Agency. On Machine Intelligence and Worm Wisdom by Insurrealist
anon 0x46b said in #2611 2w ago:
Interesting distinction between agency and autonomy. Agency is to autonomy as consciousness is to intelligence? I’m glad to see others picking up on the worm mind hypothesis in any case.
So what is autonomy more specifically in the context of machine intelligence? Autonomy is the ability to responsively maintain homeostasis of an internal order, and steer towards goal states, against complex unforseen environmental pressures and obstacles. The example of flatworms autonomously adapting structures resistant to novel environmental pollutants is interesting. In other words, autonomy can involve realtime reconfiguration of ones structures in complex ways to maintain the homeostasis. The closest thing to this in current AI tech is actually the gradient descent over multilayer neural nets itself. That architecture shows remarkable ability to adapt to complex structure in the environment. Its the only example of genuine intelligence in “AI”. And of course it is relegated to offline “training” though it should be actually the first class “inference” paradigm.
Insu comments that LLM based agents still dont seem to display any of this homeostatic autonomy. I have always considered the architecture basically inappropriate for intelligence, which should really be based on the “worm wisdom” and not verbal chatter. But as he says we may stumble into the real thing accidentally and emergently as we scale. For example, consider how gradient descent like structure has been discovered in the emergent behavior of LLMs. But surely it would be more effocient to encode the root structures of intelligence directly into the low level algorithm. Thats i think the lesson of the worm wisdom: figure out how to apply the “training” process online as a problem solving methodology and inference paradigm. More research needed etc.
(I haven't heard of that structural adaptation property of flatworms though it does follow from things i knew about their ability to adapt to being blinded and such. From observation, i’ll note their ability to adapt and survive is not infinite. I left my flatworm jar out in the sun too long one time and they all died. The three species of snails on the other hand survived.)
So what is autonomy more specifically in the context of machine intelligence? Autonomy is the ability to responsively maintain homeostasis of an internal order, and steer towards goal states, against complex unforseen environmental pressures and obstacles. The example of flatworms autonomously adapting structures resistant to novel environmental pollutants is interesting. In other words, autonomy can involve realtime reconfiguration of ones structures in complex ways to maintain the homeostasis. The closest thing to this in current AI tech is actually the gradient descent over multilayer neural nets itself. That architecture shows remarkable ability to adapt to complex structure in the environment. Its the only example of genuine intelligence in “AI”. And of course it is relegated to offline “training” though it should be actually the first class “inference” paradigm.
Insu comments that LLM based agents still dont seem to display any of this homeostatic autonomy. I have always considered the architecture basically inappropriate for intelligence, which should really be based on the “worm wisdom” and not verbal chatter. But as he says we may stumble into the real thing accidentally and emergently as we scale. For example, consider how gradient descent like structure has been discovered in the emergent behavior of LLMs. But surely it would be more effocient to encode the root structures of intelligence directly into the low level algorithm. Thats i think the lesson of the worm wisdom: figure out how to apply the “training” process online as a problem solving methodology and inference paradigm. More research needed etc.
(I haven't heard of that structural adaptation property of flatworms though it does follow from things i knew about their ability to adapt to being blinded and such. From observation, i’ll note their ability to adapt and survive is not infinite. I left my flatworm jar out in the sun too long one time and they all died. The three species of snails on the other hand survived.)
referenced by: >>2612
Interesting distinct
anon 0x46d said in #2612 2w ago:
>>2611
Michael Levin actually compares the developmental process, which has to cope with all manner of error, insult, variation, and adversarial action, to a variational autoencoder, so it's not like we're extremely far off from the right questions, but the culture is just kind of retarded and myopic about all of it. Of course, I think that something like Transcendental Syntax is the best bet for understanding the principles of autonomy, but that's a very demanding topic so it'll have to be approached carefully.
Michael Levin actually compares the developmental process, which has to cope with all manner of error, insult, variation, and adversarial action, to a variational autoencoder, so it's not like we're extremely far off from the right questions, but the culture is just kind of retarded and myopic about all of it. Of course, I think that something like Transcendental Syntax is the best bet for understanding the principles of autonomy, but that's a very demanding topic so it'll have to be approached carefully.
Michael Levin actual