Amy Dentata

@amydentata

she/her

Who knows.
—
#really real tarot

Mastodon

tech.lgbt/@amydentata

Amy Dentata @amydentata7/6/2023, 4:58 AM

Vicas @Vicas

Vicas @Vicas7/6/2023, 3:43 AM

Really good article on AI chatbots and psychics: https://softwarecrisis.dev/letters/llmentalist/

I especially like the parts about "smart" people potentially becoming even bigger evangelists when they do get fooled by it, because boy howdy does that ring true

#The Cohost Global Feed

Amy Dentata @amydentata7/6/2023, 4:58 AM

For the past couple of months, I’ve been working on an idea that I think explains the mechanism of this intelligence illusion.

I now believe that there is even less intelligence and reasoning in these LLMs than I thought before.

Many of the proposed use cases now look like borderline fraudulent pseudoscience to me.

Finally people are starting to get it

You must log in to comment.

in reply to @Vicas's post:

wuest @wuest7/6/2023, 11:43 AM

One thing that I don't know, which I would like to know is whether being aware of the mechanics of the psychic con dissuades people from their belief.

It seems to matter very little to people who are suckered by convincing text generation - which I recall learning was also true when people were getting Too Into ELIZA.

Vicas @Vicas7/6/2023, 12:22 PM

Definitely a good question. I could absolutely see someone who's aware of what cold reading is get thrown completely off-kilter by a read that hits a little too close to home and kinda spiral from there, which feels very similar to programmers being shocked that it can refrigerate basic python, but neither of those are empirical

Maersk Aerospace @Maersk-Aerospace7/6/2023, 1:31 PM

my wager, and this is totally speculation, is that this is mostly an indicator that people don't really truly understand the underlying mechanisms, which especially in LLM type things is very common, since it's functionally an entire vocation unto itself. it's much easier to comprehend "the psychic says stuff that's probably true and waits for me to flinch" than it is to comprehend a token cloud and the inner workings of the model. and i think there's a lot of smoke & mirrors in the software industry about what people actually understand vs what they know how to use.