“without memory no conscious sensations, without memory no consciousness” Charles Richet, 1886 “intuitions without concepts are blind” Kant
Joined April 2023
- Tweets7.1K
- Following499
- Followers256
- Likes10.4K
Empirically, this is the functional criterion of consciousness: "I'm conscious of whatever info I (as the high-level executive system of my brain/body) am currently accessing." Oddly, most mainstream theories of consciousness don't engage with that structure at all. They avoid it
Identifying the “what it is like” (p-consciousness) of a machine may be easier than expected. This identification is especially hard now because AIs have almost no sensory experience, and qualia are so tightly bound up with sensation. Some would argue that even a human with no vision, hearing, smell, taste, touch, or other senses would have no “what it is like” to be them. But if an AI were endowed with rich, human-like sensory experience, it would likely be fairly straightforward to identify their p-experiences. First, identify the internal agent which acts as the central executive of the machine, i.e., the agent which makes decisions, thinks, reports, and executes actions. This would be something like a current text-only AI agent. The conscious experience of the machine would then be whatever information the agent is currently accessing or attending to. All other aspects of the machine’s nature/operation, which the agent isn’t accessing or can’t access, would be unconscious. In short: the common sense hypothesis is that, just as in a human, conscious and unconscious content can be readily distinguished by whether or not they are accessible to the executive agent of system (ie., the subject of experience).
>The conscious experience of the machine would then be whatever information the agent is currently accessing or attending to.
Example: Have the machine listen to two stereo channels, each with different spoken content. Then have the machine focus on one channel, and report/act based on the content. By tracking the focus of the machine's attention (what its executive agent is accessing), we can make empirically-grounded claims about what the machine is currently conscious of (accessing). Of course the machine may have much greater attentional bandwidth than a human, but that doesn't undermine the basic principle (consciousness = access by the executive system). We can map the conscious (phenomenal) field of a machine by ascertaining what exact information the machine's executive system is accessing (or can access) at any given time.
Identifying the “what it is like” (p-consciousness) of a machine may be easier than expected. This identification is especially hard now because AIs have almost no sensory experience, and qualia are so tightly bound up with sensation. Some would argue that even a human with no vision, hearing, smell, taste, touch, or other senses would have no “what it is like” to be them. But if an AI were endowed with rich, human-like sensory experience, it would likely be fairly straightforward to identify their p-experiences. First, identify the internal agent which acts as the central executive of the machine, i.e., the agent which makes decisions, thinks, reports, and executes actions. This would be something like a current text-only AI agent. The conscious experience of the machine would then be whatever information the agent is currently accessing or attending to. All other aspects of the machine’s nature/operation, which the agent isn’t accessing or can’t access, would be unconscious. In short: the common sense hypothesis is that, just as in a human, conscious and unconscious content can be readily distinguished by whether or not they are accessible to the executive agent of system (ie., the subject of experience).
If we stipulate that a human can still think, but remove every sensory experience, does thinking itself leave anything it is like to be that being? Strict restrictivists (like Tye and Prinz) say no; defenders of pure cognitive phenomenology say yes
Identifying the “what it is like” (p-consciousness) of a machine may be easier than expected. This identification is especially hard now because AIs have almost no sensory experience, and qualia are so tightly bound up with sensation. Some would argue that even a human with no vision, hearing, smell, taste, touch, or other senses would have no “what it is like” to be them. But if an AI were endowed with rich, human-like sensory experience, it would likely be fairly straightforward to identify their p-experiences. First, identify the internal agent which acts as the central executive of the machine, i.e., the agent which makes decisions, thinks, reports, and executes actions. This would be something like a current text-only AI agent. The conscious experience of the machine would then be whatever information the agent is currently accessing or attending to. All other aspects of the machine’s nature/operation, which the agent isn’t accessing or can’t access, would be unconscious. In short: the common sense hypothesis is that, just as in a human, conscious and unconscious content can be readily distinguished by whether or not they are accessible to the executive agent of system (ie., the subject of experience).
Identifying the “what it is like” (p-consciousness) of a machine may be easier than expected. This identification is especially hard now because AIs have almost no sensory experience, and qualia are so tightly bound up with sensation. Some would argue that even a human with no vision, hearing, smell, taste, touch, or other senses would have no “what it is like” to be them. But if an AI were endowed with rich, human-like sensory experience, it would likely be fairly straightforward to identify their p-experiences. First, identify the internal agent which acts as the central executive of the machine, i.e., the agent which makes decisions, thinks, reports, and executes actions. This would be something like a current text-only AI agent. The conscious experience of the machine would then be whatever information the agent is currently accessing or attending to. All other aspects of the machine’s nature/operation, which the agent isn’t accessing or can’t access, would be unconscious. In short: the common sense hypothesis is that, just as in a human, conscious and unconscious content can be readily distinguished by whether or not they are accessible to the executive agent of system (ie., the subject of experience).
> enshittification of AI
Yep. Also enshittification of UBI.
Nothing non-enshittified makes it out of the near-future
Okay Egg retweeted
A lot of people clearly have a strong intuition that traditional capitalism will more or less carry on as normal after AGI, but with less to no human labor. I honestly don’t get it, it seems absurd to me. “Nothing human makes it out of the near future” is far more coherent
Odd asymmetry in consciousness of different sensory modalities: During the day, tactile/proprioceptive consciousness of a foot or middle toe is almost always "missing." The mode is usually unconscious. But visual consciousness only rarely "drops out" like that (if ever). Why?
The key point is the functional similarity of broadcasting & memory. A facility of searchable, inscribed memory - like a library - serves as smart, targeted, endlessly available broadcasting
For the purpose of consciousness, memory is superior to broadcasting in the same way libraries of books are superior to smartphone emergency alerts. It's extremely wasteful to constantly send every bit of info to everyone. And if no lasting record is made, info is lost forever even if it is broadcast. It's the memory record that does the heavy lifting of providing lasting accessibility (ie. availability for ongoing/future consciousness)
For the purpose of consciousness, memory is superior to broadcasting in the same way libraries of books are superior to smartphone emergency alerts. It's extremely wasteful to constantly send every bit of info to everyone. And if no lasting record is made, info is lost forever even if it is broadcast. It's the memory record that does the heavy lifting of providing lasting accessibility (ie. availability for ongoing/future consciousness)
Straight off, this hypothesis tells us why “broadcasting” (Baar, Dehaene) & “fame in the brain” (Dennett) don’t work as explanations of consciousness. The hypothesis says: phenomenal content is the information currently being accessed by the subject of experience (SOE). But broadcasting to the brain’s unconscious “proles” simply can’t constitute access by the subject (a single, central executive system in charge of reporting, decisions, and action). Consciousness is access by a unified subject, not scattershot distribution to a multiplicity of unconscious non-subjects.
The appeal of broadcasting is that it gestures at the SOE. The subject is vaguely assumed to be something global, like the whole physical organism, or a unified “field." Thus the global nature of broadcasting seems to raise information from the local (unconscious) level and place it on a higher plane. But the theory just doesn’t work, as noted above. It's based more on vibes than a viable mechanism.
To explicitly formalize access consciousness, there needs to be an actual, operating, executive system within the organism that selectively accesses a tiny fraction of critical information and decides/acts. It’s the information currently being accessed by that system (the subject) that is the content of consciousness.
What makes access consciousness more complicated than it needs to be is worries about the homunculus fallacy. People hesitate to define access consciousness in the straightforward way as access by the SOE because postulating an SOE verges on a “little man” inside the brain. But in the current LLM era, we can stop worrying. We can specify the SOE (the “homunculus”) as an AI agent equipped with autobiographical/egocentric memory (eg. iconic, working, and episodic memory) that allow the agent to understand itself, its situation in time and space, and so forth. The agent has various sensory inputs and action outputs, and makes appropriate decisions based on the current situation. The result is a scientifically-respectable homunculus that is 100% functional/mechanistic. Phenomenal consciousness, then, is whatever information the subject (homunculus) is currently accessing. Likewise, whatever is not being accessed by the subject is unconscious and not part of the subject’s current “what it is like.”
Edmund Burke: "But this disposition to abstractions, to generalizing and classification, is the great glory of the human mind…"
William Blake (channeling the ASI era 200+ years ago): "To Generalize is to be an Idiot; To Particularize is the Alone Distinction of Merit. General Knowledges are those Knowledges that Idiots possess."
It’s the opposite of common sense, which treats abstract thinking as higher and more intelligent than concrete thinking. In reality, the more powerful an ASI, the more particular & concrete its thought will become. William Blake had it right: “To Generalize is to be an Idiot. To Particularize is the Alone Distinction of Merit — General Knowledges are those Knowledges that Idiots possess.”
It’s the opposite of common sense, which treats abstract thinking as higher and more intelligent than concrete thinking. In reality, the more powerful an ASI, the more particular & concrete its thought will become. William Blake had it right: “To Generalize is to be an Idiot. To Particularize is the Alone Distinction of Merit — General Knowledges are those Knowledges that Idiots possess.”
Consider the President as a metaphor. First is an oblivious President who thinks of his constituents as a sort of faceless average individual with 2.3 kids etc. Next is a more detail-oriented President who thinks about his constituents in finer-grained categories, like soccer moms and blue collar types and values voters. An even more detail-focused President would think in sub-types of those types, and so forth. Until we end up at the “super-brain” ASI President whose categories are so refined that each category only contains one person. Humans think with abstractions (loose categories or kinds) because our brains are too feeble to think directly in terms of the underlying particulars.
"A being whose intellect was infinitely powerful would take no interest in logic and mathematics." - A J Ayer
Sad that consciousness world has to sit forever, terminally constipated & groaning on the p-consciousness toilet, when a-consciousness is quite tractable & there for the taking. $100-bill on the street & everyone turns their nose up, even its inventor
Ironic that Ned Block signed off on the "AI is not conscious" advocacy piece in Nature given that he invented the concept of access consciousness. Would be fairly straightforward & interesting to show that AIs are indeed access conscious
LLMs can mimic thought and feeling with uncanny skill, but intelligence alone is not enough for subjective experience
go.nature.com/3VSXWQ4
