This article is about opaque, proprietary algorithms that help some professions with decision-making (banks with loans, teacher rankings, university attributions, etc). As described in the book Weapons of Math Destruction, these algorithms give all the pretence of providing bias-free decisions but do the opposite. Depending on the context the algorithms may depend on hand-coded rules or on machine-learned ones, but the biases are in the code or the data and their annotations.
Waymo have an agreement to pursue self-driving cars with Lyft, the main Uber competitor. Uber have fired their principal self-driving car engineer. Meanwhile Uber is in disarray due to alleged toxic working place conditions.
The objective of this lawsuit is not for Waymo to win a settlement, they probably don't care so much about the money, it is to win time and mindshare by burying Uber in this corner of the market.
Personally I think self-driving cars are coming but the engineering challenges are still formidable, perhaps these fights are premature.
In 2013, date of the report, Deep Learning was not yet on the table except for a small number of researchers. This is the case now. DL algorithms were devised in the 1990s, with improvement in better choices of activation functions (ReLU), improvements in back-propagation algorithms with stochastic gradient descent. Also crude but effective network simplification methods were found useful (dropout). The big game changer was progress in GPUs though.
Interesting but not novel or new. For example typing patterns have been used before, this is an established biometric measurement called typing or keystroke dynamics . There are even companies that sell typing metrics as an authentication factor.
Yes but DNA does not "encode" a human. It encode a process for generating a body, which in the human case takes about 20 years. The encoding is not one-to-one either, see twins.
Nice try quoting bloomberg, like they are a perfectly neutral observer there. *Just* getting rid of the overhead of having multiple inefficient systems competing with each other in an anti-competitive market would be a bonus. Then there is the matter of the single payer being able to negotiate better medicine practices and prices. This is what nearly every country in this world does. The champion is Japan. Notice they have a lot of centenarians there.
Also, why should the US be like other countries when most people in the US want a more US-like system?
The likelihood of random noise creating a chirp consistent with a BH merger is very close to nil. The fact that two separate stations recorded the same chirp at the same time make this explanation very unlikely indeed.
Great research, Microsoft! congrats to Haiyan Zhang. That said it doesn't look like the real Emma in the video has classic Parkinson, which is a lot more than hand tremor (also slurred speech, posture changes, impaired balance, slowed movements, and more). I doubt Microsoft's Emma watch will help all of these. Nonetheless, new thinking, progress!
1- Automated translation is not yet beaten. A lot of progress has been made but not there yet. It's not a matter of building a larger computer... 2- Robots and hooliganism. Once we have a self driving car out in the streets, it will be put on its back like a tortoise so it cannot escape, get stolen and sold for parts to a competitor. Empty self-driving cars will get pushed off the road by human drivers just for fun. How do we solve that problem? 3- The energy problem. Alright, let's see your superintelligent computer solve that one. The new super-AI will need vast amounts of energy too, so it has a stake in it. Where will it get it from? We basically have an existing menu of how to produce energy from solar to nuclear but none is foreseen to solve all of our needs in this current century while simultaneously not destroying the planet, the environment, many other humans and so on. SuperAI will not be able to devise new physics by reading or even understanding our current physics textbooks. 4- Exponential grow. Right. As you know, the myth of exponential growth in silico is pretty much at an end. My 2009 computer is not significantly slower than the 2017 ones. It has 6 cores and many gigabytes of RAM and many terabytes of disk, i.e. not fewer in any measure than newer computers. I plan to use it until It dies of old age, probably in another two years or so or perhaps even longer. So how will SuperAI pull all this amazing exponential growth from? New physics perhaps? 5- Ok we have a super intelligent nasty AI. How does it prevent us from pulling the plug? See question 3.
Basically most of this AI stuff is science fiction at present.
A few things: 1- humans are actually improving, if only relatively slowly. The basic substrate is the same, sure, but the accumulated historical and scientific knowledge does add up. We do not need to solve old problems, only new ones. 2- "I don't know of any fundamental limit" does not mean there isn't one, for example a methodological limit. For me the current direction of mainstream AI research will not give us thinking computers anytime soon. Better statistical machine learning methods, yes, for sure. The dichotomy between supervised and unsupervised learning is not very productive, they do not solve the same problems at all.
Yes, of course reinforcement learning is intrinsically limited. The current state of the art is deep learning, but even this technique is still statistical machine learning. In other words, it works by devising a cost or objective function (in the case of go, winning the game, which is clearly defined), and optimises it by starting from examples. In the case of the game of go, the hundreds of thousands of games already annotated and played by humans. Reinforcement learning artificially creates many new games from old ones, and playing these, which helps finding a better local minimum in the cost function.
Now "strong" AI or whatever you want to call some version of human-like AI have to cope with situations where there is no clearly defined cost function, and very few if any example to start from. The complexity of these problems is way beyond that of Go, which is a very limited system. Note that these problems are every day ones that humans have to cope with every day, like how do I persuade my clients to pay me ? how do I explain to my work partners the value of my work ? they typically include other complex systems that are not rule-based or for which the rules may exist but are unknown. Example: we do not know enough of physics to accurately predict next week's weather. How do we improve the situation?
You will not be able to better predict next week's weather with reinforcement learning.
Concorde did not recoup its development costs due to the small number of operational aircrafts (only 20 were ever produced, 14 of which saw commercial use). However they were operationally profitable, meaning their usage generated profit over and above all the operational costs.
Frankly who votes in the US elections ? Somehow it seems the 400 richest guys in America have managed to command a massive Zombie army to vote for their reverse Robin-Hood program.
ZFS is harder to setup and manage than EXT4 but easier than LVM + md. If you wanted to do any kind of mirroring / software RAID it is pretty much the recommended way under Linux right now. The only thing that could have been a problem is deduplication. Stay away from that stuff.
No, not at all. ZFS is designed to work on bare disks with as little hardware and software between the devices and itself. No battery protected stuff; hardware RAID is a big no no. It was also started on ancient Sun hardware that is so very slow compared with modern hardware that it is not even funny.
I have a ZFS system that has been running for years on a 2010-era Pentium. It does require a lot of memory. Don't bother with deduplication (off by default) but do turn on compression.
I recommend using a RAIDZ6 on heaps of tablets then. Should be good up to two broken tablets by VDEV of 8-12 tablets. Do a vigorous scrub now and then, perhaps with a metal brush. If in doubt, convert to read-only by burying the tablets.
Slashdot Mirror
This article is about opaque, proprietary algorithms that help some professions with decision-making (banks with loans, teacher rankings, university attributions, etc). As described in the book Weapons of Math Destruction, these algorithms give all the pretence of providing bias-free decisions but do the opposite. Depending on the context the algorithms may depend on hand-coded rules or on machine-learned ones, but the biases are in the code or the data and their annotations.
It's the vibes, man. To developers VR is all wrong if Windows is underneath.
Waymo have an agreement to pursue self-driving cars with Lyft, the main Uber competitor. Uber have fired their principal self-driving car engineer. Meanwhile Uber is in disarray due to alleged toxic working place conditions.
The objective of this lawsuit is not for Waymo to win a settlement, they probably don't care so much about the money, it is to win time and mindshare by burying Uber in this corner of the market.
Personally I think self-driving cars are coming but the engineering challenges are still formidable, perhaps these fights are premature.
In 2013, date of the report, Deep Learning was not yet on the table except for a small number of researchers. This is the case now. DL algorithms were devised in the 1990s, with improvement in better choices of activation functions (ReLU), improvements in back-propagation algorithms with stochastic gradient descent. Also crude but effective network simplification methods were found useful (dropout). The big game changer was progress in GPUs though.
Interesting but not novel or new. For example typing patterns have been used before, this is an established biometric measurement called typing or keystroke dynamics . There are even companies that sell typing metrics as an authentication factor.
Yes but DNA does not "encode" a human. It encode a process for generating a body, which in the human case takes about 20 years. The encoding is not one-to-one either, see twins.
Nice try quoting bloomberg, like they are a perfectly neutral observer there. *Just* getting rid of the overhead of having multiple inefficient systems competing with each other in an anti-competitive market would be a bonus. Then there is the matter of the single payer being able to negotiate better medicine practices and prices. This is what nearly every country in this world does. The champion is Japan. Notice they have a lot of centenarians there.
Ignorance maybe ?
Nice FUD, sir.
Demonstrably socializing medicine is the better way, evidence is here.
Best.
The likelihood of random noise creating a chirp consistent with a BH merger is very close to nil. The fact that two separate stations recorded the same chirp at the same time make this explanation very unlikely indeed.
Somehow the Reps will credit this outcome to Trump's open, business-like, forward looking and win-win policies towards China.
The answer is No.
Great research, Microsoft! congrats to Haiyan Zhang. That said it doesn't look like the real Emma in the video has classic Parkinson, which is a lot more than hand tremor (also slurred speech, posture changes, impaired balance, slowed movements, and more). I doubt Microsoft's Emma watch will help all of these. Nonetheless, new thinking, progress!
A few examples in increasing order of complexity
1- Automated translation is not yet beaten. A lot of progress has been made but not there yet. It's not a matter of building a larger computer...
2- Robots and hooliganism. Once we have a self driving car out in the streets, it will be put on its back like a tortoise so it cannot escape, get stolen and sold for parts to a competitor. Empty self-driving cars will get pushed off the road by human drivers just for fun. How do we solve that problem?
3- The energy problem. Alright, let's see your superintelligent computer solve that one. The new super-AI will need vast amounts of energy too, so it has a stake in it. Where will it get it from? We basically have an existing menu of how to produce energy from solar to nuclear but none is foreseen to solve all of our needs in this current century while simultaneously not destroying the planet, the environment, many other humans and so on. SuperAI will not be able to devise new physics by reading or even understanding our current physics textbooks.
4- Exponential grow. Right. As you know, the myth of exponential growth in silico is pretty much at an end. My 2009 computer is not significantly slower than the 2017 ones. It has 6 cores and many gigabytes of RAM and many terabytes of disk, i.e. not fewer in any measure than newer computers. I plan to use it until It dies of old age, probably in another two years or so or perhaps even longer. So how will SuperAI pull all this amazing exponential growth from? New physics perhaps?
5- Ok we have a super intelligent nasty AI. How does it prevent us from pulling the plug? See question 3.
Basically most of this AI stuff is science fiction at present.
A few things:
1- humans are actually improving, if only relatively slowly. The basic substrate is the same, sure, but the accumulated historical and scientific knowledge does add up. We do not need to solve old problems, only new ones.
2- "I don't know of any fundamental limit" does not mean there isn't one, for example a methodological limit. For me the current direction of mainstream AI research will not give us thinking computers anytime soon. Better statistical machine learning methods, yes, for sure. The dichotomy between supervised and unsupervised learning is not very productive, they do not solve the same problems at all.
Basically correct.
Assuming that you had a correct model for a neuron, and the correct wiring and structure, where do you get the data to boostrap the simulation ?
Yes, of course reinforcement learning is intrinsically limited. The current state of the art is deep learning, but even this technique is still statistical machine learning. In other words, it works by devising a cost or objective function (in the case of go, winning the game, which is clearly defined), and optimises it by starting from examples. In the case of the game of go, the hundreds of thousands of games already annotated and played by humans. Reinforcement learning artificially creates many new games from old ones, and playing these, which helps finding a better local minimum in the cost function.
Now "strong" AI or whatever you want to call some version of human-like AI have to cope with situations where there is no clearly defined cost function, and very few if any example to start from. The complexity of these problems is way beyond that of Go, which is a very limited system. Note that these problems are every day ones that humans have to cope with every day, like how do I persuade my clients to pay me ? how do I explain to my work partners the value of my work ? they typically include other complex systems that are not rule-based or for which the rules may exist but are unknown. Example: we do not know enough of physics to accurately predict next week's weather. How do we improve the situation?
You will not be able to better predict next week's weather with reinforcement learning.
This is a fun thing, but the voices still sound very very artificial.
Actually it is a perfectly normal speed train. Arlanda is 45km away from both Stockholm and Uppsala. It runs at about 150km/h (100m/h)
Concorde did not recoup its development costs due to the small number of operational aircrafts (only 20 were ever produced, 14 of which saw commercial use). However they were operationally profitable, meaning their usage generated profit over and above all the operational costs.
Frankly who votes in the US elections ? Somehow it seems the 400 richest guys in America have managed to command a massive Zombie army to vote for their reverse Robin-Hood program.
ZFS is harder to setup and manage than EXT4 but easier than LVM + md. If you wanted to do any kind of mirroring / software RAID it is pretty much the recommended way under Linux right now. The only thing that could have been a problem is deduplication. Stay away from that stuff.
No, not at all. ZFS is designed to work on bare disks with as little hardware and software between the devices and itself. No battery protected stuff; hardware RAID is a big no no. It was also started on ancient Sun hardware that is so very slow compared with modern hardware that it is not even funny.
I have a ZFS system that has been running for years on a 2010-era Pentium. It does require a lot of memory. Don't bother with deduplication (off by default) but do turn on compression.
I recommend using a RAIDZ6 on heaps of tablets then. Should be good up to two broken tablets by VDEV of 8-12 tablets. Do a vigorous scrub now and then, perhaps with a metal brush. If in doubt, convert to read-only by burying the tablets.
Hope this helps.
to produce this one bitcoin.