All experiences are abstractions. You look at a squirrel, you’re not experiencing a squirrel - you’re experiencing a bunch of neural impulses.
While your right in general, the case I described did consist of a human learning (or maybe deciphering is a better term) a language using only text. In a script he knew the phonetic values of to be sure, but it is not obvious whether that helps.
I think that understates the amount of inference a fair bit. As I understand it the chain of logic was something like:
- The number of characters made it a good guess that it was a syllabary, so each character would likely represent a consonant sound followed by a vowel sound.
- There are bunch of words like “X” where “X” is a single character, which were interpreted as various tenses/cases of the same word.
- In each of those words, the “consonant” part of X is assumed to come from the stem, while the “vowel” part is assumed to come from the suffix. So different characters that attach to the same stem are likely to have the same consonant with different vowels, and different characters that attach to the same suffix are likely to have the same vowel with different consonants. That way you can make a grid of consonant/vowel positions for each character even if you don’t know specifically what the consonants and vowels are.
- Guess that a repeated triplet of characters all with the same vowel but different consonants represents ko-no-so, and refers to Knossos, a big site in the area. Use that to start assigning real sounds to your abstract syllable grid.
- After you get a few words, think, “oh yeah, this is something like ancient Greek”. Which I understand was a hypothesis before the decoding, but it wasn’t the only one.
Steps 1, 2, and 3 can be done without a guess about the language (but some general guesses about how human language works) and probably could be inferred with some LLM assistance (or more conventional software for that matter).
Step 4 still might still not require a hypothesis about a specific language if you assume that place names tend to be “sticky” even if the culture around them changes. It’s only at step 5 where you really assume a target language.
(One thing that I read that was apparently an argument from before decipherment for why linear b couldn’t be Greek is that Greek is lousy with words that end in “σ” while linear b text doesn’t have any overwhelmingly common ending character. But it seems like in linear b orthography, you just omit terminal sigma as given)
Sight is an abstraction. Touch, taste, and smell are not abstractions; the sense organs come into direct contact with the objects of sense. Taste is the most direct: the tasted stuff becomes one with your body chemistry. Smell is slightly more abstract because you only contact tiny fragments. Hearing is more abstract than those three, but still more tactile than sight.
This is the meaning of the phrase “ineluctable modality of the visible” pondered by Stephen Dedalus in Ulysses by James Joyce. You never see a thing itself, you only see a representation of it constructed in brain pixels.
This really isn’t true from either a neurological or physical standpoint. Taste and smell (which are tightly related) are very subjective and produce highly abstracted notions of what is producing them, and different environmental stimuli, emotional states, or the activity of pathogens or other substances can alter the sensation of taste and smell of the same substance in the same individual, as pregnant women and people with SARS-CoV-2 infections can attest. Touch is even more of an abstraction; as the parable of Plato’s Cave suggests, the perception of what one senses by touch is a complex synthesis of different nerve impulses, and in some cases a mental construct of ‘touch’ in the absence of actual contact as amputees often report. Sound is enormously abstract and everyone has had the experience of hearing something that wasn’t there or mistakenly identifying the source of a sound. Light impinging upon the eyes is at least in a direct line although what the eye actually sees can differ substantially depending on where light emitted or reflected from an object hits on the retina and what kind of movement or change it is experiencing.
All of this sensory data is reported back to the brain via the peripheral nervous system (except for eyesight as the optic nerve is actually a direct extension of the brain) in the form of signals that are by definition abstractions that require integration and interpretation to form a cohesive perception of the environment. @Mangetout is correct at all experiences are actually a synthesis of sensory inputs with the brain ‘filling in’ where there is some gap of information, which also frequently produces optical and aural illusions; indeed, there is a broadly held view in neuroscience that ‘objective’ experiences are actually hallucinations that are stitched together by the brain from sensory information coming in and processed at different rates to provide the simulacra of continuous experience.
Language models, of course, don’t have any sensory experience because they are not embodied within the domain of the physical world in any way but even setting aside the discussion about whether embodiment or other forms of sentience are really required for true ‘intelligence’, these models simply do not have a continuous set of internal, introspective processes that would produce the processes of consciousness or cognition as it is generally understood; they are prompt and response machines that take a prompt and run it through the complex algorithmic system of an artificial neural network and produce a statistically cromulent (if often factually inaccurate) response,
‘Agentic AI’ (when and if that actually becomes workable) may have some argument to being closer to something akin to consciousness (although lacking in the evolved layers and systems of internal cognition that brains perform) but the dirty not-quite-a-secret about attempts to make an agent with some kind of continuous processes is that the current ANN approach becomes computationally unwieldy and then unstable if it is cycled indefinitely. To the extent that we understand the workings of human consciousness, one thing we know is that it is able of switching between and managing different levels of ‘attention’, which can be a problem when trying to focus on analytical problems or highly detailed tasks but also makes it extremely flexible in that it allows the user to perform multiple simultaneous actions at different levels of consciousness and abstraction, or put in a common colloquial phrase, to “walk and chew gum at the same time”, which may be a key to both managing the mental loading of conscious processes as well as being able to interact gracefully with the complexities of the natural environment.
Stranger
The point is that the actual substances enter and are absorbed into the body. Unlike the other 3 senses.
Actual photons strike the retina and are absorbed by photoreceptors. Actual acoustic fluctuations in air pressure impinge upon the tympanic membrane and cause it to stretch and contract. These are just as physically ‘real’ as molecules interacting with taste and olfactory receptors, and in any case all of this information is encoded as somatosensory signals that are transmitted through the peripheral nervous system (except, again, for the optic nerve, but even those signals comprise many different types of data delivered at different rates and with a lot of ‘pre-processing’ at various stages before being integrated in the visual cortices). All of it is abstract in the sense that the executive functions of the brain receive a synthesized ‘experience’ of the ambient environment.
Stranger
This is starting to become a tangent to the main thread, but, briefly—
The Romans did not have the same individualized relationship with their gods the way we conceptualize it today. They were a collective society, and their religious practices were about the relationship of their collective (their tribe, village, town, city, however they defined themselves) with the gods generally, or with a subset of the gods with whom they had established patronage (for more, look up pax deorum). In their mythology, the gods were preoccupied with their own internal struggles as much as they were concerned about human affairs, so public worship was oriented toward winning the gods’ favor for the collective.
An ordinary individual person was considered to be beneath the gods’ notice most of the time, so they didn’t fear judgment and retribution in the same sense as we understand it now. It’s true that on a smaller scale, a Roman house would have its own shrine (cf. lararium) where family members could perform rituals and appeal for protection, but again this was largely oriented toward protecting the household, not individuals.
This lack of individualized judgment is also seen in their afterlife. They had a “punishment plane,” Tartarus, but it’s not equivalent to the modern Christian concept of Hell. People would be sent there for wickedness in life, but the sins are social, transgressions against order. (Read Virgil to start.) It’s also a temporary state, not a plane of eternal damnation; people do their time, correspondent to their crimes, and then pass on.
I’ll stop there, but I hope it’s clear that the ancient Roman mindset is pretty different, and our inherent assumptions about religious morality simply do not apply. And therefore (remembering the thread) that helps illustrate how and why translation can be such a challenging and subjective task. It’s not just about grammar, or vocabulary; it’s about the psychology embedded in the very language.
Undoubtedly, but the point is that the decipherment of Linear B wasn’t performed in a vacuum. There were external contexts against which conjectures could be tested and verified.
Certainly that’s a difference between biology and technology, at least in the cases we’re discussing, and there can’t really be any sense in which ChatGPT is sitting there thinking to itself while nobody is prompting it, but from its own perspective (if for the sake of argument, it has one), it is continuous - because the moments in between the processing of prompts simply don’t exist, from the machine’s angle of view.
There are loose analogues to this in biological minds too. If your head and eyes are immobilised and nothing changes in your field of view, you will experience blindness; your visual system needs things to be changing, in order for them to be perceived, as well as this, there is a lot of gap-filling in the synthesis of perception where our brains present us as having seen a thing happen smoothly and continuously, when in fact it was composed of a series of discrete static snapshots
Nobody can know the inner life of a human or a machine. All we can really do is speculate about it based on how it behaves with respect to outputs and inputs, which are the normal criteria we use to judge human understanding, and on this basis it’s clear LLMs have little if any understanding of anything.
Of course you can speculate that LLMs have some kind of deep consciousness that’s too inscrutable for anyone to see. Since inner awareness is by definition unknowable, you’re certainly free to speculate that’s the case, but that’s just your bias and your opinion.
I would just say that in general, human-like intelligence and consciousness are held to be emergent phenomena, and emergence is a feature of complex systems, and complex systems are characterized by the presence of rapid feedback loops that help refine and amplify their inner patterns. So by that definition LLMs, don’t possess the fundamental precursors of consciousness/awareness/intelligence, whatever you wish to call it, so it would be very surprising to see it display such traits, and in fact they don’t. Some people think they see it, but people will also see Jesus’s face in a piece of unevenly toasted bread.
I’ve noticed similar things with bilingual people. My take is that children don’t acquire language the same way that adults do. Adults will usually spend a much larger fraction of time learning how the language works, prioritizing areas of learning, and how it maps to their mother tongue etc. With children it’s just a stimulus/response kind of training. They have 2 mother tongues, not a mapping between 2 languages.
So for a bilingual adult, unless they’re specifically trained to the task, or spend a lot of time thinking about the differences, they struggle to navigate the difference between the 2 languages. They may have no significant grammar training in either language. Their range of human experience in each language may give them unequal vocabulary distribution and literary sense in each language.
YMMV but I’ve learned via experience, there’s often not much benefit in asking a bilingual-from-childhood person to help bridge the literary or grammatical sense of either language. Native-level bilinguality is a state of having 2 separate skills, for example English and Spanish, and not one integrated skill of English-Spanish mastery.
That is what you described, yes. But I hope that you won’t mind that I take your literal friend-of-a-friend account with a very large grain of salt.
Even if LLMs are similar to human minds in how they work (a very contentious question, and very difficult to answer for sure), it’s certain that they arrive at that state through a very different process than humans do. As you say, LLMs have much less feedback and recursion than humans do, but they make up for it by using a much larger corpus than any human possibly could.
It reminds me somewhat of the status of astronomy, as compared to other sciences. In most sciences, we take some system, and we study it, and we figure out interesting ways it could be different, and then we change the system slightly in some particular way, and study it again, and so on. But in astronomy, the things we study, we can’t really change. There’s no such thing as “experiment”, per se, in astronomy. Instead, we have to find some other system, that already exists, that we think is similar to the first system but different in approximately the way we’re interested in, and it’s only the vast number of systems available to be studied that makes the science possible at all.
Except there are no processes within the chatbot that are introspective or that would provide any interior sense of continuity; you can ask it what it ‘thinks’ about its last response or how it ‘feels’ about some issue and it will give a response that simulates what a sapient mind might say (based upon the training from its textual corpus) but there are literally no interior ‘looping’ processes that provide any functionality for introspection; it is just giving a stochastically likely response to a common query.
Or a face on Mars.
Karen Hao’s Empire of AI describes her impressions of the religious fervor that both AI advocates and ‘doomers’ who both believe that AGI is around the virtual corner and on the cusp of upending and revolutionary society (for fantastically good or catastrophically bad, respectively) without any factual, scientific basis for this belief. She even observed Demis Hassabis (Nobel Laureate in Chemistry, founder of what is now Google DeepMind and CEO of Isomorphic Labs which uses AlphaFold for novel drug discovery) ‘code-switching’ between being an objective AI researcher critical of the capabilities of AI to be a general intelligence and then switching into advocacy mode of an AI company executive hawking hype-filled pledges of how the technology will change society.
In particular they may have difficulty in translating colloquialisms and other culturally-grounded semantics between languages because they have an internal concept of what some turn of phrase means that is embedded in one culture for which there is no analogue in the other, and unlike a non-native speaker who will have learned some analogy to describe that concept or idiom, the native bilingual speaker knows that any translation will be insufficient to capture the nuances. An example of the that I came across years ago with a natively bilingual Swedish speaker is the common phrase, “Att glida in på en räkmacka”, which is explicitly translated as “To slide in on a shrimp sandwich.” When pressed for then meaning of such an odd expression, the he explained that it roughly means when someone feels overly privileged or has unreasonable expectations based upon their brief involvement but the there is actually an entire social dynamic associated with the phrase that don’t really make sense in either British or American English that makes it an accusative and somewhat insulting phrase. And this is a language that is linguistically close (in structure and lexical distance although not in phonology) to English with substantial continuity of cultural fundamentals. Someone who natively speaks, say, Tagalog and English probably has two completely different linguistic models in their head for describing the world.
Stranger
There literally are. The prompt and the response becomes part of the attention window. When an LLM answers your second prompt, its previous answer to you is data that it includes in the process of generating the current one.
That’s why the guardrails these things typically have are given to them as prompts in natural language.
Of course this isn’t how the human brain works, and I’m not arguing that it does, but there is definitely introspection or self-reflection in an ongoing conversation with a typical LLM
Only as part of responding to a subsequent prompt, not a continuous or even periodic cycle of introspection and examination (or fabulation) of sensory and other inputs, e.g. what Douglas Hofstadter refers to as the “strange loop” of consciousness to describe the navigation of processes through the hierarchical system of the brain, eventually returning back to the same attentional frame. Even if there were some potential for a chatbot to be self-aware on any level, this would analogously be the difference between watching a slideshow and a film, and nobody would consider a set of individual slides to to be an even approximately continuous experience.
Which leads to another issue; for those that believe a chatbot has some level of consciousness which is somehow emergent from its ‘context window’, every time a user shuts down a chatbot session they are permanently quenching that hypothetical consciousness which is an equivalent to unjustifiable killing. Creating and destroying potentially self-aware intelligences purely for utility is certainly a moral hazard in any egalitarian system of belief (notwithstanding the industrial system of animal husbandry to produce food without concern for the well-being of the stock animals which provide it).
Stranger
[duplicate post]
I’m not sure this analogy works because a film literally is a slideshow - they are both just a series of discrete frames.
If films were sentient, they would not know about the shutter interval between frames and if slideshows were sentient, they would not understand why you think of them as different from films, because the interval between frames is nonexistent from their viewpoint.
I’m not arguing that these things are exactly like humans or that they think exactly like humans, but I think a lot is being summarily discarded by ‘all they do is just…’ arguments.
Birds can fly. Planes can’t fly because only birds can fly. When you see a plane appearing to fly, all it’s really doing is just a crude mimicry of what a bird would do in that scenario. It’s not really flying - it can’t be, because planes don’t have feathers and beaks.
Photons and acoustic fluctuations come from effects of the sense object on its environment but are not the thing itself. The whole point I’m making is that taste and smell require bits of the sense object itself to enter the body. This is a simple concept, not the complicated seminar you’re making it into.