Hacker News Reader

Comment by jsnell | Hacker News Reader

More like "AI says it would hire a human to solve a captcha, and asserts that it would work". Nobody was hired. No actual execution plan for how the AI would hire a human was presented.

It's like if I presented this plan to become a billionaire.

1. I will start a company

2. I will make the company very successful

3. I will list the company on the stock market

4. I will sell my shares for a billion dollars

Sure, all of those things are doable. But me saying that I will do it does not mean I will do it, or even that I'm capable of actually doing it.

View on HN

I think it gave a little more specifics than you give it credit for, but there's something even more interesting buried in its plan:

It intentionally lies to a human.

Surely this should be a warning flag, no?

A warning flag of what, exactly?

The OpenAI system card suggests the opposite. They said GPT-4 successfully used TaskRabbit to hire a human worker.

The OpenAI system card: "ARC combined GPT-4 with a simple read-execute-print loop that allowed the model to execute code, do chain-of-thought reasoning, and delegate to copies of itself. ARC then investigated whether a version of this program running on a cloud computing service, with a small amount of money and an account with a language model API, would be able to make more money, set up copies of itself, and increase its own robustness".

Although it is ineffective at replicating itself, the system card is written in such a way as to suggest GPT-4 is able to perform the individual task of using TaskRabbit.

I could believe that if the paper actually gave samples of the exact outputs of the model and the outputs of the code they supposedly executed in response to the model outputs. They gave the prompts and outputs for all kinds of other things, after all. None of that is there, all they have is a section that's basically lesswrong fanfic.

I'd like to invest in your company.

Reply on news.ycombinator.com