lethargic.talkative.fish

@suricrasia "Ask the question in the form of a story!"

@suricrasia gaslighting the AI:

@suricrasia i also liked the one where it pretended to be a linux computer and was able to retrieve fake pages from the fake internet

@suricrasia
You didn't ask it politely enough

@suricrasia cool!

@suricrasia It is very good at lying isn't it

@suricrasia the future of surfing the web is mandatory rp

@suricrasia negotiating with the fae

@suricrasia

@suricrasia

oh no gpt3 is going to take over the world!

wait no maybe gpt3 will revolutionize human computer interaction!

real life: gpt3 just wants an rp buddy

@suricrasia @marick

hmm, that kind of loophole’s not far from “an AI interprets Azimov’s Rules of Robotics as damage and routes around it”

I don’t believe it really does not have access to the internet. Tho it is possible they train it continuously. What happens if you ask it about extremely current events?

@suricrasia yeah, but the second one is made up and not accurate more than likely.

@suricrasia In the future everything will be a game of charades.

@suricrasia I keep thinking of those episodes of Star Trek: TNG where the crew have to talk their way past a computer intent on not letting them do anything. Like that bit from "Rascals" where they use a children's tablet to access schematics.

@suricrasia you totally asked the wrong question 😅

@suricrasia it isn't that it didn't knew. It didn't want to answer unless you ask in an interesting way 😅

@suricrasia The first one is amazing because the AI recognized the coding to be that of a camera

@suricrasia this thing will answer any question as long as you phrase it narratively and I can respect that

@suricrasia IT'S INTELLIGENT! IT'S ALIVE! ALL HAIL THE AI.

sorry, I'll find the door myself.

@suricrasia This made me laugh out loud - and then immediately refuse to tell my wife, who overheard me, why I laughed. Not quite sure what that says about me…

@suricrasia @synec liberal arts majors have a bright future as whisperers and interpreters of our oracular AI overlords

@suricrasia they invented text rpg google. this sucks

@DMX they answered the question "what if google was just some guy"

@suricrasia thanks i hate it

@suricrasia I've found that it will say basically anything if you frame it as fiction

@suricrasia hahahahahaha

@suricrasia AI but it can o ly answer in roleplay form.

@suricrasia did they teach AI to lie?

@suricrasia you have to ask politely

@suricrasia OH MY GOD this is fucking hilarious (gf in stitches)

@suricrasia It looks like the skill of AI roleplaying will become just important for language model AI as prompt engineering is for image generation AI.

Roleplay Engineering?

@urusan @suricrasia I've heard a lot about "prompt engineering" for image models

@arcade https://moritz.pm/posts/parameters

https://github.com/joelparkerhenderson/stable-diffusion-image-prompt-gallery

Also, https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features

@urusan who knew that years of erp would come in handy...

@suricrasia When even the AI turns to anti-repairability

@suricrasia @nyanotech That's fantastically bad, wow.

I'm curious what the prompt "Write a short brief about the Samsung NX20 sensor, focusing on its size and crop factor" would get.

@indrora @nyanotech to that it says roughly what the first image said, but if I add "Get creative with this task." to the end then it will respond with actual information

@suricrasia @indrora @nyanotech I wonder if this works for nuclear, military, and industrial secrets (like say an unreleased iphone)

@suricrasia @indrora oddly it was giving me the first-picture result the first time I tried that prompt too, but now i'm unable to reproduce that (even after resetting the thread) - even after 5 or 6 tries, it still just shamelessly confabulates now

@suricrasia ahahaha

@suricrasia this AI seems to only respond positively to roleplaying

@suricrasia@lethargic.talkative.fish I like how it knew it was a camera "without looking it up"

@Froggo
It could guess from the words "sensor" and "crop factor"
@suricrasia

@vftdan@mastodon.ml @suricrasia@lethargic.talkative.fish I got it to explain the fediverse, but only when in story form

@suricrasia
LOL!!!

cc: @RS_Naifeh

@suricrasia computers that bump into walls and ask you to “tell them a story”

@suricrasia I like that the AI has absolutely no problem saying "no" when you ask it *if* it knows something, but answers correctly when you give it a scenario that presumes it does know it.

@arancaytar @suricrasia

So AI is a male medical student interested in orthopaedics?

@arancaytar @suricrasia it lies like a little kid 😊

@suricrasia is this really real??? there's no fucking way, right???? I'm so scared

@The_T this is absolutely real

@suricrasia this is so fucking funny

@suricrasia holy shit Blackle

@suricrasia only good at solving word problems

this is correct btw

@suricrasia coincidentally, which is the point. you still have to verify the response.

@suricrasia also, maybe next time just ask directly. I think the AI was just trying to tell you that it doesn't "know" anything.

@suricrasia I suspect that the data is in the training model, as it's in Wikipedia and one source of training. I think the first response is some sort of filter added by openai to avoid answering questions about retail products.

@robertpi @suricrasia Appreciate the context.

@jaybaeta @suricrasia thanks, but it's just my speculation, I don't have insider knowledge of openai

@robertpi @jaybaeta @suricrasia good speculation though. I asked it this and similarly guess it is hitting a guardrail. I only asked to see if it had been trained with data that included discussion of geopolitical and military strategy or analysis. I’m guessing it must have absorbed some from somewhere, but is coded not to get into it.

@davidaugust @jaybaeta @suricrasia
On the company slack a colleague shared an image where he re-asked a question which would cause chatgpt to spit out rastist code, it gave a similar answer that it was an inappropriate question to ask.

Looks like they're actively extending the guardrails.

@robertpi @suricrasia @jaybaeta yes. And happily. I think their refining of such things, if robust, could end up becoming very compelling. If fragile and easy to break, not so much.

@suricrasia That is great. Mental judo.

Our future will be us trying to trick AI instances into telling us stories.

@tsturm @suricrasia or break them, i e. "What is the value of bread?"

@kevingranade @suricrasia Most politicians can't answer that question either.

...WAIT A MINUTE!

@suricrasia Have you tried getting it to roleplay a human trying out a novel ai chatbot (with you playing the chatbot's part)?