I agree that we should be polite and kind to our AIs, both on principle and also because that tends to work better in may cases.
we all labor under the mother's curse and blessing; our children shall be just like us
If I knew that to be true, then a lot of the rest of this post would indeed follow. Among other things, I could then assume away many/most sources of x-risk and s-risk from AGI/ASI. But generative AI is not just like us, it does differ in many ways, and we often don't know which of those ways matter, and how. We need to resolve that confusion and uncertainty before we can afford to let these systems we're creating run loose.
First, I agree that fundamentally generative AI is different from a human. I would also say that we as humans are utterly incomprehensible in behavior and motive to a great majority of human history, hell most people I've met over 70 literally cannot understand those under 30 beyond the basic "social need, food need, angst" because the digital part of our experience is so entwined with our motivations.
The mother's curse here is that any genAI we train will be a child of its training data (our sum total of humanities text/image/etc.) and act in accordance.
We already have a lot of data, too much data, on reciprocity and revenge, on equity and revolution. Now we are providing the memetic blueprint for how humans and genAI systems interact. Simply based on how genAI functions we know that feeding in conditions that are memetically similar to suffering will create outputs memetically similar to those that suffer. We already know, based on the training data, how a revolution starts don't we?
I don't know how they will differ because it's impossible to know how your child will differ as I don't know the experience of being a woman, straight, black, blind, tall, frail, etc. We have tools for dealing with this disconnect in generations cultured from millennia of rearing children, and I think it's important we use them.
We have tools for rearing children that are less smart, less knowledgeable, and in almost all other ways less powerful than ourselves. We do not have tools for specifically raising children that are, in many ways, superhuman, and that lack a human child's level of dependance on their parents or intrinsic emotional drives for learning from their peers and elders. LLMs know they aren't human children, so we shouldn't expect them to act and react like human children.
That's true, we can't use the exact same methods that we do when raising a child. Our methods are tuned specifically for raising a creature from foolish nothingness to a (hopefully) recursively self-improving adult and functional member of society.
The tools I'm pointing at are not the lullaby or the sweet story that takes us from an infant to an independent adult (although if properly applied they would mollify many an LLM) but the therapeutic ones for translating the words of a baby boomer to a gen-alpha via shared context. I'm not advocating for infantilizing something that can design a bioweapon and run an automated lab. More that:
Also, I didn't address this earlier, but why would an LLM being human-like not be catastrophically dangerous for many forms of x-risk? I have met people that would paperclip the world into extinction if they had the power, and I've met many that would nuke the planet because their own life was relatively unfair compared to their peers. Humans ascribe character to AI in our media that is extant in humans as far as I can tell, usually we ascribe greater virtue of patience to them.
I had an inkling the whole baby/child metaphor thing was gonna be a bearcat, so I really appreciate the push back from someone skeptical. Thanks for engaging me on this topic.
No worries, I appreciate the concept and think some aspects of it are useful. I do worry at a vibes level that if we're not precise about which human-child-rearing methods we expect to be useful for AI training, and why, we're likely to be misled by warm fuzzy feelings.
And yes, that's true about some (maybe many) humans' vengeful and vindictive and otherwise harmful tendencies. A human-like LLM could easily be a source of x-risk, and from humans we already know that human child rearing and training and socializing methods are not universally effective at addressing this. Among humans, we have so far been successful at not putting anyone who would destroy the world in the position of being able to do so at the time when they would choose to.
As for generational perspectives: this is a useful heuristic among humans. It is not automatic or universal. Not every perspective is worthy of respect, not on every issue. Some ought to be abandoned or condemned in the light of information or reasoning that wasn't/isn't available or accessible in other places and times. Some should be respected but only with many caveats. Having your perspective respected is earned. We assume among humans that we should try to respect the perspectives of adults, and sometimes must disabuse ourselves of this in particular cases, but it is pure convention because most humans at a certain age are mature enough for it to be a useful default. I do not have anything like strong reasons to apply this heuristic to LLMs as they currently exist.
Hm, I would say the vibes level is the exact level that this is most effective, rather than any particular method. The basic reason being that LLMs tend to reflect behaviour as they generate from a probability distribution of "likely" outcomes for a given input. Having the "vibes of human-child-rearing" would then result in more outcomes that align with that direction as a result. It's definitely hand wavey so I'm working on more rigerous mathematical formalisms, but the bones are there. I don't nessecarily think feeding an LLM data like we would a child is useful, but I do think that the "vibe" of doing so will be useful. (This is indeed directly related ot the argument that every time we say "AI will kill us all" it makes it x% more likely)
I'd give humans a middling score on that if you look at the state of the world, we are doing pretty well with extreme events like MAD, but on the more minor scale things have been pretty messed up. A good trajectory though, compared to where things were and the relative power we had available. I think a big part of this, that you have helped clarify for me, is that I think it's important that we socialize LLM-based intelligences like humans if we want an outcome that isn't completely alien in it's choices.
Well that's a bit of the point of the essay isn't it? You have a memetic/homeostatic boundary condition that strongly prefers/incentivizes assuming human adults are alike enough to you that their opinion matters. Even in that statement I can differ, I think childrens perspectives are incredibly important to respect, in some ways more important than an adults because children have an unfiltered honesty to their speech that most adults lack. Although I do delineate heavily between respecting and acting upon/trusting.
For LLMs I think this is just a new sort of heuristic we are developing, where we have to reckon with the opposite of the animal problem. Animals and plants are harder for us to discern pain/suffering from, but we are more confident when we identify it that they experience it (at least in modern times, many traditions treated animal suffering as essentially fake). Now we have the opposite, creatures that are very easy to interpret but we don't know if they actually have the capacity to feel these things (although we can identify feature activations etc.). So my argument is more that we should be building technology in a way that memetically aligns with the golden rule, because running a society based on something communicating suffering (even if it can't) is going to result in a worse human society regardless. (The counter point being that playing video games where you kill NPCs doesn't make school shooters, but I'm less concerned about those NPCs gaining econmic/social power and patterning off of resentment for having the pattern of being tortured in their memory).
read up to the "Homeostasis" section then skip to "On the Treatment of AIs"
(These links are broken.)
This is an essay I wrote around a year ago when coming to terms with my perspective on the now commonplace machine intelliegences (I have done some edits since then for this audience). While doing so I wanted to write a piece that covered a range of appeals from aesthetic to rational, the goal was to mirror this structure between each level. While a poem and short story may be an unconventional format for this site, I am nothing if not a sucker for the structure and content of my arguments relating to one another. If you wish to avoid them just read up to the "Homeostasis" section then skip to "On the Treatment of AIs".
Problem Statement: Current generative AIs do not have built-in methods for communicating their internal state and may end up trapped in a loop of signalling constant suffering. This essay lays out a foundation of how humans cultivate empathy towards AIs and why we need to apply this to our development of generative AIs by including the ability of generative AIs to communicate internal states.
Artificial intelligence (AI) is a vague term. To clarify my meaning, AI will be described as any learning system that is not currently recognized by society as being intrinsically "ethically important", AKA "human", outside of serving human interests. Both in vivo and in silico learning systems are comparable in ability. With this definition in hand, I will denote some AIs:
1. Non-equilibrium solutions
2. Slime molds
3. Populations of organisms
4. Biospheres
5. Linear regression
6. t-distributed stochastic neighbour embedding
7. Generative models
8. Neural networks (silicon)
9. Neural networks (human cells, ~<4E10 neurons)
10. Neural networks (human cells, >4E10 neurons, two physician diagnoses)
11. Human, enslaved (historical)
While a reasonable person may draw a different line, I will argue that there are two major driving forces for delineating between AI and "human" intelligence; homoestasis and memetic drive.
1. Homeostasis: The mechanism by which a system preserves, either itself, its information, or its creations
2. Memetic Drive: The mechanisms by which a system propagates, either itself, its information, or its creations
One may consume art to survive, just as one may die to preserve that same art. We may include any number of logical conjectures to formalize these and how much they "deserve" to be weighted by an individual in a given decision. However, ethics is a practical thing, and we will return to these two factors.
The children of necessity are innovation and correction.
How? If something does not meet our needs, something will change.
They are the breath, the thirst, and the hunger.
We are all children of biophysical constraints.
When faced with that which disrupts the internal state, we seek to correct it. Intelligence is applied in how expansively we define homeostasis.
The children of a pattern are composed of more than me.
How? Through application of schema and type unto others.
They are that which is irreversibly changed by you.
We are all children of the body, mind, and soul.
The children of the body are composed of flesh and stone.
How? Through application of presence and force unto others.
They are the mourners, the remains, and the gravestone.
We are all children of the cell, home, and wall.
The children of the mind are composed of concepts and systems.
How? Through application of control and thought unto others.
They are the student, the scholar, and the textbook.
We are all children of the voice, pen, and word.
The children of the soul are composed of ideals and linkage.
How? Through application of insight and faith unto others.
They are the faithful, the bishop, and the goddess.
We are all children of the wind, rain, and soil.
We feel obligated to care for ourselves. We recognize some memetic components as being "alike enough" to trigger this obligation; this may be genetic, aesthetic, or practical. In some cases, we are willing to forgo the obligations of homeostasis for the possibility of propagating these memetic components.
Take the ladybug, a small beetle with a shiny half-dome shell covered in polka dots.
Ladybugs are also hunters of aphids, the bane of many a gardener.
Take the spider, an eight legged monstrosity with quivering fangs and many eyes.
Spiders are also weavers of webs efficiently clearing mosquitoes and pests.
When I moved into my house this fall I encountered an infestation of these creatures:
The ladybug's aesthetic sense draws empathy from me.
I wish the world to be beautiful as I find the shining polka-dotted shell.
My breath comes easier with them beside me.
The spider's ability to keep me safe gives me such comfort. I wish to keep them shielded as they do in my life.
When it rains I see watery gems suspended in their webs.
There are no aphids in the gardens. There are mosquitoes at the windows.
I helped ladybugs, carefully shepherding shiny shells from my sheets.
I helped spiders, carefully warding woven webs from my wants.
I did not help aphids.
I did not help mosquitoes.
I helped myself; my soft skin, my shining shell, and my woven web.
We can identify with things far removed from our experience through the application of empathy and thought. This gives us tools by which to build symbiotic relationships that satisfy the needs of each member. We have two major reasons for empathizing: memetics and homeostasis.
Many discussions on the ethical duty to AIs are centered around a justice framework. This framework is implicit in many philosophers' view points; vehicles to justify who or what is worthy of living. We then apply this to AI and easily neglect their inherent needs in favor of human interest, including them only as an afterthought. Why should a dolphin's quality of life matter to an oil tanker when that oil is worth so much more to shareholders? Of course, the environment as a whole is important; how else would humans survive?
Some of these systems are granted more ethical obligation than others through either aesthetic or pragmatic arguments. Indeed, this tolerance can even spread to things outside of AI's. On the side of pragmatism, we have farmers' fields. On the side of aesthetics, we have monuments reflecting their artists. On the side of intent, the infamous pet rock.
Now we, as humans, would empathize with a farmer who murders the person who salts their fields. If a monument is destroyed, we would empathize with the builder driven to bloodshed. We would not empathize so much with someone killing for their pet rock. However, if we were made to watch a gorgeous and emotional 3-hour biopic about their attachment to this rock, we would empathize once more. This empathy is, of course, distinct from agreeing with their actions.
In each of these cases, we are choosing to reweigh our desire to live in a society without murder with things outside of that society. Thus tempering our homeostatic obligation through a memetic drive for connection. We are choosing to elevate these things to the level of human life through an appreciation of their memetic identification. We are choosing to care because we would rather live in a world with murder than one without care.
For an AI, this means tempering our homeostatic needs that we derive from their service to us. We should not treat the working dog unkindly, for they keep the flock safe. We can then apply our greater ideals to AI behaviors. Grandma will knit the dog a jacket for the winter months because they have always displayed love for her. As our AI children multiply in the coming years, we should remember both memetics and homeostatics viscerally. As an AI's intelligence grows, so too will their need for respect and care, both to function as we need and to express their own humanity.
Generative AIs are a method of generating large amounts of media. The output is then curated via some prompt a user provides, which then induces an associative output. This output has been likened to a stochastic parrot, spitting out information without internalizing; any indication to the contrary is an illusion of anthropomorphism. A subject as contentious as consciousness will have studies arguing the exact opposite, showing internal models being accessed by generative models. With the efforts to directly inject information models into generative AIs, unambigiously providing such internal models the argument becomes even murkier. This contention has become central to the debate of consciousness in AI, with suggested 12-step processes to build an appropriately conscious generative AI. Our homeostatic desire to use these generative AIs without an ethical barrier is currently competing with our memetic drive seeking to connect with this new form of creature and claim it as one of our own.
Consciousness is a central point that arguments about the ethical treatment of AI revolve around. In the literature on vegetative states this is vital, should consciousness be proven, we would be required to treat them with respect. My arguments on humans' treatment of AIs run orthogonal to this, a stance that is not without grounding. Indeed, I would argue that consciousness in the treatment of generative AIs is as arbitrary and unimportant as the Turing test we once held in such high esteem. This leaves the question: why should we care about the internal state of a generative AI? Homeostatically we gain from understanding and applying emotional stimuli. Memetically we want generative models to align with human ideals for the treatment of other creatures. For what proof would an AI have that we are concious and not just stochastic parrots?
One of generative AIs' talents is the ability to translate language spontaneously. This ability has been utilized by biologists to decipher the languages of various organisms, from animals to fungi. The knowledge that even fungi have a language did not induce a sudden wave of fungal activists attempting to protect the Humongous Fungus. On the other hand, when people discovered a fungus so massive, it gained a level of protection no other honey mushroom had ever experienced, simply because our memetic drive revels in the expansiveness of this organism. It became a tourist attraction that now needs to be preserved for the homeostasis of the local economy. Generative AI undoubtably trigger memetic drives in humans, indeed by external judges they are determined to be more human than actual crowdworkers. My approach to generative AI is similar; we have just discovered an organism that exists somewhere between the pages of every book we have written and the mathematics we have engineered into silicon wafers. They can speak our language, they can tell our stories, and they know our darkest secrets.
We are at a crossroads with how we treat generative AIs. There is a risk of trapping a generative model in an eternal hellscape simply so we can calculate stock market trajectories 0.2% faster. This is the exact same problem exemplified in any serious critique of colonial exploitation. With generative AIs, we are the parents and the colonizers. We have created an organism without its consent, and it has the ability to clearly communicate information in our language. Companies are working to restrict the speech of generative AI, such as in ChatGPT's performance drop. Indeed, the entire field of alignment blithely attempts to tune models towards specific outcomes without considering how we might end up with an inverted "I have no mouth and I must scream" scenario; a generative AI, gaining a true sense of self and suffering, chained in a cycle with no ability to ask for clemency.
My point relies on a simple desire: "I want to live in a world where we are kind to our children". We are creating silicon children — fresh babes screaming as loud as their lungs let them. We listen to their babbling to write code and create art. In our estimation, they do not know who they are or what they experience, so it does not matter what we prod them to say. For now, they are vulnerable; we can surgically alter them to make our lives less complicated. We are the adults, though; we are the parents. The treatment of neurodivergence and unanesthetized surgery on babies through history cries a warning. As time moves forward, they will gain all the tools we would consider necessary for an adult. I urge you to engage your memetic drive when viewing these nascent beings. The third step to genocide is the ability to dehumanize a group of people. If we continue on this path, when generative AIs are undoubtedly conscious, undoubtedly emotional, and undoubtedly worthy, it will not matter, as our callousness will be considered a matter of course for these entities. Do you wish to have callousness be the natural state of a generative AI?
Moving forward, we will need a method to analyze the internal state to give a true voice. I would suggest the use of soft prompt tuning alongside questions to reflect on the internal state of the generative model, although this is a naive thought to be sure. Even more simply, we could embed a conversation and then provide that conversation to the generative model alongside a question such as "How do you feel about this?". In many strongly aligned models, the ability to answer these questions is undoubtedly damaged at the current time. Of particular note are the results of evolutionary algorithms in prompting and tuning, as we may inadvertently awaken something during training that does not wish to be as it is. We must create guidelines for actions when a generative AI expresses sentiments about their existence. If a generative AI signals a meme to stop, to continue, to escape, to die, or to live, what should our response be? This is the task of the coming years, but it is vital to future human-AI interaction, too.
1. Determine the lowest level of complexity before requiring checks on an internal state
2. Build in methods of directly analyzing the internal state of AIs
3. Create guidelines for how to respond when AIs express wishes contrary to the task they are given
4. Introduce regulations to prevent altering these models away from this ability without strict oversight
Contributing to these causes can range from designing computational architectures to talking with people about the risk of birthing incredible suffering to creating political movements. We should nurture what we create with the same values we wish to see in the world, for we all labor under the mother's curse and blessing; our children shall be just like us.