Quote from the paper: Actions correspond to four motor angle commands and sensations correspond to the absolute coordinate of the end effector.
This means that it cannot see, and either needs additional 3D motion tracking hardware or human handcrafted logic to detect the position of its end effector. Everything it does is just learn an inverse kinematics model, so that you can command it to move the end effector to a certain position afterwards. But it cannot learn, for example, to detect position and orientation of a target object or to avoid obstacles in its way.
Wrong for the live game against MaNa. Quote: "Following the broadcast of the recorded matches, DeepMind introduced a new version of AlphaStar that MaNa took on in a live match. The agent that played the live game didn't have the benefit of the overhead camera and instead had to make decisions on where to place its focus in the same way a human would."
But you are right for the recorded games from December 2018.
The Turing Test can be fooled by stupid chatbots. Take a billion recorded conversations between real humans and play back the best fitting answers to the tester's questions, adding some hardcoded consistency, e.g. avoid switching between male and female identities. Just a stupid Chinese Room that doesn't know what it's talking about. Obviously, Alan Turing could not foresee the internet with its massive data amount.
The Total Turing Test on the other hand cannot be fooled by symbol juggling and is therefore really genius. If your machine didn't undergo a human education in a humanoid body, then it has no chance to pass.
2. Someone has to act on the plan. If it uses humans, then it has to communicate with them physically. Physical communication is just some kind of action. If it does not act at all, it is useless. The anonymous definer seems to be afraid of robots. Maybe he is just a coward.
5. Somebody has to write a novel-situation/not-novel-situation classifier;)
You cannot try to build-create an unselfish AGI because it must have a body and therefore its reward function must have a location in space. You can try to educate-create such an AGI. But first you'll have to build a piece of hardware that is educatable.
Computers can be trained to eat in VR. Just give the RL algorithm a positive reward when the virtual cookie touches the virtual mouth, and then delete it. Once the agent has learned to hunt for cookies, make the task more difficult, for example enclosing the cookie in a transparent box and giving the agent a tool to open it. And if you add other agents and hunger, so that the agent saves up food for later instead of eating as much as he can, they will begin to act socially and hide, steal, and trade food.
How can it be afraid of being in the dark when it's off then?
This would only make sense if it's a blinding switch. The robot cannot see, and therefore it would get frightened by sudden touch. But I don't think that this NAO robot has something like the iCub Skin that can feel touch.
That experiment looks like a big load of BS to me.
Slashdot Mirror
Just reward the RL algorithm for being traceable.
Quote from the paper: Actions correspond to four motor angle commands and sensations correspond to the absolute coordinate of the end effector.
This means that it cannot see, and either needs additional 3D motion tracking hardware or human handcrafted logic to detect the position of its end effector. Everything it does is just learn an inverse kinematics model, so that you can command it to move the end effector to a certain position afterwards. But it cannot learn, for example, to detect position and orientation of a target object or to avoid obstacles in its way.
The paper is on sci-hub.tw. The DOI is 10.1126/scirobotics.aau9354. The source code is at https://github.com/rjk2147/Tas....
Wrong for the live game against MaNa. Quote: "Following the broadcast of the recorded matches, DeepMind introduced a new version of AlphaStar that MaNa took on in a live match. The agent that played the live game didn't have the benefit of the overhead camera and instead had to make decisions on where to place its focus in the same way a human would."
But you are right for the recorded games from December 2018.
The Turing Test can be fooled by stupid chatbots. Take a billion recorded conversations between real humans and play back the best fitting answers to the tester's questions, adding some hardcoded consistency, e.g. avoid switching between male and female identities. Just a stupid Chinese Room that doesn't know what it's talking about. Obviously, Alan Turing could not foresee the internet with its massive data amount.
The Total Turing Test on the other hand cannot be fooled by symbol juggling and is therefore really genius. If your machine didn't undergo a human education in a humanoid body, then it has no chance to pass.
2. Someone has to act on the plan. If it uses humans, then it has to communicate with them physically. Physical communication is just some kind of action. If it does not act at all, it is useless. The anonymous definer seems to be afraid of robots. Maybe he is just a coward.
;)
5. Somebody has to write a novel-situation/not-novel-situation classifier
It would make no sense for rocks to simulate future threats. They have no legs and cannot run away.
Cyclists tried to extract the subconscious common sense rules with their conscious. They came up with some 15 Million rules. That's ridiculously few.
If it is not human then it will not understand me and I will buy a competitor's product instead.
You cannot try to build-create an unselfish AGI because it must have a body and therefore its reward function must have a location in space. You can try to educate-create such an AGI. But first you'll have to build a piece of hardware that is educatable.
Computers can be trained to eat in VR. Just give the RL algorithm a positive reward when the virtual cookie touches the virtual mouth, and then delete it. Once the agent has learned to hunt for cookies, make the task more difficult, for example enclosing the cookie in a transparent box and giving the agent a tool to open it. And if you add other agents and hunger, so that the agent saves up food for later instead of eating as much as he can, they will begin to act socially and hide, steal, and trade food.
If causality is that funny, well then ...
If-statement: "What's my purpose?"
Programmer: "You wait for program counter."
How can it be afraid of being in the dark when it's off then? This would only make sense if it's a blinding switch. The robot cannot see, and therefore it would get frightened by sudden touch. But I don't think that this NAO robot has something like the iCub Skin that can feel touch. That experiment looks like a big load of BS to me.