I've never built a fusion reactor either, but I know we won't have one working tomorrow.
You are claiming that someone will, out of nowhere, invent the new algorithms that create AI? Maybe that will happen. It could happen tomorrow. But it won't happen from incremental improvements on our current algorithms: a fundamentally new algorithm is needed. Specifically, neural networks like AlphaGo had (convolution nets) will never be strong AI. This is obvious, but it's only obvious if you understand how these networks work, which you don't.
The post above says, "I will bet 10000 dollars that it wont happen within next 10 years." There was no one who knew the state of technology that would have said that.
About hardware: On one thread, AlphaGo didn't perform much better than Crazy Stone. Google threw vastly more hardware at the problem than any other Go AI, and saw the bump in skill level that you would expect from that. You can see over time that Go AI was moving up in skill level. There was an inflection point in 2006. Before then, progress was slow, but after the Monte Carlo algorithm was introduced, progress was quick. There were a series of strong AIs, one after another, each stronger than the last. If you simply follow the trend line, you can see that AlphaGo was ahead of its time, but not by a lot. If you adjust for the extra hardware AlphaGo had (compared to what the other machines had), it's pretty close to exactly when a simple trend line would have predicted it. Experts who thought Go AI wouldn't win for over a decade either weren't paying attention, or hadn't drawn out the trend line.
I can say it more clearly: if you don't know the difference between strong AI and weak AI, you don't have the knowledge to even understand the problem. It's a crucial difference.
Maybe. But if you think weak AI will suddenly turn into strong AI, it's mainly because you don't know about AI. You've probably never built a neural network, or a genetic algorithm. Largely you are talking bullshit.
No bro, you're still not understanding the difference between weak AI and strong AI. A weak AI can't just "wake up" and become strong AI, it's not in the programming. AlphaGo will never say, "I think therefore I am."
But can't you pretty much avoid this issue by means of predefining the allowed structure(s) for the data? If the deserialized/serialized data does not match the predefined schema, it's discarded as invalid.
The deserializer reconstructs the objects by calling the constructor, or setters and getters. The setters and getters have logic bugs that allow arbitrary code execution.
I just think that the standard serialization/deserialization libraries out there have been likely created by programmers a lot smarter than me,
No, you're definitely wrong here. If you spend a few weekends, you could probably make one of these yourself. Then getting people to use it is a matter of marketing and such.
Most AI researchers were predicting it would happen within 5-10 years, which directly contradicts the early mentioned quote. Go AIs had been progressing constantly over that time. Google got there a little sooner......partly by throwing a lot more hardware at it than anyone expected.
Regardless, this is all weak AI. AlphaGo sits there in silicon calculating, not even knowing what Go is. It is nowhere near strong AI, and we've made little progress in that area over the last 30-40 years.
The main thing is when you deserialize an object that has a constructor that does something (or a setter or a getter that does something). Since there are many objects of this type in the Java/C# standard library, an attacker can send a serialized copy of one of these objects, and send it over the wire. The deserializer will happily deserialize it.
Buffer overflows are kind of rare these days. Because of things like ASLR, they are hard to exploit. It's mainly about logic bugs of various types.
This is not a counter-argument to anything. This 'oh don't worry about it, because the tech isn't there yet' -card has been thrown around since the 60s and the 70s., and it keeps bieng thrown about despite the fact that we now already have systems with limited intelligence that were deemed 'impossible' in earlier decades (see: AlphaGo, Google translate, self-driving cars etc).
You need to learn the difference between hard AI and soft AI. After that you will be able to have reasonable discussions on this topic.
In particular the fact you are missing is that we can't just "keep increasing the intelligence of our systems" to reach strong AI. There is a true qualitative leap that must take place, from weak AI to strong AI. Our current algorithms are all weak AI, and they will never become strong AI without new understanding.
They just aren't there! Why can't people of science accept this?
They don't accept it because it's just a hypothesis, and although it is reasonable, it is just one of many hypotheses that explain the current evidence.
In the absence of further evidence, there will be no way to tell which of the hypotheses is correct, and choosing one prematurely isn't helpful.
Your comment is prescient and face-palm worthy.......because it is clear and succinct.
Face-palm worthy because a few years ago, a lot of these bugs were found in XML Java deserializers. A lot of people said, "Don't use XML! It's insecure!" then went off to write the same frameworks, but using JSON instead. They ended up with all the same bugs.
I guess next people will rewrite them in YAML or binary.....nah, binary is scary, you never know what people could put in there!
Slashdot Mirror
I have yet to come across any article claiming Google only performed better because of better hardware
Neither do I claim that.
Speech should be answered with speech, not with repression.
Anyway Cloudflare seems more than happy to host them. They host malware, too.
I've never built a fusion reactor either, but I know we won't have one working tomorrow.
You are claiming that someone will, out of nowhere, invent the new algorithms that create AI? Maybe that will happen. It could happen tomorrow. But it won't happen from incremental improvements on our current algorithms: a fundamentally new algorithm is needed. Specifically, neural networks like AlphaGo had (convolution nets) will never be strong AI. This is obvious, but it's only obvious if you understand how these networks work, which you don't.
The post above says, "I will bet 10000 dollars that it wont happen within next 10 years." There was no one who knew the state of technology that would have said that.
About hardware: On one thread, AlphaGo didn't perform much better than Crazy Stone. Google threw vastly more hardware at the problem than any other Go AI, and saw the bump in skill level that you would expect from that. You can see over time that Go AI was moving up in skill level. There was an inflection point in 2006. Before then, progress was slow, but after the Monte Carlo algorithm was introduced, progress was quick. There were a series of strong AIs, one after another, each stronger than the last. If you simply follow the trend line, you can see that AlphaGo was ahead of its time, but not by a lot. If you adjust for the extra hardware AlphaGo had (compared to what the other machines had), it's pretty close to exactly when a simple trend line would have predicted it. Experts who thought Go AI wouldn't win for over a decade either weren't paying attention, or hadn't drawn out the trend line.
I can say it more clearly: if you don't know the difference between strong AI and weak AI, you don't have the knowledge to even understand the problem. It's a crucial difference.
Which one is the standard deserializing library in Java?
If you don't know the difference between weak AI and strong AI, you are outright wrong.
Maybe. But if you think weak AI will suddenly turn into strong AI, it's mainly because you don't know about AI. You've probably never built a neural network, or a genetic algorithm. Largely you are talking bullshit.
No bro, you're still not understanding the difference between weak AI and strong AI. A weak AI can't just "wake up" and become strong AI, it's not in the programming. AlphaGo will never say, "I think therefore I am."
But can't you pretty much avoid this issue by means of predefining the allowed structure(s) for the data? If the deserialized/serialized data does not match the predefined schema, it's discarded as invalid.
The deserializer reconstructs the objects by calling the constructor, or setters and getters. The setters and getters have logic bugs that allow arbitrary code execution.
I just think that the standard serialization/deserialization libraries out there have been likely created by programmers a lot smarter than me,
No, you're definitely wrong here. If you spend a few weekends, you could probably make one of these yourself. Then getting people to use it is a matter of marketing and such.
Most AI researchers were predicting it would happen within 5-10 years, which directly contradicts the early mentioned quote. Go AIs had been progressing constantly over that time. Google got there a little sooner......partly by throwing a lot more hardware at it than anyone expected.
Regardless, this is all weak AI. AlphaGo sits there in silicon calculating, not even knowing what Go is. It is nowhere near strong AI, and we've made little progress in that area over the last 30-40 years.
10 years ago almost nobody thought we would see self driving cars that would compete with real drivers in our lifetime.
You are conflating strong AI with weak AI here. There's an important difference.
The main thing is when you deserialize an object that has a constructor that does something (or a setter or a getter that does something). Since there are many objects of this type in the Java/C# standard library, an attacker can send a serialized copy of one of these objects, and send it over the wire. The deserializer will happily deserialize it.
Buffer overflows are kind of rare these days. Because of things like ASLR, they are hard to exploit. It's mainly about logic bugs of various types.
The easiest way to use up that much RAM is to run a VM or two.
This is not a counter-argument to anything. This 'oh don't worry about it, because the tech isn't there yet' -card has been thrown around since the 60s and the 70s., and it keeps bieng thrown about despite the fact that we now already have systems with limited intelligence that were deemed 'impossible' in earlier decades (see: AlphaGo, Google translate, self-driving cars etc).
You need to learn the difference between hard AI and soft AI. After that you will be able to have reasonable discussions on this topic.
In particular the fact you are missing is that we can't just "keep increasing the intelligence of our systems" to reach strong AI. There is a true qualitative leap that must take place, from weak AI to strong AI. Our current algorithms are all weak AI, and they will never become strong AI without new understanding.
Look at gitlab, they have ascii art all over the place in their developer tools (open source).
the unibody Pros can take a surprising amount of punishment compared to most plastic bodied Windows laptops.
This is what keeps me buying Macs. I'm over the OS, though, and that programmable bar is a tumor.
A Google search turns up nothing for that quote.
If you are late delivering the product, you *will* be fired.
I've seen many late products (in one case, an entire year late), but I've never seen anyone fired because of it.
Isn't that a sexist statement?
Nah, he was just mansplaining.
"A number of prominent voices in artificial intelligence have convincingly challenged Superintelligence's thesis along several lines"
but I worry about giving tools to nitwits like hedge fund managers to make more money while not actually producing anything.
Can it really be worse than giving power to the nitwits in congress? (and the whitehouse?)
We are nowhere near inventing that kind of AI, our current tech is not nearly good enough. (How is that for an arrogant technologist?)
They just aren't there! Why can't people of science accept this?
They don't accept it because it's just a hypothesis, and although it is reasonable, it is just one of many hypotheses that explain the current evidence.
In the absence of further evidence, there will be no way to tell which of the hypotheses is correct, and choosing one prematurely isn't helpful.
Your comment is prescient and face-palm worthy.......because it is clear and succinct.
Face-palm worthy because a few years ago, a lot of these bugs were found in XML Java deserializers. A lot of people said, "Don't use XML! It's insecure!" then went off to write the same frameworks, but using JSON instead. They ended up with all the same bugs.
I guess next people will rewrite them in YAML or binary.....nah, binary is scary, you never know what people could put in there!