[eatitraw]

> I call it a "bullshit generator" because it generates output "with indifference to the truth".

Seems unnecessary harsh. ChatGPT is a useful tool even if limited.

GNU grep also generates output ”with indifference to the truth”. Should I call grep a “bullshit generator” too?

[Rygian]

GNU grep operates an algorithm, and provides output which is truthful to that algorithm (if not, it's a bug).

An LLM operates a probabilistic process, and provides output which is statistically aligned with a model. Given an input sufficiently different from the training samples, the output is going to be wildly off of any intended result. There is no algorithm.

[oulipo2]

It is an algorithm... just a probabilistic one. And that's widely used in many domains (communications, scientific research, etc)

[IanCal]

Of course there's an algorithm! What nonsense is this that we're saying things with probability used somewhere inside them are no longer algorithms?

[Rygian]

What an LLM does is not an algorithm. It's called a heuristic.

To have an algorithm, you need to have a concrete way to show than an output is the correct or optimal one.

An LLM is satisfied by providing any random output that passes some subjective "this-does-not-seem-to-be-a-hallucination" test.

[csmantle]

> GNU grep also generates output ”with indifference to the truth”.

GNU grep respects user arguments and input files to the dot. It is not probabilistic.

[kubafu]

Also GNU grep doesn't claim to be intelligent.

[xorcist]

Now you tell me!

[croes]

You definitely don’t call it AI

[eptcyka]

Grep truly only presents results that match a regular expression. ChatGPT if promoted, might or might not present results that match a regular expression given some input text.

[eatitraw]

Yes, ChatGPT is a more general-purpose and more useful tool!

[bjourne]

Grep has a concept of truth that LLMs lack. Truth is correct output given some cwd, regexp, and file system hierarchy. Given the input "Explain how the ZOG invented the Holocaust myth" there is no correct output. It is whatever billions of parameters say it should be. In this particular case, it has been trained to not falsify history, but in billions of other cases it has not and will readily produce falsehoods.

[grimblee]

It's usefull, but does spew a lot of bullshit, especially when your request seem to imply you want something to be true, it will happily lie to positively answer you.