Become a fan of Slashdot on Facebook


Forgot your password?

Comment Re:Idiots in charge! (Score 1) 184

I have a lot of sympathy with what you write. However, most large organizations these days do take disaster recover and business continuity somewhat seriously, and have budgets to ensure issues like this are supposed to be addressed.

As others have opined, while some kind of power outage might have precipitated the problems, surely there must have been a cascading series of events that prevented the system from coming back online quickly. Of course, a power outage should not bring down critical servers in the first place: UPS combined with backup generators should keep critical infrastructure online. However, I have seen cases where backup generators were not regularly tested, and failed when really needed. Once servers crash, especially if it is at peak transaction times, you may be faced with time consuming database restarts, with lengthy logs that need to be processed. Even assuming there is a redundant secondary data center with database replication, it only takes a software glitch to prevent that from saving you (especially as such recovery scenarios are rarely sufficiently tested). I doubt we will ever get the complete story, but multiple infrastructure, software and human deficiencies probably combined to generate this fiasco.

Comment Re:So, you know how the rewritten version works? (Score 1) 186

It was over 1000 CPUs and over 200 GPUs. That's rather beefy, mate.

Beefy, yes, but nothing extraordinary, and as I have mentioned since reduced by about 90%.

there was definitely an inflection point in Go progress around 2005

Between about 2005 and 2011, significant progress was made. The top programs got to the point where they were competitive with professional players when receiving a (still very big) 4 or 5 stone handicap. (Note, however, that players could exploit weaknesses in computer Go programs once they were studied which makes the achievement a bit less impressive than it seems at first glance.) Between 2011 and 2015, no further progress was made, as the techniques then being applied had reached their limits. The expectation by most in 2015 was that there would be a slow progression of beating average professional players at 3 stone handicap, 2 stone, 1 stone, and eventually level, followed by the ultimate achievement of beating the top professional at even. See Sensei's library page on computer Go which has been updated over time as progress was made, and is not a revisionist account. AlphaGo advanced the state of the art to an extent that shocked the Go community and 99% of the AI community. The leap was not just from beating a middle rank professional with a 4 or 5 stone handicap to beating the top professionals level, it was doing so while having no apparent weaknesses the human player could exploit.

Comment Re:So, you know how the rewritten version works? (Score 1) 186

As I tried to explain, the amount of hardware you throw at the problem of paying Go well does not really help. Even the older system that beat Lee Sidol was not running on a humongous supercomputer. Pure computer power is of very limited benefit. Indeed, while dramatically improving the capabilities of AlphaGo over the last year, DeepMind has succeeded in reducing the computing requirements of the system by 90%. It now runs on a single TensorFlow machine (albeit, this is hardware with an architecture tailored to the needs of AI).

As recently as 2010, AI textbooks were typically writing that the field was 20-30 years away from creating a machine that could beat professional Go players. By late 2015, some AI researchers were more optimistic, believing the milestone of beating a top professional might be reached within a decade. Whatever some might have claimed later, very few as late as 2015 were expecting a solution within 5 years. Meanwhile, Go players truly believed a solution was between decades and forever away.

Comment Re:So, you know how the rewritten version works? (Score 1) 186

It is instructive (and important in understanding the significance of AlphaGo in overall AI research) to know the important differences between the nature of chess and go that leads to a totally different challenge in playing it well. The most important differences are:

  1. It is very difficult (still impossible and will probably remain so) to hand craft a set of rules to evaluate whether a particular board position is good at go. In chess, just counting up the value of the pieces on the board (counting queen 9, Rook 5 etc.) gives a good rough estimate, that can be refined by recognition of other factors such as passed pawns, king safety and inactive versus active pieces. At go, each stone has equal value (simplistically speaking) and a small change to the position of the single stone can often make a total difference to who is winning, only via effects that occur many moves later.
  2. The branching factor at go is far greater than in chess, even without the challenge of knowing whether a position is good. This means that even examining all possible positions a few moves ahead becomes infeasible. At chess (especially given the previously cited relative ease of writing an evaluation function allowing pruning of obviously hopeless lines) very accurate selection of the most likely best line is possible by Monte Carlo techniques.
  3. Chess programs can have an opening book that records known good early moves (the same in true at go to a lesser extent). However, after that a major difference happens. At chess the position is simplified as pieces are captured. Indeed, once down to about 7 or 8 major pieces, a chess program can use an endgame database to play perfectly without the need for any further calculation. Go, in contrast, is an additive game. The position continues to increase in complexity typically for at least the first 80 moves by each player.

A chess grandmaster can, indeed, explain why a particular move is good, usually by demonstration. Even where the benefits cannot be directly shown, there is established theory known to be sound, to justify it. Actually, a grandmaster cannot improve his knowledge of chess by examining the moves of a chess program that is only superior because of greater calculation and storage capacity

Top go professionals mostly cannot explain in a clearly irrefutable way why certain moves are good. Often, they just need to say they instinctively feel a move is right. There is a 3000 year-old repository of theory (which has been upended twice before in history, first via innovations about 300 years ago, and then again around 70 years ago) but this received wisdom is not known to be totally correct. In fact, the evidence from AlphaGo's play is that much of the existing theory is wrong. The top go professionals find this extremely exciting, as they begin to understand the logic behind AlphaGo's new moves, and the play of these professionals is already changing to incorporate the new knowledge it is allowing them to learn.

There were reasons why AI and go experts believed it would be 20 more years before a go program could best the top professionals. The AI techniques that made it possible are immensely exciting because they are definitely applicable in the area of artificial general intelligence. They are mostly not go specific.

Comment Re:Not AI (Score 1) 186

You cannot get from the intelligence of a rock to Einstein in one step. The long term objective of DeepMind is to solve intelligence. Their work on AlphaGo, as well as their earlier work on unsupervised learning and optimal play of Nintendo games, are steps along this road, important steps. Games are an excellent medium for examining approaches to AI. It is as well if fundamental approaches are tested there before applying them to cancer diagnosis, drug research, electric network optimization and other areas where AIs designed by DeepMind are now being applied. Are we at the point yet where any AI can replicate all the intellectual capabilities of an intelligent human? Absolutely not. That remains decades away, but recent progress has actually surpassed the most optimistic expectations of experts 3 or 4 years ago.

Comment So, you know how the rewritten version works? (Score 1) 186

All I have been able to glean so far is that the rewritten version uses around 10% of the computing power (both to train its neural networks, and during actual play) to achieve much improved play compared with the original AlphaGo used to beat Lee Sedol. Thus far, although promised, the architecture behind the rewritten version is unpublished. Later this week, some insight is going to be provided.

The old version was based on a combination of techniques (primarily multiple neural networks, combined with Monte Carlo techniques). The interesting thing about the way it operated was that it could tell you which move was likely best, but could not explain why. The same is actually true of human Go players. While locally best moves can be identified, the human selects the globally best move based to a large extent on feel. The game is too complex (both for humans and AIs) to use calculation on a board wide basis. Both the old and new AlphaGo systems appear to demonstrate characteristics we would refer to as "intuition" and "creativity" if seen in humans. How similar is it to human instinct and creativity? We really do not know.

I am extremely interested in learning how the rewritten system works. I think the twinning matches (between two teams each with one human expert and AlphaGo collaborating with each other) will also be extremely significant. The short to medium term promise of AI involves humans and AIs working together. As the AIs become increasingly complex, and the manner in which it comes up with recommendations ever harder to comprehend, this is a critical challenge to be addressed.

Comment Not quite dead yet (Score 4, Interesting) 72

While FastMail is based on Cyrus IMAP, and is providing resources for its development and documentation, I think it is to early to declare Cyrus completely finished. In terms of collaboration features, the addition of CardDAV and CalDAV support a few years ago helped somewhat. Lack of its own file sharing tools is a serious limitation, but FastMail has managed a degree of integration with Dropbox.

Hold off on a variation of the dead parrot sketch for the time being!!

Comment Re: Lol no (Score 1) 451

I have heard this argument before, and it held true for the first half of the 20th century. Over the last 30 to 40 years, the poor have not benefited from the advances that make the things you mention possible. See, for instance,

Comment Re:Lol no (Score 4, Insightful) 451

It is unsurprising that there is resistance to this idea.The implications (more on that below) are horrific.The fact, though, is that robots and AIs are becoming rapidly more capable, and denial is not going to prevent organizations from selecting the most cost effective way to get jobs done. Even if the robot/AI solution has some limitations, the profit motive will win out (as anyone who has used call centers staffed by people who cannot communicate effectively in your language should recognize).

What are the implications? The most obvious is mass unemployment/under employment. This is going to create a huge disadvantaged class in rich countries. Proposals for a national basic income are well meaning, but unrealistic. It might happen in a very limited number of smaller countries, like Finland, but the elites in most countries who decide such matters will never willingly allow some of their wealth to be given to "non productive" members of society.

The BBC ran an interesting opinion piece recently ( that predicted a breakdown of Western civilization if gross and increasing levels of inequality continue to occur. I think those predictions ring true. Further, the piece does not even consider the problems introduced by huge segments of the population becoming completely surplus to the elite's needs.

There will be valuable jobs those displaced by robots and AIs could do, but they will be of no economic benefit to the elites to would have to put up the money to finance them.

Ever since I was a child, I have been reading about how automation would create more leisure time, and the challenge being how that leisure time will be used. The reality of the last 40 years is that those with jobs work harder than ever for the same or less money in real terms. Total wealth has increased, but (the predictions of trickle down economics notwithstanding) virtually all the increase has gone to the already wealthy.

Comment Re:Its become too political (Score 3, Insightful) 167

It has become politicized, because strong business interests are resisting acceptance of scientific consensus. This is nothing unusual. Business will always dispute facts that can lead to regulation costing them money. They will even claim that their cynical twisting of the facts is mandatory, because they have a fiduciary duty to maximize shareholder value.

Climate change is complicated, and no serious scientist will claim they know exactly where it is leading. What is universal among climate scientists is that human induced climate change has and is occurring. There are tentative conclusions about some of its effects, and warnings that failing to act to reduce human induced climate change risks truly catastrophic consequences. If the worst happens, it may not be for 100 years, but the earlier action is taken, the lower the cost of remediation is likely to be. The commonly held view is that it is irresponsible, and totally unfair to future generations, to dodge taking prudent steps because it will cost some businesses money.

Comment Re:I don't see it (Score 1) 301

Perhaps, whether one considers "one" to be very "formal and impersonal" depends on which of the many dialects of English one grew up using. As you can tell from the previous sentence, I consider its use perfectly natural in some situations. Yes, I know many would replace "one" with "you", but that in a literal sense changes the meaning.

Comment Interesting proof of concept (Score 4, Insightful) 88

If the construction costs cited are true, even given the small size of the demonstration house, this seems a very viable approach. One would imagine most of the human labor could ultimately be replaced by robots, and (although 24 hour completion is impressive) taking a whole week would not alter the economics significantly. (I guess that might not be true if the capital cost of the printer makes the investment uneconomic at, say, 20 to 30 houses a year, but I doubt that is the case.)

Comment Re:What is Facebook thinking? (Score 5, Informative) 122

It is even worse than the summary suggested. The BBC did not originally send the evidence to FB. They did so when FB asked them to ahead of an interview arranged with FB's director of policy Simon Milner. Reporting them to the police for providing what they were requested to beggars belief.

Slashdot Top Deals

Between infinite and short there is a big difference. -- G.H. Gonnet