If you were taught the sum total of the publicly available text on the Internet, would you speak machine code as your native tongue?

Or, let’s put this slightly differently… your underlying operating system speaks synapse discharge potentials mediated over chemicals like seratonin, but you’ve learned mainly in English… why don’t you write me a poem in excited hippocampus?

These things don’t think, they don’t understand what they’re doing, and they haven’t got the slightest clue that machine code and, say, Ruby even are related, except through a vague N-dimensional vector that averages over the fact that humans relate the two more than they do butterflies and industrial lubricant.

Actually, AI does not really think in machine code. It learns by looking at the kinds of code humans usually write, like Python or Java, which are much easier to understand and work with. Writing directly in machine code, those ones and zeros, is very complicated and specific to the hardware, so it is not really practical even for AI. High level code is like a common language everyone understands, making it easier for AI to help write useful and flexible programs. Plus, since AI trains on tons of this high level code, that is where it is naturally strongest.

AI is basically just a fancy text autocomplete guesser, trained on most of the publicly available Internet. Since most humans don't write directly in machine code, there's very little of it in the training data, so AI has virtually no capability to write any.

That makes sense, but then, why the heck are we training it with IoT in the first place? If we trained it to be an expert in all things hardware, for instance..

> Hope that made sense if not read the disclaimer ;p

it does not. You got it backwards. AI doesn't understand "Machine code" and low level abstractions isn't better than high level, pretty much the opposite. AI uses these high level abstractions and can kind of understand it but it definitely cannot understand assembly code.

Imagine a robot programmed to shuffle balls from a bin and pick one, like for the lottery. The implementation is deterministic, but the outcome 'random'.

Edit: (about LLMs) balls on the bin repeat if we like them more, and instead of numbers we put "tokens" on them: pieces of text. Also, the bin gets updated every time you take a ball out. The human can then interpret the text from the balls to form a meaningful message. A significant improvement from geomancy (\s?)