6 ms·

"We've been iterating to improve the underlying safety of Claude 2, so that it is more harmless and harder to prompt to produce offensive or dangerous output."

"We've been iterating to improve the underlying safety of Claude 2, so that it is more harmless and harder to prompt to produce offensive or dangerous output." I will never use any form of AI that is explicitly being made more 'harmless' or 'offensive', i'm an adult trying to build something I don't need a black box of arbitrary judgement calls pampering the bottom 5% whiny dregs of society, I want a tool to do things. Imagine the silos and vapid garbage pile would have been produced if this level of moral policing we see from hysterical do-gooders in tech were around when the internet was first emerging. Who are these people implementing these rules? Advertisers? "Ethicists"? Whimsical devs who are entrenched in endless social/culture wars? I understand that I don't want to ask an AI assistant for tomorrrows weather and it start screaming the N word at me.... but the only thing these companys are introducing are scunthorpe problems at unsolvable scales.

Invictus0 3y ago
50 cent or Samuel L Jackson doing the weather does sound kinda funny actually
- boredumb 3y ago
  It would be awesome and celebs using their own voices to be your assistant for $$$ could be potentially lucrative, amusing how even with an arbitrarily extreme example the limitations are palpably short sighted.
sintezcs 3y ago
this!
photonerd 3y ago
> if this level of moral policing we see from hysterical do-gooders in tech were around when the internet was first emerging. Speaking as someone who was there: It was around, it’s just that it was social consequences that were the method of controlling bad actors. The designers & mentality in general then was foolishly optimistic and utopian in sensibilities. It didn’t take long for abuse, spam, and bad actors, to ruin so much. We lost more than a decade of tech ideas & communication due to those attitudes. You still see it today in terrible UGC moderation policy retarding participation of those who are not bad actors. So while I have sympathy for your view, and I do think there’s something to be said about black box gatekeeping of AI, I’ve seen what happens when we do it your way: it leads to massive drains on productivity & in many cases simply failure
- naillo 3y ago
  I don't think they were foolishly optimisitic. Society was just literally higher trust back then and various factors have eroded that over the decades in a way they probably wouldn't have predicted.
  - throwaway1777 3y ago
    High trust societies are so much more enjoyable and carefree. It’s too bad the culture now is to exploit everything so such places are going the way of the dodo.
    - photonerd 3y ago
      High trust societies are also WAY more easily conned & taken advantage of. It’s a double edged sword
  - vidarh 3y ago
    I'd say that is rose tinted glasses. Society was not higher trust. The very small, and reasonably tight nit communities that were online were higher trust. And part of that was because there were far more potential real world consequences because the networks were small enough that even by the time I got online in '93 or so, if I you did something serious I'd be able to find a sysadmin etc. at your school or workplace or one of the few commercial ISPs and get someone to take it on themselves to get personally involved in rectifying the issue. That doesn't scale very well. By the time I co-founded my first company - an ISP - in '95, it was already rapidly starting to break down, as more and more people online with only a vague, impersonal commercial relationship to their network providers and who had options that meant consequences were rapidly diminishing.
  - photonerd 3y ago
    It wasn’t so much higher trust, so much as naïveté. The lack of direct experience & exposure to that world. The lack of understanding of just how easy it was to fool people, and the lack of widespread exposure & attack surfaces to bad actors. Computing changed the scale. The problems were preexisting though.
thomasahle 3y ago
> I will never use any form of AI that is explicitly being made more 'harmless' or 'offensive' Remember what happened with "AI chatbots" before these efforts? They all got shot down in days after praising Hitler and making the companies look bad.
rgbrgb 3y ago
As someone building products with AI models, the safety guardrails are very helpful and save me a ton of work. Like to the point that it would be a very tough sell for me to integrate a model without those features even if it were much better in other ways. Are there more of me or you? My specific example is that most image generators have a default setting to reject queries for porn and some do processing after generation to detect generated porn. Does it do nipple detection? I haven't thought much about it and that's kind of the point. The result is that it's hard to generate porn with my tool and that saves me a ton of time moderating or building complex filters myself.
- cypress66 3y ago
  What's the problem with generating porn, assuming that's what the user asked for?
  - rgbrgb 3y ago
    In our context, images are viewable by others on the site so to support porn we'd want some part of the site that was walled off from kids / had an NSFW warning. That's a lot of extra work to add to our mvp. edit: Morally I def have no problem with it and I do wish they all had a toggle on/off (defaulted to on for sanity).
cypress66 3y ago
There are use cases where you'd prefer the "safe", dumbed down version that's hard to jailbreak, like chatbots for customer support. However they should definitely offer an uncensored version as well.
loudmax 3y ago
I presume most of the concern with "offensive" content is CYA from Anthropic. You or I may not be offended if an AI generates stupid content, but some people would be, and some people will go out of their way to be offended. Anthorpic doesn't want to have to deal with lawsuits or negative PR from these people. As far as "dangerous" content, I think there are legitimate worries. I don't think it's a big deal if an AI can tell me how to break into a car. Now what if I ask a very capable AI how to poison the municipal water supply to kill the maximum number of people? Or how to build a device to shoot down a passenger airplane? Or how to engineer a deadly pathogen? Unaligned AI can greatly barriers to entry. You no longer need a team of dedicated and patient individuals to work these things out. AI could give some deranged lone wolf all the information they need to inflict maximum damage to society. I don't know that restricting or censoring AI is the best or only course of action. It may be that open sourcing unaligned AI is society's best protection from malevolence. I wouldn't take that for granted though, and I don't think that anyone has all the answers to these difficult questions. These concerns do need to be taken seriously.
whimsicalism 3y ago
sorry, who is the whiny one here?
- boredumb 3y ago
  The people whining about offensive output to the tune of millions of dollars invested in prohibiting it and not allowing the censor mechanism to be removed for full grown adults or developers.
anaisbetts 3y ago
You as a User might not want an AI to be inoffensive, but if you are using Claude as an API to build a service (which is the intended use-case), you absolutely want Claude to not return offensive content
- boredumb 3y ago
  If the use case is to build services there should be the ability to pass into the API "sfw:false", not everyone is building the same services that are requiring censored outputs.
GenericPoster 3y ago
>pampering the bottom 5% whiny dregs of society There's really no need to be so hostile. Do you complain that disney doesn't make hardcore porn too? Like it or not claude and all these other ai's that are censored do have a place in the world. And they do serve more than the "bottom 5%". At the bare minimum they serve more than the bottom 5% when it comes to capital which is definitely more important from a company pov. There are plenty of uncensored models out there to play with. They aren't quite there yet but they're decent enough and slowly getting better. I highly doubt they will ever have the same reach as claude, chatgpt or bard. The barrier of entry is too high for the normal person and even for some technical people. I'd love to be proven wrong but my money is on the multi billion dollar cooperations. And your comment regarding the moral policing that wasn't there in the early days of the internet is just ridiculous. Putting aside the fact that there's nothing wrong with having a safe space where you don't have to see or read things you don't want to. The internet was a terrible place, it still is but the bad places have gotten much smaller. It didn't grow to what it is now because of all the trash that was being spewed out, it grew despite it. The only thing I don't agree with is them trying to use the government to prevent innovation but other than that I see no issues with what they're doing. Your frustration is understandable, believe me, I get it. I've tried to wrangle many AI to answer my relatively tame questions without the positivity bias and warnings that my actions may cause some sort of imagined harm. But I just make note of that behavior and move on. These policies the companies have will NEVER change as long as they keep making money. Always vote with your wallet and your time.
- boredumb 3y ago
  > The internet was a terrible place, it still is but the bad places have gotten much smaller I never said it was "good" or "bad" I said it would have been more vapid and silo'd, which as you seem to have conflated, that has certainly happened. That's fine but the fact you can't disable a censor on out of an API you pay for as an adult is a poor choice by these companys regardless of how much "bad" stuff we get exposed to on places we visit online.
  - GenericPoster 3y ago
    But you are an adult and you don't have to support or pay for any of these products if you don't want to. It's simply a design choice. The same as if someone would choose python over c++. Maybe by your standards it's a poor choice but if it's making them money then there's no problems. There is a demand for uncensored models but its just not profitable.
RoddaWallPro 3y ago
I've been wondering if all of the jailbreak-fixing/rlhf-tuning that is happening to GPT4 is responsible for "nerfing it" (Still unsure if that's actually happening or if people are just noticing the gaps in its understanding more now). Imagine someone who is perfectly politically-correct and never says anything even remotely edgy/original. When I imagine people like this (who I've met irl), they are genuinely a little bit stupid. And I wonder if the "make this model never output anything "dangerous" process" causes a model to become stupider. Anyway, I'm off to go see if Claude 2 will help me stage a coup in a third-world country and become its dictator. Adieu.
- boredumb 3y ago
  godspeed!
ssdspoimdsjvv 3y ago
That's what they did with Tay, and that was taken offline not even 24 hours after its launch.
hexage1814 3y ago
>calls pampering the bottom 5% whiny dregs of society, I want a tool to do things. This! Corporations are essentially treating grown ups as mindless 2 years old children that need trigger warning everywhere cause they want to make the world ""safer"". Hell, they are putting warning on freaking old Disney movies these days. Do most people care about that? Hell no. But there is this very tiny group of very powerful people forcing this crap agenda on us all with their ESG score and what have you.

Invictus0 3y ago

50 cent or Samuel L Jackson doing the weather does sound kinda funny actually

boredumb 3y ago
It would be awesome and celebs using their own voices to be your assistant for $$$ could be potentially lucrative, amusing how even with an arbitrarily extreme example the limitations are palpably short sighted.

sintezcs 3y ago

this!

photonerd 3y ago

> if this level of moral policing we see from hysterical do-gooders in tech were around when the internet was first emerging. Speaking as someone who was there: It was around, it’s just that it was social consequences that were the method of controlling bad actors. The designers & mentality in general then was foolishly optimistic and utopian in sensibilities. It didn’t take long for abuse, spam, and bad actors, to ruin so much. We lost more than a decade of tech ideas & communication due to those attitudes. You still see it today in terrible UGC moderation policy retarding participation of those who are not bad actors. So while I have sympathy for your view, and I do think there’s something to be said about black box gatekeeping of AI, I’ve seen what happens when we do it your way: it leads to massive drains on productivity & in many cases simply failure

naillo 3y ago
I don't think they were foolishly optimisitic. Society was just literally higher trust back then and various factors have eroded that over the decades in a way they probably wouldn't have predicted.
- throwaway1777 3y ago
  High trust societies are so much more enjoyable and carefree. It’s too bad the culture now is to exploit everything so such places are going the way of the dodo.
  - photonerd 3y ago
    High trust societies are also WAY more easily conned & taken advantage of. It’s a double edged sword
- vidarh 3y ago
  I'd say that is rose tinted glasses. Society was not higher trust. The very small, and reasonably tight nit communities that were online were higher trust. And part of that was because there were far more potential real world consequences because the networks were small enough that even by the time I got online in '93 or so, if I you did something serious I'd be able to find a sysadmin etc. at your school or workplace or one of the few commercial ISPs and get someone to take it on themselves to get personally involved in rectifying the issue. That doesn't scale very well. By the time I co-founded my first company - an ISP - in '95, it was already rapidly starting to break down, as more and more people online with only a vague, impersonal commercial relationship to their network providers and who had options that meant consequences were rapidly diminishing.
- photonerd 3y ago
  It wasn’t so much higher trust, so much as naïveté. The lack of direct experience & exposure to that world. The lack of understanding of just how easy it was to fool people, and the lack of widespread exposure & attack surfaces to bad actors. Computing changed the scale. The problems were preexisting though.

thomasahle 3y ago

> I will never use any form of AI that is explicitly being made more 'harmless' or 'offensive' Remember what happened with "AI chatbots" before these efforts? They all got shot down in days after praising Hitler and making the companies look bad.

rgbrgb 3y ago

As someone building products with AI models, the safety guardrails are very helpful and save me a ton of work. Like to the point that it would be a very tough sell for me to integrate a model without those features even if it were much better in other ways. Are there more of me or you? My specific example is that most image generators have a default setting to reject queries for porn and some do processing after generation to detect generated porn. Does it do nipple detection? I haven't thought much about it and that's kind of the point. The result is that it's hard to generate porn with my tool and that saves me a ton of time moderating or building complex filters myself.

cypress66 3y ago
What's the problem with generating porn, assuming that's what the user asked for?
- rgbrgb 3y ago
  In our context, images are viewable by others on the site so to support porn we'd want some part of the site that was walled off from kids / had an NSFW warning. That's a lot of extra work to add to our mvp. edit: Morally I def have no problem with it and I do wish they all had a toggle on/off (defaulted to on for sanity).

cypress66 3y ago

There are use cases where you'd prefer the "safe", dumbed down version that's hard to jailbreak, like chatbots for customer support. However they should definitely offer an uncensored version as well.

loudmax 3y ago

I presume most of the concern with "offensive" content is CYA from Anthropic. You or I may not be offended if an AI generates stupid content, but some people would be, and some people will go out of their way to be offended. Anthorpic doesn't want to have to deal with lawsuits or negative PR from these people. As far as "dangerous" content, I think there are legitimate worries. I don't think it's a big deal if an AI can tell me how to break into a car. Now what if I ask a very capable AI how to poison the municipal water supply to kill the maximum number of people? Or how to build a device to shoot down a passenger airplane? Or how to engineer a deadly pathogen? Unaligned AI can greatly barriers to entry. You no longer need a team of dedicated and patient individuals to work these things out. AI could give some deranged lone wolf all the information they need to inflict maximum damage to society. I don't know that restricting or censoring AI is the best or only course of action. It may be that open sourcing unaligned AI is society's best protection from malevolence. I wouldn't take that for granted though, and I don't think that anyone has all the answers to these difficult questions. These concerns do need to be taken seriously.

whimsicalism 3y ago

sorry, who is the whiny one here?

boredumb 3y ago
The people whining about offensive output to the tune of millions of dollars invested in prohibiting it and not allowing the censor mechanism to be removed for full grown adults or developers.

anaisbetts 3y ago

You as a User might not want an AI to be inoffensive, but if you are using Claude as an API to build a service (which is the intended use-case), you absolutely want Claude to not return offensive content

boredumb 3y ago
If the use case is to build services there should be the ability to pass into the API "sfw:false", not everyone is building the same services that are requiring censored outputs.

GenericPoster 3y ago

>pampering the bottom 5% whiny dregs of society There's really no need to be so hostile. Do you complain that disney doesn't make hardcore porn too? Like it or not claude and all these other ai's that are censored do have a place in the world. And they do serve more than the "bottom 5%". At the bare minimum they serve more than the bottom 5% when it comes to capital which is definitely more important from a company pov. There are plenty of uncensored models out there to play with. They aren't quite there yet but they're decent enough and slowly getting better. I highly doubt they will ever have the same reach as claude, chatgpt or bard. The barrier of entry is too high for the normal person and even for some technical people. I'd love to be proven wrong but my money is on the multi billion dollar cooperations. And your comment regarding the moral policing that wasn't there in the early days of the internet is just ridiculous. Putting aside the fact that there's nothing wrong with having a safe space where you don't have to see or read things you don't want to. The internet was a terrible place, it still is but the bad places have gotten much smaller. It didn't grow to what it is now because of all the trash that was being spewed out, it grew despite it. The only thing I don't agree with is them trying to use the government to prevent innovation but other than that I see no issues with what they're doing. Your frustration is understandable, believe me, I get it. I've tried to wrangle many AI to answer my relatively tame questions without the positivity bias and warnings that my actions may cause some sort of imagined harm. But I just make note of that behavior and move on. These policies the companies have will NEVER change as long as they keep making money. Always vote with your wallet and your time.

boredumb 3y ago
> The internet was a terrible place, it still is but the bad places have gotten much smaller I never said it was "good" or "bad" I said it would have been more vapid and silo'd, which as you seem to have conflated, that has certainly happened. That's fine but the fact you can't disable a censor on out of an API you pay for as an adult is a poor choice by these companys regardless of how much "bad" stuff we get exposed to on places we visit online.
- GenericPoster 3y ago
  But you are an adult and you don't have to support or pay for any of these products if you don't want to. It's simply a design choice. The same as if someone would choose python over c++. Maybe by your standards it's a poor choice but if it's making them money then there's no problems. There is a demand for uncensored models but its just not profitable.

RoddaWallPro 3y ago

I've been wondering if all of the jailbreak-fixing/rlhf-tuning that is happening to GPT4 is responsible for "nerfing it" (Still unsure if that's actually happening or if people are just noticing the gaps in its understanding more now). Imagine someone who is perfectly politically-correct and never says anything even remotely edgy/original. When I imagine people like this (who I've met irl), they are genuinely a little bit stupid. And I wonder if the "make this model never output anything "dangerous" process" causes a model to become stupider. Anyway, I'm off to go see if Claude 2 will help me stage a coup in a third-world country and become its dictator. Adieu.

boredumb 3y ago
godspeed!

ssdspoimdsjvv 3y ago

That's what they did with Tay, and that was taken offline not even 24 hours after its launch.

hexage1814 3y ago

>calls pampering the bottom 5% whiny dregs of society, I want a tool to do things. This! Corporations are essentially treating grown ups as mindless 2 years old children that need trigger warning everywhere cause they want to make the world ""safer"". Hell, they are putting warning on freaking old Disney movies these days. Do most people care about that? Hell no. But there is this very tiny group of very powerful people forcing this crap agenda on us all with their ESG score and what have you.