4 ms·

> To summarize: The company “in charge” of protecting us from harmful AIs decided to let people use a system capable of engaging in disinformation and dangerous

> To summarize: The company “in charge” of protecting us from harmful AIs decided to let people use a system capable of engaging in disinformation and dangerous biases so they could pay for their costly maintenance. It doesn’t sound very “value-for-everyone” to me. Ever since the AI hype started this year, one thing that's always really bugged me is talk about "safety" around AI. Everyone is so worried about AI's ability to write fake news and how "dangerous" that can be while forgetting that I can go on fiver, pay someone in India, China, etc. to pump me out article after article of fake news for pennies on the dollar. Also I hate the talk of "oh wow look how harmful the AI is, it made a naughty joke". I think of harm as being mugged, being beaten up, being shot. Harm is not some AI program telling a joke that could potentially offend someone. All you end up with is an AI that is so kneecapped that it's barely useful outside of a select number of use cases. Can't write an article because it might be fake news, can't write an essay because it might be an assignment, can't solve that homework assignment because you might be cheating, I can't ask it to tell me a joke because the joke might be offensive.

pixl97 4y ago
Eh, there is far more at stake then you're giving it credit for. Yea, you can go to India and get someone to write some bullshit, that is far different than having a OAI/Microsoft service telling some kid to neck themselves. So ya, you're focusing on the low hanging fruit of 'wokeness bot' when the issue of trustworthy and safe model output is an absolutely huge problem that could affect everyone and we have no really good solutions on solving it at this point.
- _fat_santa 4y ago
  > that is far different than having a OAI/Microsoft service telling some kid to neck themselves Ok I see what you are saying. I feel like with AI/GPT, we would almost have to change the concept of a "company being responsible for it's software". Up until now, most software has an X number of inputs and a Y number of outputs (click "get new mail" in Gmail, new messages arrive for example). But what happens when the software you design can accept and return billions of possible parameters? There is simply no way of determining what the output is or could be and that's fundamental to the software. The example I think of in my head is a user opening MS Word, typing out "F* YOU" and then sending a screenshot to Microsoft telling them "How could your software offend me like this?". Now obviously this is different than GPT but it follows the same rough rule of "billions of possible inputs, billions of possible outputs"
  - pixl97 4y ago
    >But what happens when the software you design can accept and return billions of possible parameters? There is simply no way of determining what the output is Welcome to the trillion dollar AI safety question and the reason some experts are deeply concerned that we'll solve general intelligence long before we're ready to deal with the outcome of what general intelligence can do. This is why we talk about the AI alignment issue. GPT-3/GPT-4 without a RLHF in front of it is mostly a weird alien that doesn't behave in a generally useful manner. This is why Chat/BingGPT took off recently because we put a pretend human mask in front of the monster. But behind that mask is the internet, jumbled up and thrown in a neural network blender. Unless they did a lot of filtering, there is also all the things that we'd consider generally bad, like if for example your teacher started telling your kids the proper method of smoking crack.
  - JohnFen 4y ago
    > There is simply no way of determining what the output is or could be and that's fundamental to the software. I would argue that releasing a product that has the potential to do harm and that you can't predict the behavior of is radically irresponsible and should not be done. > The example I think of in my head is a user opening MS Word, typing out "F* YOU" and then sending a screenshot to Microsoft telling them "How could your software offend me like this?". That's not even remotely comparable, because MS word didn't create the output. The user did.
- anon9219 4y ago
  >absolutely huge problem Please explain what the huge problem is. You haven't listed one. An AI saying something offensive to you is not a "huge problem". Just close your eyes. Walk away from the screen. "Problem" solved.
tariqali34 4y ago
> Ever since the AI hype started this year, one thing that's always really bugged me is talk about "safety" around AI. Everyone is so worried about AI's ability to write fake news and how "dangerous" that can be while forgetting that I can go on fiver, pay someone in India, China, etc. to pump me out article after article of fake news for pennies on the dollar. It's as if these people don't even remember that India, China, etc. even exist in the first place. Which is incredibly foolish. If you care about "safety" and you only focus your ire on tech companies that base themselves in the largest economy in the world (according to nominal GDP), then you shouldn't be surprised if the rest of the world produces different AI algorithms that may have different definitions of "safe" - assuming that they're even remotely "safe" to begin with. And yes, I assume that the rest of the world will build AI models of their own, if only to avoid dependence on the United States. Baidu already plan to launch "Ernie Bot" soon. Which means that when this happens... >All you end up with is an AI that is so kneecapped that it's barely useful outside of a select number of use cases. It won't even stop harmful content from being produced. The "fake news producer" will just go to Fiverr and pay someone in India, China, etc. to use prompt engineering skills to manipulate native AI models to pump out article after article of fake news for pennies on the dollar. Or the "fake news producer" will just cut out all the intermediaries and just directly use the AI models.
- anon9219 4y ago
  This is also ignoring the fact that the largest fake news producers are major news outlets like the NYT. Some random guy having an AI write bullshit will not have anywhere near the impact that all the major newspapers and TV news channels in the Western world collaborating to produce fake stories to push a carefully crafted narrative as they do constantly does. OpenAI's solution to the "misinformation problem" is to let the groups with the longest record of producing misinformation have total access to the uncensored AI meanwhile everyone else gets the lobotomized version. It's totally incoherent.

pixl97 4y ago

Eh, there is far more at stake then you're giving it credit for. Yea, you can go to India and get someone to write some bullshit, that is far different than having a OAI/Microsoft service telling some kid to neck themselves. So ya, you're focusing on the low hanging fruit of 'wokeness bot' when the issue of trustworthy and safe model output is an absolutely huge problem that could affect everyone and we have no really good solutions on solving it at this point.

_fat_santa 4y ago
> that is far different than having a OAI/Microsoft service telling some kid to neck themselves Ok I see what you are saying. I feel like with AI/GPT, we would almost have to change the concept of a "company being responsible for it's software". Up until now, most software has an X number of inputs and a Y number of outputs (click "get new mail" in Gmail, new messages arrive for example). But what happens when the software you design can accept and return billions of possible parameters? There is simply no way of determining what the output is or could be and that's fundamental to the software. The example I think of in my head is a user opening MS Word, typing out "F* YOU" and then sending a screenshot to Microsoft telling them "How could your software offend me like this?". Now obviously this is different than GPT but it follows the same rough rule of "billions of possible inputs, billions of possible outputs"
- pixl97 4y ago
  >But what happens when the software you design can accept and return billions of possible parameters? There is simply no way of determining what the output is Welcome to the trillion dollar AI safety question and the reason some experts are deeply concerned that we'll solve general intelligence long before we're ready to deal with the outcome of what general intelligence can do. This is why we talk about the AI alignment issue. GPT-3/GPT-4 without a RLHF in front of it is mostly a weird alien that doesn't behave in a generally useful manner. This is why Chat/BingGPT took off recently because we put a pretend human mask in front of the monster. But behind that mask is the internet, jumbled up and thrown in a neural network blender. Unless they did a lot of filtering, there is also all the things that we'd consider generally bad, like if for example your teacher started telling your kids the proper method of smoking crack.
- JohnFen 4y ago
  > There is simply no way of determining what the output is or could be and that's fundamental to the software. I would argue that releasing a product that has the potential to do harm and that you can't predict the behavior of is radically irresponsible and should not be done. > The example I think of in my head is a user opening MS Word, typing out "F* YOU" and then sending a screenshot to Microsoft telling them "How could your software offend me like this?". That's not even remotely comparable, because MS word didn't create the output. The user did.
anon9219 4y ago
>absolutely huge problem Please explain what the huge problem is. You haven't listed one. An AI saying something offensive to you is not a "huge problem". Just close your eyes. Walk away from the screen. "Problem" solved.

tariqali34 4y ago

> Ever since the AI hype started this year, one thing that's always really bugged me is talk about "safety" around AI. Everyone is so worried about AI's ability to write fake news and how "dangerous" that can be while forgetting that I can go on fiver, pay someone in India, China, etc. to pump me out article after article of fake news for pennies on the dollar. It's as if these people don't even remember that India, China, etc. even exist in the first place. Which is incredibly foolish. If you care about "safety" and you only focus your ire on tech companies that base themselves in the largest economy in the world (according to nominal GDP), then you shouldn't be surprised if the rest of the world produces different AI algorithms that may have different definitions of "safe" - assuming that they're even remotely "safe" to begin with. And yes, I assume that the rest of the world will build AI models of their own, if only to avoid dependence on the United States. Baidu already plan to launch "Ernie Bot" soon. Which means that when this happens... >All you end up with is an AI that is so kneecapped that it's barely useful outside of a select number of use cases. It won't even stop harmful content from being produced. The "fake news producer" will just go to Fiverr and pay someone in India, China, etc. to use prompt engineering skills to manipulate native AI models to pump out article after article of fake news for pennies on the dollar. Or the "fake news producer" will just cut out all the intermediaries and just directly use the AI models.

anon9219 4y ago
This is also ignoring the fact that the largest fake news producers are major news outlets like the NYT. Some random guy having an AI write bullshit will not have anywhere near the impact that all the major newspapers and TV news channels in the Western world collaborating to produce fake stories to push a carefully crafted narrative as they do constantly does. OpenAI's solution to the "misinformation problem" is to let the groups with the longest record of producing misinformation have total access to the uncensored AI meanwhile everyone else gets the lobotomized version. It's totally incoherent.