Leela Zero and gradual rise of AI

[CoinCube]

A group of volunteer developers recently decided to decentralize the techniques used by the DeepMind Google team and create a chess program that teaches itself chess. As they were volunteers they lacked the enormous computational resources of Google. To solve this problem they decentralized the work it asking the chess community to contribute a GPU or a CPU towards helping Leela learn chess.

The computational power required to do this is immense. The process uses Markov chain general reinforcement learning algorithms. Essentially the program plays itself millions of times and remembers favorable positions and tactics gradually improving over time. There is no human involvement in the learning it teaches itself from the basic rules of the game via trial and error.

As more games are played the algorithm gradually learns the game. Millions of games must be played to master the game with this technique making the computational power needed to immense. The load is more manageable, however, when spread across hundreds of volunteers.



The Graph above shows the playing strength of the Leela algorithm over time. Was is notable is the gradual but continuing improvement. The Elo rating in the table above represents the improvement in play over random moves. Leela is not as strong yet the very best traditional chess programs but it is already playing at the GrandMaster level.

How good will it get with time? No one really knows but it appears to be continuing to improve by the day. If you want to see if you can outplay a decentralized self learning algorithm you can challenge it to a game at the link below. Word of caution I am not a bad chess player and it utterly destroyed me on the hard difficulty.

http://play.lczero.org/

Industrial robot sales are growing at a rate of 14% a year etting the stage for 3.1 million industrial robots in operation globally by 2020. ARK Investment Management, a leading researcher in this market, says that industrial robot costs are expected to drop a solid 65% between 2015 and 2025. Impressively, the cost per robot will plunge from $31,000 to $11,000 over that decade of time.

Why Industrial Robot Sales are Sky High
https://www.zerohedge.com/news/2018-05-29/why-industrial-robot-sales-are-sky-high



Chess is likely just the beginning. It's simple rules and clearly defined outcome make for an ideal cradle for such learning algorithms. There is no imagining how far this type of technology will progress with time. As the cost for labor drops like a rock due to an increasingly automated workforce true competition in the economy will for better or worse likely shift towards the computational race to solve new and complex challenges with every higher quality learning algorithms. These algorithms once fully developed will likely far surpass human efforts in the fields they are directed towards.  

As the saying goes we live in interesting times. One wonders what kind of money will be valued in a future economy that is increasingly dominated by the ability to efficiently compute solutions to complex problems via computational work. Something to think about before selling your bitcoins.

If you want to help Leela Zero learn to play chess it is easy to dedicate a GPU or a CPU towards this. Instructions to do so can be found here:

https://github.com/glinscott/leela-chess/wiki/Getting-Started

[OROBTC]

...

Very interesting, CC.  Thanks for sharing.

Google's AI bot recently beat the world champion in the Oriental strategy game of Go, which "they said" was a much harder game for a computer than chess.

I wonder how much AI and algorithms would affect a Knowledge Age, as predicted by one of our old friends here.

*   *   *

I saw robots in Korea at a couple of their ball bearing plants.  Those factories were very advanced.

And robots were everywhere at the BMW auto assembly plant we toured a year and a half ago.

The USA is apparently way behind in the race to build high-quality robots.  It's an important industry, our country will be missing out if we do not engage in this race.

[Steamtyme]

So I had the opportunity a few years ago to work with an automated cell. Very rudimentary robots you could say, every position was plotted and programmed; this along the support infrastructure for placing and centering relied entirely on sensors. Every action required these sensors to either open or close that part of the circuit for operation. If I remember correctly these robots cost about 250 000 each, and had no thought process just a program to follow and HMI for when things went sideways.

The introduction of AI to a simple system like this would have easily eliminated 3 of the 4 jobs required to run that cell. Not necessarily a bad thing, wieghing the cost benefit of these inovations will fall on society to deal with and find a balance.

Where an AI system does tend to make the hair on my neck stand, is the idea of unleashing it into society. I think of examples such Tay, Microsofts twitter chat bot. The "learning" led it down a path of racism and misogyny.  Using this example, I feel like despite these trial runs, it may be another advancement in technology we push further without fully understanding the possible ramifications of our actions.

Like I said I can see application specific benefits to AI systems, but I also think there is a very fine line that we should not cross I'm just not sure where that is yet.