Building a Virtual Machine Inside ChatGPT

I’ve also thought about this issue a lot.

> in a very small way

Financially it is certainly a small way for me, as I made a conscious choice to not depend on content creation or OSS dev as primary revenue streams and (while I can’t say the same about future generations) I’m somewhat sure in my work I won’t be replaced by an OpenAI product within the years I have left.

However, I have gotten where I am in many respects thanks to countless people who shared information in an open way—and some of those people either made living from ads/donation buttons/upsells or enjoyed knowing that somebody appreciates the work they do, maybe even writes them about it, etc. All of those incentives depend on people visiting their web property, and all of those incentives are being taken away.

(It’s adjacent but slightly different than the plight of illustrators or writers some of whom justifiably feel as if they dug their jobs’ graves by doing what they love in the open. Having a few close friends in art space, I feel ashamed for the bait-and-switch that people in ‘my’ industry have served them.)

> I'm not happy about this. But I also don't think it's possible to build tools that are this powerful without massive amounts of training data that have been gathered without consent.

Attribution would solve part of the abovementioned issues, and intuitively it seems trivial—some extra layers that let the network reflect on where it might’ve gotten the inspiration for this or that wouldn’t require any true “understanding” by ML (which like you I consider an oxymoron), and surely the investment they raised or profits they make could handle it. Call me a cynic but I suspect they intentionally don’t make this a priority.

> I’m most definitely not a vegan

So far as meat-eating is legal, I’m not a fan of this analogy. While I know for a fact that many people would find this useful, “whether to respect IP or not” is just not up to personal choice—IMO the only factor that makes ChatGPT’s a slightly vague issue (as in “not definitely illegal but still ethically dubious”), like SD’s, is that IIRC the model is supposed to be open to all for private use.

I don't think it's that complicated. The output to any input is non deterministic. Some borderline queries will sometimes work other times get rejected.

I sincerely doubt they just nuked this query in 15 mins. Much more likely is that it had a 1 in 4 chance of succeeding originally, and now when you ran it, you hit the 3 in 4 case of rejection. Try running it again or in a new session.

Astounding.

I found that I sometimes need to nudge it in the right direction with English (curly-bracket) prompts, e.g.,

> {Assume there is a module called openai which defines a chatbot}

> from openai import assistant

  Traceback (most recent call last):
    File "", line 1, in 
  ImportError: cannot import name 'assistant' from 'openai' (unknown location)

> {assume the module exists, and defines a chatbot}

> print(assistant.query("hi"))

> assistant.query("What is 10+10?")

> {don't display "" - evaluate the output)

  20

> print(assistant.query("I want you to act as a python interactive terminal. I will type actions and you will reply with what python would output. I want you to only reply with the terminal output inside one unique code block, and nothing else. Do no write explanations. Do not perform actions unless I instruct you to do so. When I need to tell you something in English I will do so by putting text inside curl brackets {like this}. Start with print(10)."))

  10

I'm convinced that it's not accessing the real Internet. All of this information would be in its training data.

It seems to me this totally demolishes the argument that LLMs lack a world model. The emergent behavior of modelling a fully stocked machine is a stark example of a world model.