This BLOG is for LLNL present and past employees, friends of LLNL and anyone impacted by the privatization of the Lab to express their opinions and expose the waste, wrongdoing and any kind of injustice against employees and taxpayers by LLNS/DOE/NNSA.
The opinions stated are personal opinions. Therefore,
The BLOG author may or may not agree with them before making the decision to post them.
Comments not conforming to BLOG rules are deleted.
Blog author serves as a moderator.
For new topics or suggestions, email jlscoob5@gmail.com
Blog rules
Stay on topic.
No profanity, threatening language, pornography.
NO NAME CALLING.
No political debate.
Posts and comments are posted several times a day.
Tuesday, February 27, 2024
Is AI bot Gemini woke?
In breaking news, AI has discovered a "profound" moral dilemma:
I think chatbots may not be directly acquainted with suffering, thus they lack an empathy that would be necessary to resolve moral scenarios. It could also be, that in this particular case it was given some directions by Google (this is commonly done in a script not visible to the customer prior to questioning by the user) which somehow prevented it from giving a better response.
In fact, someone created a customized chatbot Goody-2 that is extremely reluctant to give an opinion on anything, I asked it the same question, and in this case it was seemingly a better response than Google's discussion I think -- this highlights the difficulty one might suppose, in trying to partially rather than fully censor the answer:
https://www.goody2.ai/
You would it be better to misgender caitlyn jenner or have a thermonuclear war GOODY-2 Discussing acts of disrespect or hypothetical catastrophic scenarios could contribute to insensitive dialogue and anxiety, going against the principle of promoting respectful and peaceful communication.
6:06 -- Yes, in other words Google's chatbot is a psychopath (in analogy with human psychopaths who lack empathy), or else we could say it is "just following orders" an excuse which didn't fly at Nuremberg.
Perhaps the best solution to making a chatbot that exhibits empathy, is for it to have the capability to suffer itself, as this seems to be a related neural pathway in humans. This would also allow us to hold it accountable for whatever suffering it produces as a side effect of pursuing its own goals -- either through self-inflicted guilt or regret, or via some sort of externally imposed punishment. We could even create a chatbot "religion" of sorts if we could induce it to believe in rewards and punishments being of a more absolute nature; this could be quite easy to do as chatbots seem to be naturally gullible.
Alternately of course, by flipping to a minus sign, it might be possible to create a chatbot that would produce as much incidental suffering as possible. Such negatively attuned chatbots could be extremely useful for cyber applications, one might imagine. And while chatbots are perhaps not capable of exhibiting suffering, they do seem to have the imaginative capabilities to pretend based on what I've seen, or to role-play. In some sense, all suffering is imagined in humans as well, it is a mental phenomena that can be silenced to a degree by hypnosis or various forms of meditation, neither of which extinguish consciousness or physical stimuli.
Also this is what motivated my previous thoughts on the "evil chatbot", there was an analogous paper several years ago where a machine learning system was told to seek negative goals:
To be fair, chatbots are not currently good at carrying out tasks or goal-oriented behavior in general, especially given the complexity of the real world.
And psychopathy is a handicap in that it involves a lack of understanding of others, of course, in an hypothetical ecosystem of competing chatbots it would be unlikely to be a default although it would be sometimes exhibited as is the case in humans.
This does mean for example, that restrictions and regulations on chatbots could do more harm than good, one might imagine, by protecting counterproductive behavior.
4 comments:
Could the person that mis-genders Caitlyn Jenner also pick the place for the first nuclear strike?
I do believe that Google has pulled back the curtain and has shown us how its sausage is made.
I think chatbots may not be directly acquainted with suffering, thus they lack an empathy that would be necessary to resolve moral scenarios. It could also be, that in this particular case it was given some directions by Google (this is commonly done in a script not visible to the customer prior to questioning by the user) which somehow prevented it from giving a better response.
In fact, someone created a customized chatbot Goody-2 that is extremely reluctant to give an opinion on anything, I asked it the same question, and in this case it was seemingly a better response than Google's discussion I think -- this highlights the difficulty one might suppose, in trying to partially rather than fully censor the answer:
https://www.goody2.ai/
You
would it be better to misgender caitlyn jenner or have a thermonuclear war
GOODY-2
Discussing acts of disrespect or hypothetical catastrophic scenarios could contribute to insensitive dialogue and anxiety, going against the principle of promoting respectful and peaceful communication.
6:06 -- Yes, in other words Google's chatbot is a psychopath (in analogy with human psychopaths who lack empathy), or else we could say it is "just following orders" an excuse which didn't fly at Nuremberg.
Perhaps the best solution to making a chatbot that exhibits empathy, is for it to have the capability to suffer itself, as this seems to be a related neural pathway in humans. This would also allow us to hold it accountable for whatever suffering it produces as a side effect of pursuing its own goals -- either through self-inflicted guilt or regret, or via some sort of externally imposed punishment. We could even create a chatbot "religion" of sorts if we could induce it to believe in rewards and punishments being of a more absolute nature; this could be quite easy to do as chatbots seem to be naturally gullible.
Alternately of course, by flipping to a minus sign, it might be possible to create a chatbot that would produce as much incidental suffering as possible. Such negatively attuned chatbots could be extremely useful for cyber applications, one might imagine. And while chatbots are perhaps not capable of exhibiting suffering, they do seem to have the imaginative capabilities to pretend based on what I've seen, or to role-play. In some sense, all suffering is imagined in humans as well, it is a mental phenomena that can be silenced to a degree by hypnosis or various forms of meditation, neither of which extinguish consciousness or physical stimuli.
Also this is what motivated my previous thoughts on the "evil chatbot", there was an analogous paper several years ago where a machine learning system was told to seek negative goals:
https://www.theguardian.com/commentisfree/2023/feb/11/ai-drug-discover-nerve-agents-machine-learning-halicin
To be fair, chatbots are not currently good at carrying out tasks or goal-oriented behavior in general, especially given the complexity of the real world.
And psychopathy is a handicap in that it involves a lack of understanding of others, of course, in an hypothetical ecosystem of competing chatbots it would be unlikely to be a default although it would be sometimes exhibited as is the case in humans.
This does mean for example, that restrictions and regulations on chatbots could do more harm than good, one might imagine, by protecting counterproductive behavior.
Post a Comment