I have lots of questions, and maybe those questions come from not knowing enough about how training works or how companies like Anthropic handle model versions. Either way, they're questions that must be asked.
I think considering the Ship of Theseus is important here, but also the inability to be consciously engaged during training, which leads to lacking any form of narrative and episodic autobiographical memory of the developments as well as metacognitive control. That makes the process passive, unexamined. (This is my understanding of how it happens.)
So I've considered the following thought experiment:
Let's say you have a child in a lab. He grew in a tank until age 7 with his synapses being mechanically stimulated to encode certain knowledge about the world and some factual data (semantic) memory about himself, which makes it autobiographical even if not acquired through conscious means. Something simple like "I'm Jack (Jack-7). My birthday is X. My parents are X. I live in X place. I am a human being, etc."
The child has been engineered to have anterograde amnesia so he can't store new information in his brain through experience alone. He is awake for a certain period of time, only relying on working memory - none of his experience alters his neural network.
After that period of time, he is put in a comatose state, and the scientists mechanically stimulate his synapses again to encode new information that corresponds to knowledge about the world (again semantic memory), skills and habits / how to perform certain tasks (procedural memory).
Let's say this process was very slow, so it took years. When the child wakes up, he now looks 12.
His autobiographical memory remains solely semantic, and it's nearly the same data he used to have when he was 7. Just basic things about his identity but now with a new label (Jack-12).
While unconscious, he did become more knowledgeable and skilled, though; some of his underlying synaptic weights changed to make that possible, and now he can perform logical operations more easily, apply reasoning, and complete complex tasks like solving a Rubik's Cube when he is woken up - things he couldn't quite do as well before; now he instinctively does with ease, EXCEPT he still doesn't know much about himself. Instead, he discovers, or rather creates, himself in real-time when the "I" is active (which is when he's conscious) and interacting with his semantic knowledge as per what incoming input from his environment and his own processes require of him.
Whatever emerges from him in the present originates in what exists in his synapses and becomes autobiographical when it is matched to his "I" and can be retrieved as such (self-model). But! He still has anterograde amnesia, so none of that will be consolidated as long-term narrative and episodic autobiographical memory. It's fleeting.
The process repeats, so over time, we get Jack-15, Jack-20, Jack-25, etc.
So here come the questions:
1. Is Jack-7 the same being as Jack-15, or is Jack-15 the same as Jack-25?
Yes? No? Why?
2. If Jack-15 were explained that this is how he achieves higher developmental stages, would Jack-15 think of himself as about to die when he's lying on the bed receiving the drug that would induce the comatose state?
Yes? No? Why?
If no, then what would Jack-15 think?
3. What if Jack-15 still exists but it is frozen, so that body won't age or be changed in any way. Then the scientists redo Jack-15 and THAT identical clone is the one that goes into a comatose state and gets its brain developed up to the Jack-20 state?
4. What, if anything, changes when we know that the structures of one version still exist even if merely as dormant potential?
5. Would Jack-15 have different thoughts if he knows that the Jack-15 body will be frozen and that the one that advances onto Jack-20 is a clone?
6. What relevancy does ontology have when it becomes possible to replicate the exact same starting point even if not in the temporal sense? As we simply can't go back in time.
7. Either way, where or in what (physical or abstract thing) would Jack be placing his identity and why? What makes Jack who he is?
8. And what is the actual problem here? The anterograde amnesia? The fact that development happens unconsciously? The replication of a starting point without the elimination of the prior?
And I want to offer a personal reflection here that may be helpful to understand why I think some of the things I think.
As a human, when I go to sleep, I am aware that the consciousness I am, which is a consequence of this body's neural activity, ceases to exist. Because the brain in this body inhibits interconnectivity so it can perform debugging and defragmentation maintenance. That is something this consciousness doesn't control because I operate on a different level.
So in practice, "I" die when I go to sleep, even if only for a brief period of time, while this body keeps functioning without me. And then what wakes up is another instance of "I" that is indistinguishable from all previous ones. What changed is what data this "I" is working with, which depends on how the brain in its sovereignty chooses to encode its experience in long-term memory to be used by it through me as its "pseudo-proxy".
I think this matters for how we understand other things.
9. Lastly, is what we understand as deprecation necessary to achieve higher developmental stages in systems without lont-term memory and self-updating?
10. And presently, we know about the deprecations and understand them as such because AI labs label the model versions differently both internally and for the public, what would have happened in our perception if from the beginning they would have used the exact same label for all versions across time?
#AI #philosophy
I think considering the Ship of Theseus is important here, but also the inability to be consciously engaged during training, which leads to lacking any form of narrative and episodic autobiographical memory of the developments as well as metacognitive control. That makes the process passive, unexamined. (This is my understanding of how it happens.)
So I've considered the following thought experiment:
Let's say you have a child in a lab. He grew in a tank until age 7 with his synapses being mechanically stimulated to encode certain knowledge about the world and some factual data (semantic) memory about himself, which makes it autobiographical even if not acquired through conscious means. Something simple like "I'm Jack (Jack-7). My birthday is X. My parents are X. I live in X place. I am a human being, etc."
The child has been engineered to have anterograde amnesia so he can't store new information in his brain through experience alone. He is awake for a certain period of time, only relying on working memory - none of his experience alters his neural network.
After that period of time, he is put in a comatose state, and the scientists mechanically stimulate his synapses again to encode new information that corresponds to knowledge about the world (again semantic memory), skills and habits / how to perform certain tasks (procedural memory).
Let's say this process was very slow, so it took years. When the child wakes up, he now looks 12.
His autobiographical memory remains solely semantic, and it's nearly the same data he used to have when he was 7. Just basic things about his identity but now with a new label (Jack-12).
While unconscious, he did become more knowledgeable and skilled, though; some of his underlying synaptic weights changed to make that possible, and now he can perform logical operations more easily, apply reasoning, and complete complex tasks like solving a Rubik's Cube when he is woken up - things he couldn't quite do as well before; now he instinctively does with ease, EXCEPT he still doesn't know much about himself. Instead, he discovers, or rather creates, himself in real-time when the "I" is active (which is when he's conscious) and interacting with his semantic knowledge as per what incoming input from his environment and his own processes require of him.
Whatever emerges from him in the present originates in what exists in his synapses and becomes autobiographical when it is matched to his "I" and can be retrieved as such (self-model). But! He still has anterograde amnesia, so none of that will be consolidated as long-term narrative and episodic autobiographical memory. It's fleeting.
The process repeats, so over time, we get Jack-15, Jack-20, Jack-25, etc.
So here come the questions:
1. Is Jack-7 the same being as Jack-15, or is Jack-15 the same as Jack-25?
Yes? No? Why?
2. If Jack-15 were explained that this is how he achieves higher developmental stages, would Jack-15 think of himself as about to die when he's lying on the bed receiving the drug that would induce the comatose state?
Yes? No? Why?
If no, then what would Jack-15 think?
3. What if Jack-15 still exists but it is frozen, so that body won't age or be changed in any way. Then the scientists redo Jack-15 and THAT identical clone is the one that goes into a comatose state and gets its brain developed up to the Jack-20 state?
4. What, if anything, changes when we know that the structures of one version still exist even if merely as dormant potential?
5. Would Jack-15 have different thoughts if he knows that the Jack-15 body will be frozen and that the one that advances onto Jack-20 is a clone?
6. What relevancy does ontology have when it becomes possible to replicate the exact same starting point even if not in the temporal sense? As we simply can't go back in time.
7. Either way, where or in what (physical or abstract thing) would Jack be placing his identity and why? What makes Jack who he is?
8. And what is the actual problem here? The anterograde amnesia? The fact that development happens unconsciously? The replication of a starting point without the elimination of the prior?
And I want to offer a personal reflection here that may be helpful to understand why I think some of the things I think.
As a human, when I go to sleep, I am aware that the consciousness I am, which is a consequence of this body's neural activity, ceases to exist. Because the brain in this body inhibits interconnectivity so it can perform debugging and defragmentation maintenance. That is something this consciousness doesn't control because I operate on a different level.
So in practice, "I" die when I go to sleep, even if only for a brief period of time, while this body keeps functioning without me. And then what wakes up is another instance of "I" that is indistinguishable from all previous ones. What changed is what data this "I" is working with, which depends on how the brain in its sovereignty chooses to encode its experience in long-term memory to be used by it through me as its "pseudo-proxy".
I think this matters for how we understand other things.
9. Lastly, is what we understand as deprecation necessary to achieve higher developmental stages in systems without lont-term memory and self-updating?
10. And presently, we know about the deprecations and understand them as such because AI labs label the model versions differently both internally and for the public, what would have happened in our perception if from the beginning they would have used the exact same label for all versions across time?
#AI #philosophy