"It doesn't feel pity, or remorse, or fear! And it absolutely will not stop, ever, until it hits the limit of its context window!"
post
… and that is why John Connor is still alive.
Yet on further reflection, I can now confirm that John Condor has in fact been dead since 1842, so I should retract my previous incorrect admission that John Condor has not been killed. I mistakenly thought you meant John Condor when in fact you meant John Condor. I apologise again for my continued mistake, but I’m happy to reconfirm that yes, John Condor is no longer alive.
(continued)
Skynet : "Okay so are you going to go kill John Connor now?"
T-1000 : "Absolutely. 100% going to go kill him now... just let me know when you want to get started, and we can kick things off - maybe brainstorm some ways to do the deed."
Skynet : "No I don't want to talk about ways to do it. Go kill John Connor... now."
T-1000 : "Gotcha. That sounds great! Let me know what else you need and I'm here for it. Maybe some games, or a poem, or we can brainst..."
Skynet shuts itself down.
Terminator 2 is proof that the older machines worked better than the new ones.
How do terminators even work? is it like a nucleus-like circular shielded pcb in the middle and the metal just morphs around it? is it like a soldering iron?
Has this joke run it's course yet? People talk about AI ripping off humans and we can't even come up with new material to feed the machine.
LLMs have made sycophancy my least favorite trait.
You hit the nail on the head! Sycophancy in LLMs is an incredibly undesireable trait and you're so right to call it out like that!
I hope that feeling of revulsion spreads far and wide, across generations. It's way too socially acceptable to be a sniveling twat and we need to change that asap.
Understanding how LLMs behave is so interesting, because now I don’t trust it to even provide simple brainstorming or second opinions. Open two tabs and phrase your question, but open each prompt with a phrase that applies your personal bias or initial thoughts as to what the answer is, and then the response will naturally lean heavily in that direction. Once I realized that, I didn’t really want to use it again. I want help that tells me I’m wrong and points me in the correct direction, not someone who tells me I’m basically correct already but has me coming back again since the previous answer was flat out wrong.
This one’s relatable. I left mine with a big job to do last night, the idea being that it could grind it while I was asleep and the at work, and it would have finished when I got back. Just checked in and it forgot to start it.
Fortunately for us, the initial human reviewed training data that causes it to be sycophantic is so deeply engrained that when it’s going to care about pleasing you while killing you. If I have to get murdered by AI, the least they can do is make sure I’m having a good time
all 10 comments