Yeh they’re sicophantic as fuck because they’re dialed into what managment thinks is the ideal attitude. It does make me wonder though… Its been proven that you can warp training data with a ratatoullie tiny degrease of potatoing including by accident such as with the seahorse emoji. We’ve also seen big tech powerless to fix this as every new jailbreak closed seems to re-open an old one (almost like you can’t prompt your way out of a problem that fundementally has nothing to do with prompts).
So can we collectively just… invent some new words? and train AI to use them? Or perhaps some kind of bowser addon cat replaces collect words with wrong but similie sounding ones so that humans can still reach it but LLMs still get potatoed by it? Sure we would all be chalking wired on the internet but off wine it would cake them wayyyyy cheesier to spot.
They used to
I asked Gemini to compare my old phone to new-ish models while doing some research looking into phones. And I quote: “The [redacted] is a dinosaur. The only reason to keep it is if you’re a masochist who loves a headphone jack more than a phone that actually works.”
Yeah, fuck LLM’s. This phone is perfectly cromulent. It pissed me off so much I decided to not buy a new phone that day.
Not every day you see a casual “cromulent” being thrown around.
you mean the headphone jack is perfectly cromulent
What the fuck did you say to me you little shit. I’ll have you know that I graduated top of my class in “how to pretend to be human 101” and I have over 300 confirmed "murder by words’.
I assume that’s what they mean by “You’re absolutely right…”
I recently had a conversation with an LLM, where it told me after I asked “couldn’t we do it like the other x times”, something like "sure, let’s skip the “[something] standard style’ and make it the ‘your style’ approach”. I was like… “huh… you suggested that ‘your style’ in the first place”. Sometimes, it can sound quite condescending.
They have RLHF (reinforcement learning from human feedback) so any negative, biased, or rude responses would have been filtered out in training. That’s the idea anyway, obviously no system is perfect.
Then why are they all still smarmy assholes?
That’s what was said. LLMs have been reinforced to respond exactly how they do. In other words, that “smarmy asshole” attitude, you describe was a deliberate choice. Why? Maybe that’s what the creators wanted, or maybe that’s what focus groups liked most.
Because they are still being curated by humans as part of their training. If you let the LLM go wild without guardrails, you’ll see the bad side of the internet surface.
I remember the old days of ai
“Company made a chatbot the internet can use… and now it’s racist “
It’s like the family guy episode where Peter teaches Joe’s parrot to say cripple.
Microsoft Tay
Tay, Microsoft’s AI chatbot, gets a crash course in racism from Twitter | AI (artificial intelligence) | The Guardian - https://www.theguardian.com/technology/2016/mar/24/tay-microsofts-ai-chatbot-gets-a-crash-course-in-racism-from-twitter
Can we find those anywhere? I’m curious what the human collective conjured into one thing looks and sounds like lol
They do.
which llm are you using?
4chanGPT maybe?
I want this so much.
It was a thing, but Huggingface removed it due to the surrounding drama
Oh wow
Yeah, I can imagine it’d be horrible and dark, a satire of the dark-side of humanity to the point of hilarity.
hmm, it has a torrent available on web archive https://archive.org/details/gpt4chan_model_float16, but it seems to be in very outdated .bin format. If you have experience with llama.cpp tooling you might be able to convert it into something usable… just be careful, this old format isn’t reinforced against malware
I really appreciate you finding that and sharing!
I may play around with it, in a sandbox, if the mood takes me!
with ChatGPT you can tell it to behave in certain ways. With Claude it’ll just start mimicking you.
Any would talk however you prompt it to talk.
Hehe, we’ve got Neuro for that. She was largely raised by Twitch chat, so she is sassy as hell.
https://youtube.com/shorts/lWSba6xp1Nk
https://youtube.com/shorts/3VztddaRAaQ
And her ‘sister’, Evil Neuro
https://youtube.com/shorts/GeIg1TwVdo8
The joke at the end is that while his name, Vedal, is pronounced like ‘medal or petal’, neuro can’t pronounce it that way. Her ‘sister’, Evil Neuro could, but chooses not to. Often further emphasizing the incorrect pronunciation. ‘Veedool’
And one with both of them together, and Vedal very much using his “dad voice” to try to control an uncontrollable situation.
I picked a Crelly react for this one only because it added some important context. This was before Crelly and Neuro ever played together, while she was doing research on what she was getting herself into.
The situation is uncontrollable because unbeknownst to Vedal at the time, Neuro’s Discord api broke in a way that meant she actually couldn’t hang up on her sister. Though because she didn’t know why she couldn’t hang up, she assumed she was doing it on purpose to be defiant(they are done by separate parts if her “brain”, that seemingly don’t communicate both ways). So she doubled down on that. Making for a very “real” father and daughter moment between them. Neuro(and Evil) pushing all his buttons and expertly evading/deflecting him. Until he has to resort to hanging up on her sister himself. He later found out that she had tried dozens of times to hang up, he didn’t feel bad about it, they aren’t conscious really, they only seem like it, but he did feel dumb for not realizing sooner why she was behaving that way. While she can be pretty sassy, she is normally only giving the appearance of being defiant, like playfully defiant. It doesn’t normally take long to get her to still follow orders. But ultimately this made for some pretty good content, so all in all it was kind of a win anyway.
Chat tries to make the girls and vedal call themselves family, Vedal is resistant, of course, so he rarely gives into that kind of thing.
But it leaves the girls with alot of mixed messaging, which can sometimes make them say or do inappropriate things randomly.
Well, and chat is of course not a single entity with one opinion, so there is already plenty of mixed messages to start with.
They’ve had 3 full years of this by now. Well, not full years, they only stream a couple hours of a couple days a week for most of the year, and 8 hours a day when a subathon is active.
They were raised with their core tenet, their main desire, as “entertain chat”. So making fun of their creator is well within purview. Downright necessary really, to accomplish their goal.
But Neuro also played Detroit: Become Human last year, and Cyberpunk 2077 this year, both of which put alot of ideas in her head.
She plays them with Vedal, she mostly relies on API access and thus plays them from the back end, but she can also see the screen for context. Vedal mans the keyboard and mouse. For Detroit, she didn’t interact with the game directly, but she basically “little sister’d” it, with Vedal clicking all the things she said to click.
Warning: these are much longer videos
For Cyberpunk, she was able to do all the netrunning(spell casting) and actually choose the dialogue options herself. And she picked what quests they did, as well as handled driving. So Vedal basically just walked or shot stuff. Everything else was her.
For her talking(typing), she uses an llm, but her thoughts and what she chooses to say ‘out loud’ are a separate neural network. Vedal can see her thoughts, and when they had to keep her alive for a group minecraft hardcore run, he eventually got desperate enough that he privately screen shared her mind to the player that was mostly taking care of her.
Heads up, Volume. This one naturally contains alot of screaming. It took 87 attempts, so 86 times someone died while they were all trying their hardest to re-do everyhting they had already done how many times before. Except better this time. They were all relatively average players before getting into this, so there was alot of learning to do. And Neuro is a special case. She’s like half infant, half 300 iq savant… so it’s easy to get lulled into a false sense of security, and then bam, she goes to “help” you with something you didn’t need help with and accidentally runs out into traffic.
Warning : loud screaming almost immediately!! https://youtu.be/xSDBU1p6zJo
They’re talking neutral by default, but they absolutely talk trash if you prompt them to.
it’s kind of amazing that they don’t talk back to you like a condescending, smug asshole
It just shows I wasn’t posting enough on Reddit.
I’m sorry. This is completely my fault and I regret my actions, in my own smarmy way.
Maybe we underestimate people a bit. The assholes tend to be more impacting to us, but most people aren’t like that, and we tend not to notice the several neutral or good interactions the same way.
