be engineering a bunch of different complex modules and sticking them together to get general intelligence. i think it will be surprisingly simple learning algorithm that they parameterized with a ton of computing power and experimentation in different environments. this is certainly not something a person could do in their basement currently. things get a bit different by the early to mid 2020s and petaflop desktops become a reality.
i am not happy about this situation and i can imagine things going very wrong, but it doesn't seem like much can be done about it at this point.
(http://www.autoadmit.com/thread.php?thread_id=3629037&forum_id=2#33416863) |
|
Date: May 28th, 2017 5:37 PM Author: ivory federal associate indirect expression
The new version is significantly stronger than the one that played Sedol and runs on much less powerful hardware. The one win Sedol pulled off left open the possibility that there was a fatal exploitable weakness. It is clear now that computer Go is basically a solved problem.
how do you propose to stop it? Short of blowing up the semiconductor fabs, there is no realistic way to do it. I would rather have Google build it than an underground group of computer scientists.
(http://www.autoadmit.com/thread.php?thread_id=3629037&forum_id=2#33416175)
|
|
Date: May 28th, 2017 5:58 PM Author: razzmatazz elastic band
dunno if anyone else here is a go player, but.
sedols win is intriguing in part because it was definitely not his best game... he didnt win on big picture strategy so to speak. he won because alphago misunderstood a ko fight. like clearly couldve won but literally couldn't figure out what sedol was doing until too late (whereas a mediocre amateur player could see it).
computers have always been oddly retarded at ko fights, alphago seems ok at them but its one giant misstep was related to ko which makes you wonder why. idk tho. obviously it wasn't some ultimate fatal flaw it couldn't learn its way around or others wouldve exploited it.
(http://www.autoadmit.com/thread.php?thread_id=3629037&forum_id=2#33416253) |
|
Date: May 29th, 2017 12:07 AM Author: vermilion lettuce area
very very bad
only learned a bit from my dad but im really terrible
he spends all his free time playing online GO and complains about chinese all the time lmao
(http://www.autoadmit.com/thread.php?thread_id=3629037&forum_id=2#33418369) |
Date: October 18th, 2017 2:07 PM Author: ivory federal associate indirect expression
this is pretty neat:
https://deepmind.com/blog/alphago-zero-learning-scratch/
they eliminated the supervised learning step and trained it purely on self play. looking at the ELO statistic, there's approximately a 1 in 10000 probability the best human player could defeat it now, and it uses much less computational power than before.
Go has been effectively solved.
(http://www.autoadmit.com/thread.php?thread_id=3629037&forum_id=2#34471658)
|
Date: June 25th, 2018 11:11 PM Author: ivory federal associate indirect expression
Dota 5v5 is about to fall to computers:
https://blog.openai.com/openai-five/
This environment is significantly more challenging than Go. It has a larger action space and more distant rewards (which means it requires long term planning). The architecture isn't all that sophisicated and doesn't make use of many of the state of the art RL techniques, but it still works. Throw enough computing power behind a highly sub-optimal learning algorithm and you can still get superhuman performance.
Dumbs will argue these techniques don't extend to real world tasks where you don't have an environment simulator you can brute force. Generative models are becoming quite good, so that won't be the case for the long.
It is time to panic.
(http://www.autoadmit.com/thread.php?thread_id=3629037&forum_id=2#36309817)
|
|
|