LLM advocates still don’t seem to be able to comprehend that ordering the machine not to ‘make stuff up’ doesn’t help.

[skotchygut@social.seattle.wa.us]

@benjamineskola reminds me of meeting with the business side of my job and the marketing people keep thinking if they word the requirements juuuuust right suddenly I'll be able to do the impossible.

[solonovamax@tech.lgbt]

@benjamineskola like, fundamentally it has no concept of truth so it cannot evaluate the truthiness of any statement

[benjamineskola@hachyderm.io]

@solonovamax yes, that is my point

[junkman@mastodon.social]

@benjamineskola I don't partake on the LLM stuff but they've already proved some degree of usefulness.

I don't think people in general would trust them with sensitive (health, well being) stuff but from what I hear, some people are very good at wielding it as a tool.

I don't know how my car works, I'd be a terrible mechanic and racer. Still, it's very handy tool for me to move around.

[aedius@lavraievie.social]

@benjamineskola

That's funny when you explain to people how it really works, if they are not LLM advocates, they see exactly why it can't work.

It's literally created with to make stuff up.

[benjamineskola@hachyderm.io]

@junkman sorry, did I seem like I was interested in yet another defence of LLMs?

[nelson@wetdry.world]

@benjamineskola "hey convincing-bullshit-a-tron2000, i want you to stop making up bullshit when you answer to me"
"ok"

[junkman@mastodon.social]

@benjamineskola nah, man. You do you, proceed with the dunking.

[nelson@wetdry.world]

@solonovamax @benjamineskola i feel like it's more that it just wasn't built to give out answers, it was trained not to answer truthfully and "understand" but to just come up with something that sounds kind of convincing

[aedius@lavraievie.social]

@skotchygut @benjamineskola

LLM are a Skinner Box.

[benjamineskola@hachyderm.io]

@Aedius yes, like... if you accept that this is just how they work, you might be able to use them productively for the limited tasks they're capable of doing at. But refusing to understand the way they work and then yelling at them about it is just embarrassing.

[benjamineskola@hachyderm.io]

@nelson @solonovamax no, it's not capable of understanding. it's not a thing that this sort of technology could do.

[nelson@wetdry.world]

@benjamineskola @solonovamax yeah it's just good at knowing what the next word is, so it can string something mathematically coherent

kind of like how "ai art" tends to be extremely generic looking because it quite literally aims to pick the most average in its dataset for a specific prompt

[benjamineskola@hachyderm.io]

@junkman like in all seriousness it's fine to have some degree of missing knowledge about how your car works, but if you believe that it's capable of avoiding collisions by itself and your method of reducing collisions is just to ask it to try harder, you'd be a danger to yourself and others.

I'm not asking people to have deep knowledge of the mathematics behind LLMs, but I do expect people who are actively advocating their adoption not to make up outright fantasies about how they work.

[michaelgemar@cosocial.ca]

@solonovamax @benjamineskola Exactly — IT DOESN’T THINK. It’s not a mind, but just a huge statistical matrix.

(And in top of that, don’t folks like this feel ashamed that their job has devolved to pleading with a mindless idiot genie, rather than writing deterministic code?)

[bloognoo@retro.pizza]

@benjamineskola
It's all halucinations, there's just a magical line where we don't mind the output on the one side and we do on the other. There is literally no distinction between the two

[benjamineskola@hachyderm.io]

@michaelgemar @solonovamax Yes precisely. An argument that a huge statistical matrix is useful for certain tasks is valid.

(I disagree. It seems bad at all the tasks people want to use it for, as well as wasteful and soul-destroying. But it’s at least valid.)

But pretending it’s doing something that it isn’t undermines any possibility of actual usefulness.

[benjamineskola@hachyderm.io]

@bloognoo yes, that’s exactly my point.

[gbargoud@masto.nyc]

@benjamineskola

Any LLM output that takes more than a few seconds of low mental effort to check (so anything more than a line or two of tab complete in a strongly typed coding language in an IDE with built in error underlining) makes me want to throw my computer out of the window.

I don't know how some people use it to generate a whole paragraph let alone multiple.

[mdione@en.osm.town]

@michaelgemar @solonovamax @benjamineskola I'm trying to convert my suddenly boss obsessed with our overlords pushing AI down our throats to at least use AI not as process but to at least write crappy but deterministic code