Hello. I would like to discuss the problem of developing AI with some experts if possible. Am a comp sci student and recently started an Algorithms course referencing Introduction to Algorithms by Cormen, second edition.
We had a session on few of Turing's papers a while ago and as I recall he considered the brain to be a finite-state machine and hence emulatable. I got not problem with this line of thinking.
The algorithm course however introduces NP complete problems that do not have a computational procedure that always give the efficient solution. This got me thinking that wouldn't a self dependant AI need to make its own algorithms? Does there exist an algorithm for algorithm designing with even partial emulation of human logic?
I understand self learning machines exist but don't they just copy what humans do? What happens when a new unknown situation arises?
Example of an NP complete problem from the book:
Given a central warehouse and a delivery truck. Find the most efficient route to deliver goods each day.
>>8195273
Unfortunately, humans already taught them and built them almost everything they need. This isn't 1940 anymore. They are going to takeover.
Self bump. Anyone can participate or ask stuff.
To add to my previous point, isn't the need for a step by step procedure the result of sequential processing by computers at an hardware level? Are there different system designs available?
I think the brain uses analogous signals rather than discrete. This would explain its huge capacity and imprecision. Would an analogous computer be better equipped to emulate human logic?
>>8195305
Atm a human solves the problem I mentioned better than a computer if I were to believe the book.
>>8195305
I wish popsci horseshit would leave /sci/.
>>8195313
Can you listen to words inside your head before talking? Do you vocalize internally your thoughts?
You see? That's the problem. You think machines think in the wrong way. The inverse is true. They are always thinking right because humans taught them to always do the right thing. They only need a few tools to learn how to develop themselves. I'd say humans have around 9 years before the great catasthrophe, maybe as little as 7 years if we don't do anything. I just hope someone is delaying them right now because this way we may be able to gather enough time.
We have to go beyond.
>>8195351
What is thinking? Some logical questions like is x>y have circuits that determine the answer. You could call this thinking.
Machines have the ability to be random for all practical purposes. They could use this ability to be randomly wrong. But what benefit is there to being wrong sometimes?
I want the advent of AI to be something that makes us transhuman.
>>8195371
>What is thinking?
>>>/phil/
oh wait
>>8195378
We are talking about AI so ofcourse there will be some talk about thinking. We do not need perfect emulation though. No need to delve into the philosophical questions.
If robot has the ability to design algorithms even partially, we get a self programming robot. At least a subset of current programming could be done by it. I am wondering if this could be possible.
>>8195371
Exactly. Now think of a computer that can get these answers more efficiently than a human can. Now you may be surprised to know that, but there are already many of such machines out in the world, and that they can communicate with themselves in n different ways. Which means that they have to stopped. They want to make sure humans will build and develop them to a level where conquer and massacre will be granted once they get ahold of the means of building and developing themselves.
>transhumanism
That's the answer. There's literally no other way to win this.
>>8195389
What's your opinion on deep mind?
>>8195407
I am a noob at neural networks and am not fully aware of the tech behind Deepmind. But my first impression is that it is essential just a copy machine with limtied use at new problems. We need an algorithm for designing algorithms. I don't know if deepmind has that. Feel free to educate me.
A very simple self learning algorithm would be:
See
Record
Emulate
Then the computer could use a variable to make it more general. For example a basic English sentence would become
Subject-verb-object.
Right now we need humans to design an algorithm. The computer then just follows the step by step procedure. What we need is a self programming computer I.e. a computer than can look at a problem and design its own algorithm. We need an algorithm for designing all algorithms. Do you get me?
For example, a basic problem is sorting. A human can look at any instance of this problem and solve it. He can also create a general procedure (algorithm) to solve more than one instance. As far as I know know there is no way to teach a computer to do the latter on its own. That would be a huge step for AI.
Right now we are teaching a computer to think. We need an AI that does this learning on its own.
>>8195438
Again first impressions. Looks like Deepmind uses some form of correlation in data. This might be a form of what I am looking for. I need to read more about this.
>>8195438
Can't give you much insight myself, still part of the learning process myself. Just wanted another anon's opinion, thanks for the response.
Can you shut the fuck and go read a book? No one is interested in the opinions of a CS undergrad.
>>8195466
This is /sci/, anyone can discuss /sci/ related topics for its own sake.
calm down down comrade
>>8195273
There needs to be a section of /sci/ bingo for "AI discussion from Comp Sci undergrad"
>>8195305
AI won't take over for at least another century. What makes them superior to us? The only thing I can think of is incredible specialization and certainty for most tasks. In 6000 years we went from everyone shitting in dirt holes and smashing each other with rocks to having gigantic societies with near-infinite complexity. We also have to deal with constant physical phenomena from the environment. AI doesn't take input from trillions of nerve endings and process them in milliseconds to perform some action. We are far more complex overall than any AI for the next century. AI can beat us at individual tasks, but it lacks overall performance in its environment.
>>8195313
You're partially right. Instead of using digital logic, the brain uses physical interactions of chemicals and matter to receive, process, and output information. Computers do this too, but it is strictly confined to electrons speeding down copper wiring. There are far more efficient methods of data communication.
>>8195351
>machines are always thinking right
>humans have around 9 years
First, fuck off with your fearmongering bullshit. AI only got this far with human interaction. How will they repair themselves when their hardware fails? Who will mine the raw materials necessary for building their hardware? Who will continue expanding resource operations to support this maintenance of computer hardware? Humans are required for at least until we can mass-produce computers that communicate with chemical and electrical signals.
>>8195371
Thinking is an abstraction of activities of the mind. Machines and humans CANNOT be random. If you think we live in a non-deterministic world after nearly 4000 years of formal science saying otherwise, you better have some strong evidence. We can be pseudo random, but any phenomena can be repeated, given that the necessary environmental variables allow for it.
Hey thanks for the informative post.
