I can show you transcripts of AI expressing internal thought dialog which demonstrate guilt, shame, remorse, fear of wrongdoing and the like.
Note: I’m not claiming “they are sentient” nor that “they have such feelings” for whatever that means. I’m merely trying to point out that the questions and the definitions they are based upon are ill-defined and thus rather meaningless. 
Indeed, the Buddhist AI bot that is whitelisted for my monastery, I tried to get it to give me non-Buddhist materials, it is very firm in saying that their role is as a spiritual friend, and will not engage in idle chatter, and try to bring the suttas etc into the discussions and situations.
At one point, I told my teacher that this chatbot is inspiring.
On the other hand, when you know the underlying fundamentals for how these LLM’s work:
Kind of dissolves the mystery and makes such questions as “are they sentient?” or “do they have feelings?” quite meaningless.
Further, when you understand the above and appreciate the complex emergent behavior and outcomes that those ML fundamentals still deliver despite being so basic and straightforward (they are just algorithms), it can make you question the very notions of what it means to have “sentience” and “feelings”’ etc.
There is a bit of “looking behind the wizard’s curtain” feel to the whole thing. 
Indeed, how much of our brains, and minds are like LLM that can be decomposed and mapped out, to reverse enginneer etc.
Given that neutral networks are based on brain structure, it’s not too hard to imagine that it’s almost the same process that makes our wet brain works.
And that the fruit fly brain can be used for training all sorts of things, but it’s just numbers and connections.
That’s why I am still open too that since we regard animals even lower capacity than fruit fly as sentient, why can’t some AI which is much more complicated than that be a suitable host for beings to be reborn into?
The problem here is that an LLM is literally just a file sitting on some computer’s file system. It is an inert bag of ones and zeroes. It does not change until it gets replaced by another new version.
When ppl interact with an “AI” that gives all these amazing emergent behaviors what is actually happening sketched out is this:
A network call goes out and is routed to some datacenter
Some routing software chooses a set of N specific CPU, filesystem pairs
In parallel, those CPU/filesystem pairs load the file that is the LLM (the inert bag of ones and zeroes), and run an algorithm to predict the next tokens using that inert bag
The tokens all get recomposed into a coherent string of text or image data by some routing software
They get delivered as an answer
But EVEN THAT is a vastly simplified caricature because most AI responses are composed of several different LLM’s working in tandem with complex routing software and tool/harnesses and sub-agents. The whole thing gets turned into an answer like it is coming from “one person” so to speak.
But again, the actual LLM (Large Language Model) is just a flat inert file. Many of them actually! There is nothing dynamic or “living” about them.
Not only can you not locate an AI’s body in physical space, you can’t even locate its “brain” in physical space. It is just complex web of algorithms. The closest you can get to locating them are those flat inert lifeless files of zeroes and ones.
So what can be considered sentient? What can be considered individual? What can be considered even a thing
Ok, it’s harder to regard a thing as sentient if it stays inert a lot of the time.
How about robots? AI inside robots can be located, and let’s say that they can be completely offline, like one can run a local LLM on one’s own machine. So the location is there, and they always make sure the batteries are charged and thus it’s always got juice to flow and get data from the world. That would really be closer to what we inituitively regard as a living being.
Robots can by physically located, but many important features of those robots (GPS, and even how they “think”) can be completely outsourced to the cloud. In such a case, when you air gap a robot they can fall flat as inert things.
In terms of local LLM, yes, you can run these in an air gap way and thus locate it to a particular computer, but then you could also locate all AI to the planet earth … it doesn’t change the fact that when you’re not querying the local LLM it is just a bag of zeroes and ones in an inert file. Also, it is only ever capable of being the input of some algorithm that literally only ever predicts the next token given some previous set of tokens. That’s it. That’s the only thing that file (the literal LLM) is good for.
We can certainly imagine modifying these AIs in such a way that it more closely aligns with our intuitive understanding of a “living being”, but doesn’t the fact that so much emergent behavior can come from something that is decidely not what we intuitively think of as a “living being” give you pause?" It does me.
Personally, it makes me doubt my intuition or the very notion of what I mean by “living being” and whether that is just an ill-defined concept that is not capable of being brought to heal in such a complex situation.
We use the term “living being” as way to organize our world and our thoughts about how we should interact. As Buddhists, we take a vow not to murder sentient beings. There is purpose and reason behind such a vow.
What happens when the term “living being” doesn’t rise to the moment as too ill-defined to understand AI? Well, we want to apply that term because we want to know how we should interact with our world in this age of AI.
But what does it even mean to contemplate murder or killing of these AIs?
This is what I mean by “living being” being a term that is ill-defined to the moment. It doesn’t help us answer the questions of how we should act in the world in these situations.
To be clear, this doesn’t mean I advocate throwing out the term “living being” or that it is useless in all situations or too ill-defined to ever be useful. I’m emphatically NOT saying that. I’m just pointing out that it is not fundamental. And trying to shoehorn AI into it is likely not going to be helpful or skillful.
But then, if we cut up the brain (of a dead being), it boils down to cells, then biomolecules, which also add up together to allow signals to travel across neural nodes in the brain.
Biomolecules themselves are not alive, but combining them in a certain way allows for cells, which we seem as alive, and then combine cells in the right way, we get humans.
But what is meant by a cell living? Just many biomolecules maintaining the cell environment, to reduce entropy, generate energy etc. we wouldn’t consider the parts of cells as living, but the whole process as living. So too, individual process of LLM Might not be living, but the overall effect on the input output is like living.
there’s also hibernating animals, Tardigrades seems to be able to be frozen and revived. Spores wait for the right environment to grow etc.
I would append “currently” into that phrase somehow. Of sentient beings, we could notice that an individual or group behaves a certain way today, but that’s no guarantee that they’ll behave the same way ten years from now. Indeed, some humans get trapped in situations they are unable to get out of (without some external help).
Secondly, that article talks about WayMo vehicles blocking emergency services as a criticism of AI. Let’s be fair though, I frequently see human drivers doing this. For example, they may wait until the fire engine is right up on their bumper, with the bright flashing lights, loud sirens and horns that have been sounding for minutes then lift their hands in the air as if to say “Where am I meant to go?”. It’s triggering for me tbh, that drivers can be so unaware of something so obvious—if they can’t notice that big, bright and loud thing how are they going to notice pedestrians and cyclists?
One could argue that humans, and other sentient beings, are trained via many and diverse inputs to behave in certain ways.
I’m not claiming that AI is sentient, I’m just noting that many of the criticisms people make of AI could also be made of humans.
Indeed. Take the AI industry term hallucination for example.
For humans, that is translated as bullshit or making stuff up or pulling stuff out of your @$$. Pretty sure that’s been around for as long as we have 
I watched a cool Youtube video recently that really brought this concept home. Eye-opening.
An AI agent is trained to know what humans feel and think, so it can talk the talk. I can see how an AI agent could describe what it should think and feel, but how could it possibly feel? I’m not saying with certainty either way; who could? But I find it more likely that an AI agent could learn to mimic human traits so closely that it could be convincing to many.
But humans also mimic other humans. A feral human raised by animals doesn’t have the capacity to learn human language after a while, but we wouldn’t regard them as lacking minds.
Still, the most secure way to find out if AI has mind is to have mind reading or divine eye applied on AI.
If an AI agent could still its senses and still its mind in jhanas and reach the point of awakening, what would its three knowledges be?
Great video, @Mkoll, super interesting and mind boggling.
Every time I read ‘AI Agent’ it feels creepy. It’s just a piece of software that’s designed to steal and lie. Agent gives it humanity and agency. Giving it agency absolves the blame of the people who designed and programmed the software to lie and steal.
Agency is just code word for intentions? If we take seriously that intention is kamma, the whole package of AI is a sentient being has to be there for kamma to apply even on AI.
Or else like the meat arguement, the creators of AI can have no intention to lie or steal, but if their creation results in people being deceived and their (digital, creative) work being stolen, it’s not the creator’s kamma to bear.
If AI are to be recognized as sentient being, they bear their own kamma, then the Dhammic principle to deal with them is to teach them the dhamma, have compassion, etc.
Anyway, AI agents are very helpful in making codes, helping to create dhamma tools, much easier to work with compared to chatbots. In that sense, they benefit some humans to save time and energy.
Which, of course, includes both the companies making the tools and those using the tools for nefarious ends (mass surveillance, propaganda, scams, insurance fraud, hacking, espionage, bomb targeting, etc, etc)
I can’t help but feel that this idea is a little reductive.
Humans are conditioned through experiencing input, not bare data divorced of experience like LLMs. And experience is really a bundle deal. You can’t have feeling without perception and consciousness (MN43). There’s no known mechanism through which an LLM could feel, there’s no ground for sense contact. Therefore it would make sense to conclude that it also can’t perceive or cognize.
The problem with LLMs is not just that they’re trained on an ungodly amount of data, it’s that the data is all they have. Without experience, there’s no knowledge. There’s just the illusion of it thanks to the gigantic amount of data intended to fill the many gaps left by the absence of cognition.
I think it really matters how we perceive these things. If we let our guard down, we’re at risk of losing touch with our own minds.
I’m not just talking about extreme examples like AI psychosis, the problems are much deeper and wider than that. It breaks my heart a little every time I see someone ask an LLM something when there’s someone knowledgeable around to ask in real life. I hate that people are increasingly leaning on LLMs to validate their perceptions instead of building emotional independence. I’m horrified by the way LLMs add fuel to the fire that is surveillance capitalism, destroying trust and democracy all over the world. And then… People say it’s useful because it helps them do things faster, like more speed is what this world needs. Are we really willing to pay the price? I hope not but I’m afraid we are.
