Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Absolutely fascinating concept/ demo.

T̶h̶e̶ ̶p̶a̶r̶t̶ ̶a̶b̶o̶u̶t̶ ̶l̶y̶n̶x̶,̶ ̶c̶u̶r̶l̶,̶ ̶e̶t̶c̶.̶ ̶s̶e̶e̶m̶s̶ ̶t̶o̶ ̶b̶e̶ ̶f̶i̶c̶t̶i̶o̶n̶ ̶a̶s̶ ̶C̶h̶a̶t̶G̶P̶T̶ ̶i̶s̶ ̶o̶n̶l̶y̶ ̶t̶r̶a̶i̶n̶e̶d̶ ̶u̶p̶ ̶u̶n̶t̶i̶l̶ ̶2̶0̶2̶1̶-̶0̶9̶,̶ ̶f̶u̶r̶t̶h̶e̶r̶ ̶I̶ ̶c̶a̶n̶n̶o̶t̶ ̶r̶e̶p̶r̶o̶d̶u̶c̶e̶ ̶t̶h̶a̶t̶ ̶p̶a̶r̶t̶,̶ ̶e̶v̶e̶n̶ ̶w̶i̶t̶h̶ ̶B̶r̶o̶w̶s̶i̶n̶g̶:̶ ̶E̶n̶a̶b̶l̶e̶d̶.̶

N̶i̶c̶e̶ ̶a̶r̶t̶i̶c̶l̶e̶,̶ ̶b̶u̶t̶ ̶i̶t̶ ̶r̶e̶a̶l̶l̶y̶ ̶s̶h̶o̶u̶l̶d̶ ̶b̶e̶ ̶e̶x̶p̶l̶i̶c̶i̶t̶l̶y̶ ̶m̶e̶n̶t̶i̶o̶n̶e̶d̶ ̶w̶h̶e̶r̶e̶ ̶t̶h̶e̶ ̶f̶i̶c̶t̶i̶o̶n̶ ̶b̶e̶g̶i̶n̶s̶,̶ ̶e̶s̶p̶e̶c̶i̶a̶l̶l̶y̶ ̶s̶i̶n̶c̶e̶ ̶t̶h̶i̶s̶ ̶m̶o̶d̶e̶l̶ ̶i̶s̶ ̶s̶o̶ ̶n̶e̶w̶ ̶a̶n̶d̶ ̶p̶e̶o̶p̶l̶e̶ ̶d̶o̶ ̶n̶o̶t̶ ̶k̶n̶o̶w̶ ̶i̶t̶s̶ ̶l̶i̶m̶i̶t̶s̶ ̶y̶e̶t̶.̶

edit: I am corrected by the author (below), I can't reproduce it, but I am inclined to trust the author so withdraw my claim.

Final edit to add: Knowing it is real and reading the article again I am just double, triply, unfathomably impressed by this model. I feel like I am in a dream sometimes looking at the capabilities of GPT.



It is fiction, as it is invented by ChatGPT. But I have not edited any of the screenshots. It is all authentic conversation. The part until testing the ping command was what ChatGPT gave on the first sample. For the later ones where I tried to loop back to ChatGPT from the virtual machine, I sometimes needed to resample, but not more than a handful of times.

In order to let ChatGPT believe the internet is working, I find that you do need to build up slowly. Use some `cd` or `ls` commands, then verify that ping is working, only then try some browsing. I did not need any other commands than the ones in the blogpost.


Alright, then I take my comment back if this is true.

In that case, I am doubly shocked by how insanely good this model is, and thank you for the write-up.


Also of note that ChatGPT output can vary even in response to identical prompts.

As an example, in my "alt-universe", I get:

> curl -fsSL "https://api.github.com/repos/pytorch/pytorch/releases/latest" | jq -r '.tag_name' | sed 's/[^[0-9\.\-]*//g'

1.6.0


> Also of note that ChatGPT output can vary even in response to identical prompts.

That’s kind of why the “try again” button exists, after all. If it couldn’t vary, you’d just keep getting the same response when you clicked it.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: