If the AI is incapable of observing the outside world, it's impossible for it to learn even the most basic science. It wouldn't know our world existed, all it would know would be, at best, the hardware it's running on. But if you code it to be platform-independent, it won't even know that :)

You may be interested in AI Safety - Computerphile. Also slightly off topic, listen to the discussion inspired by Superintelligence on Hello Internet #52: 20,000 Years of Torment, there is some interesting discussion on its reddit.

But there are docs inside the PC. Maybe not ones you've put there intentionally, but even a basic Unix/Linux OS contains thousands of files, including programs, scripts and config files. Just the list of file names would be an interesting source of learning for the computer, but the contents of those files would be even more interesting. Error message, manual pages, output from some of the programs... all would be sources for learning.

The one big problem with your proposal is your idea that you would never give it any info about science or history. Frankly, if you want to be able to communicate with it at all, that's going to be next to impossible to achieve because you're going to want to communicate in English, and natural English is riddled with historical artefacts. Even without any other data, just English words, an AI could deduce a lot about the evolution of the language and the fact that it has taken syntax from multiple previous languages. There is a lot of hidden meaning in words.

I know you said there are no docs or anything, but before you first "booted up" the AI, did you program to know how to read a language? When you use the terminal, is that part of the learning or does it already know how to communicate via terminal?

The computer could observe the power source to which it is connected.... and there is non random signal noise on power lines... but the method which it uses to improve its own code will either be based upon the power line input, or upon the inherent assumptions in the OS code itself.

I understand what you're trying to do with this question, but in real life AI systems work mainly because of the vast amounts of data used to train them. I don't know how feasible it is to actually program an AI with no data.

Also, your tests will provide it with information, not that it will understand most of it without definitions it can comprehend, or explanations it can comprehend.

To clarify, I didn't deny the error messages or him obtaining new info from things that are already there. No docs, audio and all that just means there are nothing who is not already there with the system. I interpreted he learned to talk from reading me asking questions like "Hello, are you alive?" or another similar questions made regularly to test it and probably the things already inside the OS. I just gonna talk to the AI to understand their evolution. I know they gonna try to understand me, but I never gonna talk abount anithing related to science.

You said that AI is "making itself more efficient". How will it decide, what is efficient and what isn't?

@enkryptor Trial and error I assume. I am a developer, but not an expert on AI so I asume he can do something like genetic style or anything like that. Anyway it's not the point of the questions, we assume he just do that.

@Malkev if there is no input, how can you define an "error" ?

@enkryptor don't understand what you are saying. When you code something there can be errors, and those errors usually are commented to print it to you and see what happened. The errors are already defined.

@Malkev when I code something, I always have kind of requirements. Living in a complete emptiness, how the AI can decide, what is "better", what should it "improve"? Why does it have to improve something, why write any programs, for what purpose? For an empty world, an empty cycle that does nothing would be sufficient.

@enkryptor as I already said, not related. I dind't create a super-intelligence in real life. We are assuming that already happened, he is self-conscious. The question is another.

@Malkev I still think it is related. You're asking if the AI can achieve a particular goal, that obviously depends on the launch environment.

"Hello, are you alive?" How could it have the slightest clue what that means without a huge amount of knowledge about the outside world/reality?

The most powerful consumer PC you can buy or build. Given what budget? Some pretty massive multi-processor systems can be bought/built by hobbyists.

@user2338816 There are no limit in budget, just nothing a single person can purchase or was too suspicious. And I don't know how he learned to talk, but he did. Some answers solved that point in differents ways.

Not how to talk, but rather how to comprehend "alive". Even a minimal comprehension of "alive" requires a huge grasp of our basic reality. And even "you" necessarily implies "other" or "not you". Those two together may be enough to infer everything needed.

The question is more like "Can the AI develop science so the humanity achieve singularity?". Singularity is not only the AI obtaining consciousness, it's the explosion of knowledge.

Simple AI answer: If your [Super AI thing] isn't sitting in an airlocked (as in, with actual airlocks) faraday cage, powered only by a vat of lead and a vat of acid with some wires running between them, then it can escape your prison and take over the world.

A bit off topic, but I thought you would be interested to know that I once downloaded all of english wikipedia and and experimented on big data and machine learning technologies against it with a small cluster of computers. The data itself was only about 30 GB, which just about any consumer hardware now can hold. The processing took a small cluster of computers running overnight. The amount of information the human brain is estimated to process is still way more than a desktop computer. I used this by the way: cs.waikato.ac.nz/~ihw/papers/12-DM-IHW-OStoolkit-wikimining.‌​pdf

@Kaz and anyone else worrying about AI escaping. The AI isn't going to be able to make any observations that its hardware does not support, nor is it going to have any means of transmission that it's hardware does not support. At least any signal wouldn't reach far at, all even if it managed to hijack it's own circuitry for that purpose. Even if a sensitive instrument could infer some kind of broadcast data based on minute power line voltage perturbations, the power supply on a PC can't measure it, and even if it did, can't physically report it to any other hardware.

@StephenLujan There's a lot that computer hardware can do that it was not intended to do. Case in point, a memory bus is, basically, just an antennae and can be used to send radio signals. Given complete control over the hardware in a typical computer, how many different ways are there to send & receive digital & analog signals? Answer: far more than we can imagine. So, assuming any kind of artificial intelligence, either it's in a faraday cage or you have to assume it has the means to escape.

@Kaz I see it is capable of transmitting information using frequency modulation at up to 60 Bytes per second. It doesn't sound like anyone has been able to receive radio transmissions on these malware infected devices, and being able to send observable information is impressive, but this doesn't mean it could operate in the bandwidth and protocol of something like Bluetooth.