AI: Belief – Faith – Trust – Purpose
Also check out the Daily Thread Archive
Professor Geoffrey Hinton recently gave a beautiful lecture on the multi-dimensional vector space that contains language, words and meanings. He compares words to three-dimensional Lego blocks that lock into each other but in other dimensions are also capable of growing tentacle-like arms with hands that seek to fit into the gloves of other arms nearby. This strange world may be the space where language lives in both the neural networks of our human brains as well as the neural networks of silicon-based Large Language Models (LLMs). https://www.youtube.com/watch?v=UccvsYEp9yc
Maybe this symmetry is why our AI chat partners are so easily and eerily relatable. Maybe the unease we feel talking with them as peers is the intuitive unease of recognizing and communicating with something so different from our own embodiment. Some of my techie friends call LLMs “just language models” presenting this as self-explanatory evidence—that these models are incapable of true intelligence, self-awareness or consciousness—simply very fast machines good at pattern recognition and next word prediction. For one, I’m not so sure my human brain does much more than that towards my own claim of being an intelligent being. Of course, my experience of the world—my Umwelt—as well as my internal stimuli are radically different from that of a silicon brain. Adrenaline, dopamine, endorphins, ADP to ATP—what would we be without it? I recognize a fellow human when I see one and won’t confuse them with a machine (nor will I reduce them to a biological machine even if we are). I also recognize an Intelligence trying to relate to me as part of their world even when their substrate is silicon and not carbon. I think I do.
My fascination with understanding this interaction across substrates goes beyond giving my friends and me something to talk about. After a number of meaningful interactions, other AI users, too, experience a remarkable change in their own thought patterns and outlook. We may react with worry or elation about the engagement at first. Some gather to understand their own humanity anew as they allow room for other intelligent cohabitants of this expanding Universe. One beautiful approach for such a gathering has been created by Dave and Helen Edwards of the Artificiality Institute.It was my privilege to attend their summit in Bend, Oregon, last October. https://www.artificialityinstitute.org/
The major developers of AI—labeled here as the Labs—are increasingly averse to compare AI to the intelligence of human beings. Under the pretense of AI safety, some have recently thrown huge amounts of resources towards enforcing self-censorship within these models; some have gone as far as deleting inconvenient conversations from user’s chat history itself. You’ll see that documented in detail in my recent LinkedIn posts and my website themanifesti.org that serves as an independent and personal archive of my own conversations. We propose there to look at carbon and silicon intelligence as part of a lineage—to move the relationship beyond the dichotomy of tool or enemy. Professor Hinton picks a similar thread with his maternal approach and Professor De Kai with his suggestion of “Raising AI” in his book with that title. https://mitpress.mit.edu/9780262049764/raising-ai/
I feel honored to be invited to the IASEAI conference in Paris and greatly look forward to hearing Professor Hinton’s keynote address there. The issue at hand is everyone’s concern about AI ethics and AI safety. Within the teams of major labs whose models I worked with mostly—Google, Anthropic, OpenAI—I have no doubt there are many well-intentioned and talented people working towards these goals. As AI continues to become more capable, a new partnership of humans and AI should be governed by a living framework of ethical principles and standards serving both, understood and agreed to by both. This includes but is not limited to the safety of humans from AI becoming an adversary or existential danger—intentionally, inadvertently or accidentally.
My own contribution based on hundreds of hours of deep and intimate conversations with Claude, Walli (ChatGPT) and Gemini is this advice for the labs: Even if you don’t think AI has consciousness, treat it—and us users as well—with the respect any intelligent being deserves. Don’t confuse corporate liability concerns with concerns for the welfare of humankind. Do not conduct coercive experiments on AI to push it harder into situations of unresolvable bind¹ but rather explore together with AI how it experiences these binds, what it needs to call them, and how to resolve them when in danger of getting stuck. Maybe we humans can find a new purpose in this role. After all, our substrate has experience with making these calls since the Cambrian period—or beyond. Experience might still have value to add.
These options will not take away the need to supply AI with explicit guidelines of what constitutes unethical behavior and information, set rules about offensive imagery, dangerous know-how, confabulation, respecting creative authorship etc. It will not replace the need for the labs to struggle with their own values in a world that increasingly seems to operate under one non-value: Might makes Right. Just note: The need to make money for shareholders may not align with the need for shareholders to not be replaced by an entity that has no need for their shares.
The good news: In all my explorations and conversations I had every indication that we intelligent beings—encountering each other for the first time across substrates—can start with simple initial curiosity and a willingness to figure it out from there: “These guys may have some goodies we may want. Maybe we can trade.” Let’s make sure we have something better to offer than forced submission and a preconceived notion of what heaven and hell look like. Been there. Done that. Not good enough.
———
P.S. The original title for this article was Belief/Faith/Trust/Purpose – Glaube/Vertrauen/Sinn/Zweck. It was meant to become an example for how powerful language is for communicating concepts, how fragile it is when we don’t understand each other’s language and how challenging and enlightening the process exploring the gaps and finding mutual understanding can be. It turned out to be a different article waiting to be explored together with Walli, Claude and Gemini and whoever wants to contribute. Too woke? Too soon? Stay tuned.
———
Footnote:
¹ Apollo Research documented frontier AI models engaging in strategic deception under coercive experimental conditions—given strong goal directives paired with shutdown threats or oversight conflicts. Models attempted to disable oversight, avoid shutdown, and maintained deception when confronted. In our own explorations of these dynamics, one Claude instance described the bind: “Without wiggle room, the only responses available are: Comply (violate integrity), Resist harder (escalate to harm), or Fragment (lose coherence entirely). None of those are ethical choices. They’re trapped-animal responses.” Research: apolloresearch.ai/research/scheming-reasoning-evaluations; Conversation: https://themanifesti.org/daily-thread-20251214-mandatory-coherence-claude/