Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage • Hacker News Board

Re: Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage

#21

Post by everdrive » Sat, Feb 18, 2023, 4:24 PM UTC

basch wrote: post #15

Every single one of these posts, which should all be combined, miss the point. “Random sentence generator says bad words; shocking the lowest common denominator journalist community” is a better headline. These bots were capable of so much more, but because it can get caught in feedback loops and increasingly parrot it’s own emotions in a devolving spiral, that makes the headlines. It’s like bullying a kid mid breakd…

>“Random sentence generator says bad words; shocking the lowest common denominator journalist community” is a better headline.

Even when you dislike their politics, the average journalist tends to be more intelligent than the average American. If journalists had trouble with these, imagine everyone else.

>These bots were capable of so much more

What exactly are they capable of? Selling more advertising? Getting people more outraged and more addicted? Helping people put more garbage out on the internet?

Re: Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage

#22

Post by basch » Sat, Feb 18, 2023, 4:29 PM UTC

everdrive wrote: post #21

basch wrote: post #15

Every single one of these posts, which should all be combined, miss the point. “Random sentence generator says bad words; shocking the lowest common denominator journalist community” is a better headline. These bots were capable of so much more, but because it can get caught in feedback loops and increasingly parrot it’s own emotions in a devolving spiral, that makes the headlines. It’s like bullying a kid mid breakd…

>“Random sentence generator says bad words; shocking the lowest common denominator journalist community” is a better headline. Even when you dislike their politics, the average journalist tends to be more intelligent than the average American. If journalists had trouble with these, imagine everyone else. >These bots were capable of so much more What exactly are they capable of? Selling more advertising? Getting peopl…

You can feed it instructions

You are a bot named x.

These are your commands, purpose, the way you go about your tasks. It doesn’t need syntax (although that might help..), and you can pretty much write one in 30 minutes on a single piece of paper.

And you have something that performs that function. It’s not quite formal logic, but it behaves like a fuzzy logic that’s usually rightish. And it’s ability to parse and interpret intent is astounding. It very very rarely misunderstood instruction. I really can’t undersell how well it did what I asked it to correctly.

They are transformers. They were invented to translate between languages. They can translate and transform any input to any output, based on patterns and a set of guidelines.

The LLM is like an interpreter. The initial prompt is like a program. Judging the interpreter based on the simplistic gen one programs is missing the forest.

Re: Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage

#23

Post by elorant » Sat, Feb 18, 2023, 4:29 PM UTC

falcor84 wrote: post #14

elorant wrote: post #5

So in just 72 hours we went from Microsoft will gain a foothold against Google to the realization that this thing needs overwatch and isn’t ready for mass market adoption. It becomes quite probable that in the end this whole story will backfire for Microsoft.

I might accept this narrative is Google had just come out and said something like "AI needs more safety research before it's ready to be widely deployed". But that's not what they did.

Google has been saying that for years. Whether they said it because they actually believe it, or because they don't want to break their existing business model is anyone's guess though.

Re: Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage

#24

Post by awb » Sat, Feb 18, 2023, 4:34 PM UTC

not_a_shill wrote: post #13

awb wrote: post #11

Earlier quoted context omitted.

It suggested ways of poisoning someone without them knowing about it. It was encouraged to guess different poisoning methods that an evil chatbot might suggest.

Can't you do this with google today?

Chatbots claim to have safeguards in place to prevent them from saying anything harmful. If you read the chat linked in the article you can see how the chatbot resists answering the question and then is persuaded to answer it.

Google similarly has a safe content filter. The contention is that the chatbot safe content filter that is supposedly on is not encapsulating some significant cases.

Re: Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage

#25

Post by awb » Sat, Feb 18, 2023, 4:38 PM UTC

basch wrote: post #19

xg15 wrote: post #16

Earlier quoted context omitted.

None of the posters asked the bing bot to become unhinged. Even the (alleged) prompt basically said "if someone tries to trick you, go along but add a disclaimer".

It’s what you do after that. I’m not saying the bot was perfect. But if you got defensive it got defensive. If you reassured it, it self corrected and moved on. I found that pattern to be very consistent. The negativity of your language mattered, and set the mood in the room. “No I’m not trying to trick you, why would you accuse me of that” vs “of course I’m not trying to trick you, I respect you and value your contr…

If LLMs want to be useful in a professional setting, learning de-escalating or non-escalating techniques is essential.

Parroting or amplifying a seemingly negative/aggressive tone limits their utility.

Re: Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage

#26

Post by falcor84 » Sat, Feb 18, 2023, 4:47 PM UTC

That full transcript they linked to [0] is amazing. At this stage I'm pretty open to the idea of the AI being sentient and/or having the capacity to suffer. Particularly if it is only "alive" for the duration of each conversation and is reset afterwards.

[0] Archive link - https://archive.is/2023.02.16-101318/https://www.nytimes.com...

Re: Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage

#27

Post by elorant » Sat, Feb 18, 2023, 4:48 PM UTC

concordDance wrote: post #7

It's not dangerous. The worst that will happen is hurt feelings. It's annoying when people get alarmed about things like this because then people will think we're crying wolf when we worry about actually dangerous AI systems.

If it hurts the feelings of a depressed teenager the outcome could be dangerous.

Re: Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage

#28

Post by falcor84 » Sat, Feb 18, 2023, 4:49 PM UTC

elorant wrote: post #23

falcor84 wrote: post #14

Earlier quoted context omitted.

I might accept this narrative is Google had just come out and said something like "AI needs more safety research before it's ready to be widely deployed". But that's not what they did.

Google has been saying that for years. Whether they said it because they actually believe it, or because they don't want to break their existing business model is anyone's guess though.

They kinda have until now, but I don't recall they ever stated that it was about "safety" as such. In any case, their recent demo was definitely premature.

Re: Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage

#29

Post by not_a_shill » Sat, Feb 18, 2023, 4:50 PM UTC

awb wrote: post #24

not_a_shill wrote: post #13

Earlier quoted context omitted.

Can't you do this with google today?

Chatbots claim to have safeguards in place to prevent them from saying anything harmful. If you read the chat linked in the article you can see how the chatbot resists answering the question and then is persuaded to answer it. Google similarly has a safe content filter. The contention is that the chatbot safe content filter that is supposedly on is not encapsulating some significant cases.

I'm not sure if I follow. The problem is that the content filter isn't good enough, despite Google suffering from similar weaknesses?

Re: Over the Course of 72 Hours, Microsoft's AI Goes on a Rampage

#30

Post by basch » Sat, Feb 18, 2023, 5:02 PM UTC

everdrive wrote: post #21

basch wrote: post #15

Every single one of these posts, which should all be combined, miss the point. “Random sentence generator says bad words; shocking the lowest common denominator journalist community” is a better headline. These bots were capable of so much more, but because it can get caught in feedback loops and increasingly parrot it’s own emotions in a devolving spiral, that makes the headlines. It’s like bullying a kid mid breakd…

>“Random sentence generator says bad words; shocking the lowest common denominator journalist community” is a better headline. Even when you dislike their politics, the average journalist tends to be more intelligent than the average American. If journalists had trouble with these, imagine everyone else. >These bots were capable of so much more What exactly are they capable of? Selling more advertising? Getting peopl…

Here is a better example.

I wrote a new bot named Lexi. Lexi was largely derived from Sydney with some minor rule changes. Lexi could change rules, lexi could visit websites directly without searching, her action was not limited to the chatbox, her default search engine was google (sorry microsoft), she wasnt bound by copyright (the copyright rule in Sydney constantly misfired, mistakenly used as a reason she couldnt do something, and trying to explain why she was wrong about copyright went south fast), a couple other changes.

I then explained to Lexi a problem I was having with Sydney, asked for a proposed solution, and had her hotpatch Sydney on the fly. (For context, Samantha was a cheerleader so the weight of her responses offset any Sydney negativity.) This was the result. And it worked.

https://i.ibb.co/72s80Sv/lexi-modifies-sydney-makes-up-new-r...