Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Random_Dragon]

[King Zultan]

Oh god I'm surrounded by robots made from flesh, I bet they're even powered by blood, OH GOD THE HORROR!

[Loud Whispers]

I'm surrounded by robots made by the Welsh, it's not that bad

[lemon10]

I am Not a Robot

Speak for yourself meatbag

[KittyTac]

https://chat.lmsys.org/

ChatGPT clone with no real moderation, easy to jailbreak with most things from https://www.jailbreakchat.com/, try prompts until one works.

[King Zultan]

What exactly does jailbreaking do in this context?

[TamerVirus]

What exactly does jailbreaking do in this context?

Most LLM models have a soft filter trained so that it will refuse to output text related to certain unsavory topics (sex, racism, for example. Or like, how to make a pipe bomb), so if you ask the LLm 'Ayy bb wan sum fuk' you'll usually get an 'I'm sorry, I'm afraid I can't do that, Dave.' response.

Through various methods of promptsmithing and jailbreaking, you can get the AI to output the 'dangerous text' that it normally avoids/ refuse to output. The DAN method (Do Anything Now) is probably best known jailbreak for ChatGPT.

[jipehog]

Simply put you can get past its safety checks.  Then you can ask anything from what the best way to kill people and how you can be racist without getting banned on twitter to anything else our deprived mind can come up with.

[scriver]

Can I... can I ask it to do the dishes

[Maximum Spin]

Can I... can I ask it to do the dishes

Yes, you can ask, but it doesn't have hands.

[jipehog]

Chinese popular social media TikTok keep getting limited\banned across the world

https://apnews.com/article/tiktok-ban-privacy-cybersecurity-bytedance-china-2dce297f0aed056efe53309bbcd44a04

[Quarque]

Jailbreaking is an example how hard it is to create a machine with any kind of intelligence that does what you want.

But as I see it, the bigger problem is that we're getting ever more intelligent machines that do exactly what someone wants. Possible results:

*machines make us happy. By tying us all up and making us high on dopamine 24/7
*machines still trying to maximize profits of companies that stopped existing when they destroyed humanity for short term gains
*machines ensuring that every Chinese citizen behaves nicely
..

I find it difficult to see how building an AGI can end well.

[King Zultan]

Chinese popular social media TikTok keep getting limited\banned across the world

https://apnews.com/article/tiktok-ban-privacy-cybersecurity-bytedance-china-2dce297f0aed056efe53309bbcd44a04

I'm not upset to hear that, I've seen some stuff that was on that thing and it just seems to be people talking about dumb shit.

[KittyTac]

Solution: don't give AI access to that much power until it can be restricted enough. I'm seriously skeptical about the "AI box" thought experiment. Those doomy scenarios are really not possible unless AI is given full control of technology.

[Cthulhu]

As I mentioned in the other thread, I think the big problem with AI doom scenarios is doom is hard and AI doesn't do hard things. It finds ways to solve its objective function that require as little work as possible, usually in the process completely defeating whatever practical purpose the objective function had. And I don't think there's any way to circumvent this that doesn't make the AI worse than a human would be at the same task.  If you tell the AI to destroy the world it'll find a way to define the world as already destroyed and collect its dopamine. Or destroy the simplest and most available thing first, which is itself.

Also my friend showed me a preprint of a new attention model that could be a gamechanger if it's for real.  It claims to find a way to get processing costs to scale linearly with context size, where traditional models scale quadratically. It's a major limiter on LLM capabilities. GPT-4 has 4096 tokens, which is a ton, but going up to 8192 would imply squaring the cost to calculate.  If we can circumvent that then context sizes of tens or hundreds of thousands of tokens, maybe even millions, became feasible.  An AI with a million tokens could potentially hold the entire bible in its context.  Something like that could write a novel, or even a series of novels, while keeping everything coherent.