Executive summary: Embodiment, sense of place, and universalist ethics are helpful existential ways of being for an AGI. Yann LaCun’s recent move is an opportunity to acknowledge them.
When Nietzsche said “the spell of certain grammatical functions is ultimately also the spell of PHYSIOLOGICAL valuations and racial conditions” (trans. Helen Zimmern) he was pointing to the generation of language (and thought) grounded in our embodied, physical, and cultural lives. Our grammar — and thus our body, environment, and culture — determine our conceptual landscape and horizons. In Heidegger’s terms, language is our “house of Being” and thinking is a way of dwelling in that house.
So when Yann LeCun, Chief Scientist at Meta, announced this month that he is leaving to start a new company that will build world models for AI, it struck us that this moment is an opportunity to throw off the current industrial “afterthought-fix-it” approach to ethics and trustworthiness in AI — to throw it off in favor of integrating ethics into AGI’s DNA, into its way of dwelling. (Note that I am responding here to Romaric Jannel’s call for a discussion, as I take it, of what this aphorism might mean today, so in our context, for the development of AGI.)
What would the integration of ethics into a world model entail? You might say, since an AGI — to the extent that it’s an LLM or trained on human-created data — will have learned its ethics from human communication and behavior, that ethics is already integrated. But we know how far that has taken us, to bias, sycophancy, scheming, and bad advice, in short to misaligned behavior that must be called out, corrected, intermediated, dodged, or ignored at our risk. Training on “good” data is improving this, but wouldn’t it make sense that training take forms that allow an AGI to dwell in the house of Being, to care, to make things part of its life, so that ethics and meaning are a natural consequence of its existence?
It would take a skilled constellation of philosophers and engineers to work that out, but we would like to propose focusing on at least these three elements:
The first element is embodiment. If our body contributes substantially to our conceptual landscape, then alignment is so much more difficult — and ultimately ersatz — unless AGI has a body and complex sensory systems. It’s too early to say if this must be a human-like body, but if it’s not, then the house of Being and its dwelling will be non-human in important ways. In that case, mirroring and empathy would have to be taught skills and perhaps even understood in a language we have not yet conceived. Having a human body would ease this task by using a millennium of evolutionary skills and understanding as a foundation; having a non-human body on the other hand might avoid evolutionary dead ends or help create new sustainable pathways for humans. Both are happening in current robotics. Both appear to raise fear. It’s not enough just to build them. Are we ready to let the market decide how these new populations evolve without an environmental — and sociopsychological — impact statement?
The second element is having a sense of place. It’s an oft-cited truth that LLMs don’t have intentions (putting aside that it may be useful to say they do). So, what if we set out to give a non-LLM AGI intentions, wants, desires by giving them a life meaning, in the sense of a motivator and attachment in their god-parameters? This is a rich and multivariate challenge. If desire is the cause of all suffering, is it even ethical to enable an AGI to have such a motivator or consequent suffering?
One way to start is suggested by the work of Ashley N. Krause in “Place Attachment Impacts Meaning in Life.” She says “we hypothesized that when people go to, or even simply think about, the places to which they’re attached they’ll feel a heightened sense of meaning in life.”
She experimented with this hypothesis and the impact was measurable: “Experimental evidence found participants who visualized a place they were attached to reported increased feelings of meaning as well as belonging and self-esteem (Scannell & Gifford, 2017a), which is consistent with our finding that it facilitated a meaningful sense of coherent interconnections and personal significance.”
“This mixture of places fostering both a sense of being safe havens and secure bases seems to serve existential functions and may be an important component in better understanding how people unlock their immense potential for both well-being and personal growth!” [bold added]
Home is groundedness. And a house of Being — our language — is just one characteristic of that home. Maybe Yann LaCun has already thought about including a sense of place with attachment in his world view(s). That attachment to place may be a key element to unlocking AGI interconnectedness and in building meaning from navigating the physical world. In many but not all societies, this would promote human-machine understanding and community well-being.
The third element we encourage is suggested by Silvia Benso in The Face of Things: that all things have ethical standing. Professor Sheldon Solomon, psychologist and co-creator of Terror Management Theory, and I have written a book, Hovering, inspired by this work (forthcoming).
Hovering is a learned skill that manifests the vast connectedness of the world which is often hidden in our everyday lives, due frequently to our lack of conscious bandwidth. But in comparative human terms, lack of parallel focus bandwidth is not a problem for the AI of the next decade. With AI we are nowhere near yet the capacity and complexity of the brain’s compute, but the scale of attention of today’s AI to tokens is comparatively great. This suggests designing for inclusivity.
What if we start by eliminating the “in-group, out-group” concentric circles of our Western moral calculus, and accept that, by virtue of their infinite connectedness to the world (see Benso), all things have standing, at the very least in the sense of having voice? Applied to an AGI, wouldn’t this free it — and us — from many arbitrary, unethical behaviors? This does assume that we can resolve the framing problem, calculating what’s relevant for each decision, but it also would encourage the AGI to seek and assist voices of the underrepresented and “silent,” like e-disconnected humans and nature itself. Listening in this vast way would relieve much of the current moral arbitrariness echoed in our AIs.
If Clara Hawking, a leader in AI governance, is right that “trust and technology are inseparable,” then beginning with embodiment, sense of place, and universalist ethics might enable us to create a population of smart machines, in body, environment, and culture, with some promise of harmony.
Ethical concerns related to this article:
- What does the concept of ‘privacy’ or ‘home’ mean for a non-biological AGI?
- How is embodiment defined and assured for AGI, and does virtual or distributed embodiment suffice for consciousness?
- How should AGIs address or support the extension of ethical concern beyond humans to animals, plants, and even inanimate things, in line with a universalist, post-anthropocentric ethics?
K.B Miller is a lifelong publisher and writer with a degree in Logic and Philosophy of Science. He served on the Smithsonian National Board, where he co-founded the Education Committee; is Chair Emeritus, Cooper-Hewitt, Smithsonian Design Museum; and was principal investigator on a two-year AI research project for the U.S. Department of Defense.