Nicely done, and the neat/scruffy framing is the right spine for it.
The thing I keep catching is an assumption Friston and Summerfield seem to share without arguing for it: that embodiment is necessary. Friston wants active inference, a system acting on the world, not passively predicting it. Summerfield wants bodies, recurring needs, motivation. Different routes, same premise underneath: the sensorimotor loop is load-bearing for genuine intelligence.
But it's worth asking what the body is actually contributing there. A sensorimotor stream is, in the end, information about the body, signals, not the body itself. If that's right, then privileging embodiment looks like a fact about one kind of realizer's input rather than a condition on intelligence as such. The body is a rich and well-structured information source, not obviously a different category from any other stream a system has to predict.
I don't think the debate settles this, because neither side puts the assumption on the table to defend it. Which is maybe the more interesting question hiding under the neat/scruffy one: is embodiment a requirement, or just the particular way the one example we know of happens to get its information?
You’re right to ask this question, for sure embodiment play’s a part in Chris’ most recent book, he talks about the main difference between human and AI learning is type of input. Humans learn through vision and touch and they understand tone of language, extra channels of information current llms don’t have. I don’t think consciousness is needed for intelligence, maybe there’s parts to your question I haven’t fully understood. Does this cover it and what do you think?
Yes, that's the gist of it, and your examples are the right ones. Here's my angle: vision, touch, tone are extra channels, but each one is still information. Vision is information about light hitting the retina, touch is information about pressure on the skin, tone is information about a speaker's state. The body doesn't reach the brain, it's the signals about the body that do. The brain only receives the stream, never the world itself.
So humans clearly have more channels, and richer ones, than a text-only model, no argument there. My question is whether that's a difference in amount or a difference in kind. If it's amount, then a body isn't a requirement for intelligence, it's just the particular, very rich set of inputs that one kind of system, us, happens to have. A different system with different inputs could still be running the same process.
That's the assumption I think Friston and Summerfield both leave unexamined: they treat the body as necessary, but they might only be describing how the one example we know of happens to get its information. Curious what you think from the neuroscience side, since that's where the answer probably lives.
reading Sapolsky and watching an old Lex Fridman podcast with Andrej Karpathy, something clicked that i want to add to this debate. neuroscientists have already documented it — prefrontal metastability is not a transitional state, it is the working regime of intelligence. a fully synchronised oscillator system is locked — no new information enters. intelligence lives in incomplete synchronisation, where the error signal is nonzero and still changing. both Friston and Summerfield describe what intelligence optimises — neither describes what it preserves. the invariant is not the solution. it is the productive distance from it. a system that has converged has stopped being intelligent. it became a record ⊛
One of the most interesting parts of this discussion to me is how often intelligence seems less connected to having information itself and more connected to how an organism or system navigates uncertainty, context, and adaptation over time.
Amazing read! Thankyou for sharing. It also seems to me that the discussion had the essence of whether we need one system to solve all ( a general model, AGI) or a bunch of systems solving individual pieces( a mixture of specialist models, ASI) to achieve genuine intelligence.
Minimizing surprise is certainly common among the less alive and.old people but in the most alive and intelligent people, adventure seeking , their life choice exponentially increase surprises. Minimizing surprises is for the ensurance business not for the living ,at least not for what is exciting in life.
Nicely done, and the neat/scruffy framing is the right spine for it.
The thing I keep catching is an assumption Friston and Summerfield seem to share without arguing for it: that embodiment is necessary. Friston wants active inference, a system acting on the world, not passively predicting it. Summerfield wants bodies, recurring needs, motivation. Different routes, same premise underneath: the sensorimotor loop is load-bearing for genuine intelligence.
But it's worth asking what the body is actually contributing there. A sensorimotor stream is, in the end, information about the body, signals, not the body itself. If that's right, then privileging embodiment looks like a fact about one kind of realizer's input rather than a condition on intelligence as such. The body is a rich and well-structured information source, not obviously a different category from any other stream a system has to predict.
I don't think the debate settles this, because neither side puts the assumption on the table to defend it. Which is maybe the more interesting question hiding under the neat/scruffy one: is embodiment a requirement, or just the particular way the one example we know of happens to get its information?
You’re right to ask this question, for sure embodiment play’s a part in Chris’ most recent book, he talks about the main difference between human and AI learning is type of input. Humans learn through vision and touch and they understand tone of language, extra channels of information current llms don’t have. I don’t think consciousness is needed for intelligence, maybe there’s parts to your question I haven’t fully understood. Does this cover it and what do you think?
Yes, that's the gist of it, and your examples are the right ones. Here's my angle: vision, touch, tone are extra channels, but each one is still information. Vision is information about light hitting the retina, touch is information about pressure on the skin, tone is information about a speaker's state. The body doesn't reach the brain, it's the signals about the body that do. The brain only receives the stream, never the world itself.
So humans clearly have more channels, and richer ones, than a text-only model, no argument there. My question is whether that's a difference in amount or a difference in kind. If it's amount, then a body isn't a requirement for intelligence, it's just the particular, very rich set of inputs that one kind of system, us, happens to have. A different system with different inputs could still be running the same process.
That's the assumption I think Friston and Summerfield both leave unexamined: they treat the body as necessary, but they might only be describing how the one example we know of happens to get its information. Curious what you think from the neuroscience side, since that's where the answer probably lives.
reading Sapolsky and watching an old Lex Fridman podcast with Andrej Karpathy, something clicked that i want to add to this debate. neuroscientists have already documented it — prefrontal metastability is not a transitional state, it is the working regime of intelligence. a fully synchronised oscillator system is locked — no new information enters. intelligence lives in incomplete synchronisation, where the error signal is nonzero and still changing. both Friston and Summerfield describe what intelligence optimises — neither describes what it preserves. the invariant is not the solution. it is the productive distance from it. a system that has converged has stopped being intelligent. it became a record ⊛
One of the most interesting parts of this discussion to me is how often intelligence seems less connected to having information itself and more connected to how an organism or system navigates uncertainty, context, and adaptation over time.
Yes 100%, it seems you have more of the scruffy point of view, which I agree with too, i feel its more practical
I think so too. Maybe because real life is usually messy enough that adaptation ends up being just as important as knowledge itself!
Woah, That's new year to me Thanks!!
Amazing read! Thankyou for sharing. It also seems to me that the discussion had the essence of whether we need one system to solve all ( a general model, AGI) or a bunch of systems solving individual pieces( a mixture of specialist models, ASI) to achieve genuine intelligence.
Minimizing surprise is certainly common among the less alive and.old people but in the most alive and intelligent people, adventure seeking , their life choice exponentially increase surprises. Minimizing surprises is for the ensurance business not for the living ,at least not for what is exciting in life.