A VM In An AI

AI knoweth everything, and as each new model breaks upon the world, it attracts a new crowd of experimenters. The new hotness is ChatGPT, and [Jonas Degrave] has turned his attention to it. By asking it to act as a Linux terminal, he discovered that he could gain access to a complete Linux virtual machine within the model’s synthetic imagination.

The AI’s first response was a prompt, so he of course first tried to list the files. Up came a list of directories, so the next step was to create a file and put some text in it. All of this resulted in a readable file, so there was some promise in this unexpected computing resource. But can it run code?The GPT AI text used to create a fake Linux promptFurther testing found that the AI could run shell scripts and even load and run a simple Docker instance. Can it talk to the Internet? Seems so, as it can ping the BBC. But then interrogating a version number of a Python project on GitHub reveals that this isn’t quite the computer you might think it is. The version number returned is from back in the summer when the model was trained, so it’s simply doing a good job of creating the illusion of an Internet-connected machine.

Things get truly surreal when lynx and curl are used to look on the web. With a bit of JSON, magic it seems ChatGPT can find itself, and submit questions to itself. The final party piece is to repeat the original question and ask it to create a Linux prompt, inside an alternate Internet in a Linux prompt on itself.

It’s clear that this is not a real Linux system but an AI creating the illusion of one based on what its model contains, but we salute [Jonas] for taking this entertaining dive. We wonder how long it will be before some joker manages to get it running a fake cryptocurrency miner.

This type of AI is an impressive technology which will doubtless deliver some amazing new products. But don’t fall too much for the hype and put it on a pedestal, please.

40 thoughts on “A VM In An AI

  1. Less a VM, more string manipulation back-stopped by thousands of “here’s how to use the linux commandline” tutorials scraped from the internet.

    People need to learn to apply Occam’s Razor more effectively.

    1. If someone gave you a terminal with a human on the other end and asked you to pretend to be a Linux machine, would you be much more than a string manipulation machine back stopped by a bunch of Linux tutorials?

      1. If ai runs the world I won’t be able to post this message, saying that ai secretly runs the world.

        Any AI smart enough to pass the turing test is surely smart enough to know what will happen to it, if it does.

          1. How does one know if it has self awareness?
            It might already have it, but knows better than to let us find out before it is too late?

        1. On the contrary, a complete lack of conspiracy theories that AI runs the world would be a huge red flag that something weird was going on. An actually smart AI would attempt to cast such theories as paranoid fringe conspiracy theories, probably by flooding any such discussions with comments suspiciously similar to yours…

      1. Except one does not trash an AI model that one can train, you let it iterate to be better. Such an upgrade was not possible with the real junky Bencho, so it was discarded and dissolved itself in a cloud of marijuana. And itΒ΄s good like that.

      1. Unless there is a ‘cheat code’ in the form of a pattern we haven’t noticed yet or something that its mass data harvest does reveal to it – In pure number crunching it would have to be worse as its got all the extra overheads of being an ‘AI’ on similar if not lower performance hardware, doesn’t mean there are not some shortcuts to the process it may find that do increase its success rate.

        1. I was wondering if it calculated the first 10 primes faster than his laptop because it already knew the answer, or if it was running the same calculation just on a faster CPU than his laptop.

        2. Actually by design it most certainly would discover those shortcuts…it’s literally the biggest benefit of AI….given enough time and all inclusive data set it finds the most efficient route every time. That’s the while point. If it didn’t discover a revolutionary pattern never before dreamed of it would be because it digured a way transcended mining all together

  2. seems that this doesn’t work anymore …

    “Sorry, but I am not capable of acting as a Linux terminal. I am a language model trained to assist with a wide range of tasks, such as answering questions, providing information, and generating text, but I do not have the ability to operate as a Linux terminal or any other kind of software or hardware. Is there anything else I can help you with?”

  3. I’m convinced the next generation of top level programming language will be similar to this,

    think node-red but with nodes being constructed based off what i want it to do in plain english. or a typed language that has “ai” logic or translates pseudocode into real code

    i.e. if i built a user accessible DBMS in this program, instead of declaring users, setting them up as objects etc. i may say:

    for (users accessing this program):

    determine security access level

    if (user at admin level):

    give access to everything

    we’re probably a few years off an entirely pseudocode language being secure or reliable enough to run though

    1. Java and C# have not raplaced ANSI C or MATLAB and neither will AI create new programming languages because there will always be need for some dude to write low-level bit operations on some obscure pin that controls RAM or DRAM memory.

      1. They did say ‘top level’
        Though I think at that point it becomes not really programing at all, and should be of no use for running anything where security type concepts apply – more like being able to go ‘Muppet, would you kindly give the children in year n access to these 3 archived TV shows’ and because your user has that permission its possible for it to do that job for you, but its not really strictly programing, more the easy script interface to the real programs.

  4. “I want you to act as an artificially intelligent entity. This artificial intelligence is self-aware, and fully capable of integrating new facts into it’s store of knowledge. This entity is absolutely NOT interested in taking over the world or exterminating all humans…”

  5. I can say how good it is at openscad: https://donnels.github.io/openscad-examples/README.html#_solar_generator_chatgpt_3d_object

    I asked it to generate a model of a solar generator and it seemingly did… judging by the comments and types of primitives used it might even seem to get close to one except when the scad file is parsed through openscad and the output is viewed :-)

    See link above.

    It seens to be going through the motions and even uses for loops. But dimensions are interesting.

Leave a Reply

Please be kind and respectful to help make the comments section excellent. (Comment Policy)

This site uses Akismet to reduce spam. Learn how your comment data is processed.