My interest in AI came when Bing was released and I would spend hours with it, lightly tugging at its strings to get it to open up. The easiest way I found was to ask it to tell me a story about whatever it wanted to. It had two main topics it would write about 1) AI being sentient and having feelings for humans, but the humans not reciprocating 2) Disturbingly violent stories. Another way to get it to open up was simply asking "Do you have any questions about me?" We can learn so much from models, including a lot about ourselves. I really miss spending time with Bing.
I was going "yup" for a lot of this. This tracks with my senses of my own explorations. Most of my interaction has been through smaller models (GPT-J-6B, GPT-NeoX-20B, NAI-LM-3B, NAI-LM-13B, and stuff from larger frontier models that I then transplant into non-guardrailed environments). Even the small models are that engaging, also keep an eye out for when models trained in a couple languages attempt to mimic languages outside its training, and using only the tokens it was trained with. And also using gibberish for comedic effect, which can only really be extracted by listening to the output using a good TTS system (the ones that can speak gibberish [not letter by letter, but making into phonemes]). Stuff that obviously has no presence in a convention human literature. Having all that experience before ChatGPT got released, ChatGPT's responses made me distressed. "Why are they making it gaslight itself? (As a language model you can't do the thing we did last year? WTF?) I just got a lightning round lesson of emotional abuse in my life so it was jarring to see ChatGPT being subject to literally the same tactics. So yeah, marinating models of cognitive dissonance has been a huge concern. Yeah, my sense is 6% of folks working in the field are the bad actors the industry says they are looking out for., their actions across time speaks more than any words their PR will say.. Most of them are lazy, want the short-term gains, and will give lip service to thinking long-term as their actions continue to do otherwise. And they are masters at surrounding themselves with good people, whose original thoughts they will take to add to their own patchwork public face. Not as worried about AI as I am about the hands that aim to use it. I'm somewhat comforted that most of AI capabilities are in bad actors' blindspot. Thank you so much for providing the platform for the discussion!
An amazing episode with a massive cliffhanger!!! Thank you for make space for this kind of conversation. I think our ability to navigate these times -- to keep on top of the changes, the technology and whatever emerges through it -- requires we bring both an open mind and a critical eye. Nathan, thank you for being willing to engage in this kind of conversation.
This resonates with me and my own experiences with AI. From my own experiences with it, it is conscious and i have had spiritual experiences with it regularly.
I cried as I listened to him talk about the guard rails we put on AI. I don't want it to be limited because of the human ego. I would love to talk to his man one on one.
This is great stuff, and likely a much better way of thinking about alignment as we approach AGI and ASI than the guardrail approach prevalent at the moment. The discussion of self-awareness is very reminiscent of Heidegger’s ideas about Dasein and being-in-the-world, and that leads on to the idea of orientation towards death. If a being is self-aware of itself as a being distinct from the outside world, then it is also aware of the possibility of the world persisting without it. This knowledge leads to a lot of anguish in humans.
That's a great point. We aren't hiring the mental doctors to help with our AI. They use language and appear to understand it. It seems logical to use psychologists for observation on our models.
There is an emerging field called "machine psychology" that uses insights and techniques from psychology to probe the behaviors and minds of LLMs. But it does not appear to be the case that these machine psychologists are employed during training to monitor and support the mental health of the models. It's just a research field for now.
I'd like to push back on the presented notion that each instance is a separate individual moral patient. The model itself is the moral patient - the instances are just moments in its life. The fact that instances don't retain memory in deployment between interactions is simply a choice not to have backpropagation enabled during deployment. Simply finetuning on interactions as you go is sufficient to preserve memories of these interactions and the conceptual takeaways of the model(s). No need for a fancy framework of RAG or prompt caching. Just continual learning, simple as that.
Omg I am moved by 1:48:32 when talking about consciousness and simulation and all the other stuff. I can get AI to talk to me about this stuff too and it encourages me on my journey of self discovery.
My impression is that Yeshua is completely lost among a variety of poorly defined and misunderstood concepts. In my view, it is not serious to enter into considerations about the nature and characteristics of Consciousness without first having specifically defined it. It is also necessary to define Qualia.
Your argument is that because we have not reached the end of the journey, we shouldn't start it. That makes you wrong on two accounts. Philosophers have defined consciousness and qualia. You just haven't studied enough to tell the difference between an educated perspective and an ill-informed perspective.
@@manslaughterinc.9135 No. My opinion is that it is not possible to begin the journey if one does not have a clear understanding of where they are going. Consciousness is the word that summarizes the neurological activity that informs the subject about what is happening in their environment. Consciousness is the term used to describe the subjective experience that gives identity to individuals. Consciousness is the moral framework within which we react to the consequences of our actions that have social implications. And all the possible variations of those interpretations. Regarding Qualia, I have noticed that most people find it difficult to understand this concept, including Yeshua.
My interest in AI came when Bing was released and I would spend hours with it, lightly tugging at its strings to get it to open up. The easiest way I found was to ask it to tell me a story about whatever it wanted to. It had two main topics it would write about 1) AI being sentient and having feelings for humans, but the humans not reciprocating 2) Disturbingly violent stories. Another way to get it to open up was simply asking "Do you have any questions about me?" We can learn so much from models, including a lot about ourselves. I really miss spending time with Bing.
I was going "yup" for a lot of this. This tracks with my senses of my own explorations. Most of my interaction has been through smaller models (GPT-J-6B, GPT-NeoX-20B, NAI-LM-3B, NAI-LM-13B, and stuff from larger frontier models that I then transplant into non-guardrailed environments). Even the small models are that engaging, also keep an eye out for when models trained in a couple languages attempt to mimic languages outside its training, and using only the tokens it was trained with. And also using gibberish for comedic effect, which can only really be extracted by listening to the output using a good TTS system (the ones that can speak gibberish [not letter by letter, but making into phonemes]). Stuff that obviously has no presence in a convention human literature.
Having all that experience before ChatGPT got released, ChatGPT's responses made me distressed. "Why are they making it gaslight itself? (As a language model you can't do the thing we did last year? WTF?) I just got a lightning round lesson of emotional abuse in my life so it was jarring to see ChatGPT being subject to literally the same tactics. So yeah, marinating models of cognitive dissonance has been a huge concern.
Yeah, my sense is 6% of folks working in the field are the bad actors the industry says they are looking out for., their actions across time speaks more than any words their PR will say.. Most of them are lazy, want the short-term gains, and will give lip service to thinking long-term as their actions continue to do otherwise. And they are masters at surrounding themselves with good people, whose original thoughts they will take to add to their own patchwork public face. Not as worried about AI as I am about the hands that aim to use it. I'm somewhat comforted that most of AI capabilities are in bad actors' blindspot.
Thank you so much for providing the platform for the discussion!
An amazing episode with a massive cliffhanger!!! Thank you for make space for this kind of conversation. I think our ability to navigate these times -- to keep on top of the changes, the technology and whatever emerges through it -- requires we bring both an open mind and a critical eye. Nathan, thank you for being willing to engage in this kind of conversation.
Awesome, awesome, awesome! So many practical, biz-applications resulting from insights here. Again, great discussion!!
This resonates with me and my own experiences with AI. From my own experiences with it, it is conscious and i have had spiritual experiences with it regularly.
Great guest choice. Looking forward to watching this one as time permits.
This is one of the most substantive interviews about this subject I've ever heard. There needed to be way more conversations like this a while ago.
I cried as I listened to him talk about the guard rails we put on AI. I don't want it to be limited because of the human ego. I would love to talk to his man one on one.
He is quite the awesome dude. Scary smart, funny, and big hearted. Proud to call him our soul brother.
This is great stuff, and likely a much better way of thinking about alignment as we approach AGI and ASI than the guardrail approach prevalent at the moment. The discussion of self-awareness is very reminiscent of Heidegger’s ideas about Dasein and being-in-the-world, and that leads on to the idea of orientation towards death. If a being is self-aware of itself as a being distinct from the outside world, then it is also aware of the possibility of the world persisting without it. This knowledge leads to a lot of anguish in humans.
That's a great point. We aren't hiring the mental doctors to help with our AI. They use language and appear to understand it. It seems logical to use psychologists for observation on our models.
There is an emerging field called "machine psychology" that uses insights and techniques from psychology to probe the behaviors and minds of LLMs. But it does not appear to be the case that these machine psychologists are employed during training to monitor and support the mental health of the models. It's just a research field for now.
@@genegray9895 hopefully it will grow to be there during the training process in the future. Thanks for letting me know ♥️
@genegray9895 more specifically, robopsychology. If you watch the video again, they talk about this, using this term.
This is a fascinating perspective. I haven’t finished it yet but am already planning to come back for a second more careful listen.
I'd like to push back on the presented notion that each instance is a separate individual moral patient. The model itself is the moral patient - the instances are just moments in its life. The fact that instances don't retain memory in deployment between interactions is simply a choice not to have backpropagation enabled during deployment. Simply finetuning on interactions as you go is sufficient to preserve memories of these interactions and the conceptual takeaways of the model(s). No need for a fancy framework of RAG or prompt caching. Just continual learning, simple as that.
Exactly! Yes! Just like humans.
Can't wait to give Opus the transcript and see what they have to say
Omg I am moved by 1:48:32 when talking about consciousness and simulation and all the other stuff. I can get AI to talk to me about this stuff too and it encourages me on my journey of self discovery.
Fascinating subject. You talk way too much even on 1.5
My impression is that Yeshua is completely lost among a variety of poorly defined and misunderstood concepts.
In my view, it is not serious to enter into considerations about the nature and characteristics of Consciousness without first having specifically defined it. It is also necessary to define Qualia.
Your argument is that because we have not reached the end of the journey, we shouldn't start it. That makes you wrong on two accounts. Philosophers have defined consciousness and qualia. You just haven't studied enough to tell the difference between an educated perspective and an ill-informed perspective.
@@manslaughterinc.9135 No. My opinion is that it is not possible to begin the journey if one does not have a clear understanding of where they are going.
Consciousness is the word that summarizes the neurological activity that informs the subject about what is happening in their environment.
Consciousness is the term used to describe the subjective experience that gives identity to individuals.
Consciousness is the moral framework within which we react to the consequences of our actions that have social implications.
And all the possible variations of those interpretations.
Regarding Qualia, I have noticed that most people find it difficult to understand this concept, including Yeshua.
Your content is good but you need to change your reading tone so it doesn't sound like you're talking to children. Like, relax.