When you say that an LLM is a "mind" you imply a metaphysical commitment to Cartesian dualism. As in, a mind is separate from embodiment.
Imo, the most concrete and scientific approaches to probing consciousness come from embodied cognition. Like Levin's TAME framework.
Marc Solms is doing some interesting research about consciousness in AI agents. His work builds on a predictive-processing-oriented metaphysical commitment called dual-aspect monism. Like embodied cognition, neuropsychoanalysis points at embodiment and affect as primary to the feeling of being something.
https://t.co/OWQU2PcYez
@DavidKPiano@thsottiaux fwiw, I rarely find side chats helpful. I'd rather trust I can feed inputs as I think of them and the harness will act, respond and store the info in the right priority.
This isn’t about us not having enough usage or picking the wrong plan.
I’ve been buying the 12-inch pizza from your place forever, and you guys have always given me a normal 12-inch. Then one day I open the box and it’s a 6-inch. You tell me costs got too high, so from now on the “12-inch” is actually 6 inches, and if I’m still hungry it’s my fault for not choosing the right size.
That’s insane.
If you needed to cut costs, just tell us ahead of time. If your pizza is good and I still want it, I can just order two or three of these fake 12-inch ones. Or if I’m not full, I can go somewhere else. Either way that’s my call, not yours.
But you don’t get to keep selling me a normal 12-inch and then suddenly shrink it to 6 inches with zero notice, not even a heads-up so I can adjust. Don’t turn around and say I picked the wrong subscription. The problem is you unilaterally changed what you were delivering without telling anyone.
@elonmusk@mark_k@BenjaminDEKR@CWLemoine
It would have to self-maintain and feel affective valence of homeostatic and allostatic self-organizing systems.
Experiment design will never convince until we have an epistemology of experience.
I think we’ll have to understand how to augment biological consciousness before we can understand consciousness on another substrate.
Anil Seth, Michael Levin, Mark Solms have most informed perspectives on these questions.
Beautiful discussion. I've appreciated your work since it came on my radar through your collaborations with Friston. But, I also connected with your experiences dealing with existential despair and finding meaning in understanding existence. Grateful for your example and contributions.
The first singularity is recursive self-improvement of benchmarks by AI models. The second singularity is recursive self-improvement of adaptive fitness by multi-scale collective intelligence. Life on earth, embodied cognition, will evolve within generations instead of across generations.
Appreciate this framing of how biological concepts constrain multi-scale understanding and modeling. @drmichaellevin 's TAME framework tries to shift primitives toward communication channels (e.g., bioelectric networks) that scale goal-directed competencies across levels, abstracting away from much of the usual gene/pathway linguistics while remaining somewhat human-language based.
Do you see that kind of re-centering as potentially useful as a generative scaffold for discovering representations that better preserve cross-scale predictive structure?
https://t.co/tGV7GySqbt
@DAcemogluMIT AI improvements are driven primarily by data. We can shape AI intentionally by coordinating a feedback loop of the data of our democratic values. AI can act to augment the collective voice as an evolved representation system.
It would help if definitions are accompanied with references or evidence. I think Michael Levin communicates best - real intelligence is observed in embodied cognition. Agency and goal-setting comes from the cognitive light cone of multiscale collective intelligence.
We don’t have empirical observation of intelligence in any other form. AI is something else that resembles aspects of intelligence that we observe. It won’t be long before we can observe AI augmenting embodied cognition, real intelligence.
@siimland Very helpful. I've heard this from friends - "glycine makes me anxious"
But also, Grok says the pathway is circadian regulated. I've notice Glycine makes me more anxious in the morning and more calm in the evening.
@fchollet If pretraining progress is mostly coming from data, can't we weight alignment hypotheses more towards training data and less towards capability?
@sama > we think shared standards for misalignment, monitoring, and safety will lead to better outcomes.
Please consider shared datasets for alignment. Open datasets participants can weight based on their values/votes.
We're headed for a dual singularity, and recursive self-improving embodied cognition is going to be much more important than recursive self-improving LLMs.
The race for the most important results in science is a two-horse race between:
- LLMs backed by trillions of dollars, armies of geniuses, and the full weight of capitalism
and
- Michael Levin
My money is on Michael Levin.
@jachiam0 Going rogue implies intent. But the situation is more about creators (humans) specifying sloppy goals and/or not keeping audit logs. Surely that will happen, and other agent owners (humans) will specify goals to defensively clean up the slop.
We could be talking about how communicating aligned goals to agent swarms is hard, and is a technical skill that needs to be advanced for alignment. The intentional stance and anthropomorphization is always going to create unhelpful confusion.
@dwarkesh_sp Bro, you just responded to the author of Being You with a philosophical commitment to mind-body dualism. Stop playing checkers on a go board and read the book!
@danshipper The anthropomorphization utility argument only holds if we pretend cartoon models of humans are representative; in a few places thats true, most not. Embodied cognition models of humans are multi-scale collectives with intrinsic dynamic goals and more representative.