not GP but this seems like quite an attractive idea that many people have reached: a brain of a given "complexity" cannot comprehend the activity of another brain of equal or higher complexity. I'm positive I'm cribbing this from scifi somewhere, maybe Clarke Or Asimov, but, it's the same idea as the Chomsky hierarchy, and the Godel theorems seem like a generalization of that to general sets of rules rather than mere "automata".
For example, you can generalize a state automata to have N possible actors transitioning state at discrete clock intervals. The machine never terminates until all actors have reached a termination state. That machine is probably impossible to model on any kind of a Turing machine in polynominal time. And a machine that operates at continuous intervals is of course impossible to model on a Discrete Neural Machine in polynomial time.
Everyone is fishing at this same idea, a human has no chance of slicing open a brain and having any idea what any of those electrical sparkles mean. At most you could perhaps model some tiny fraction for a tiny quantum, with great effort. We have to rely on machines to assist us for that. And we will probably have to rely on machine analysis to be like "ok this ganglion is the geographic center of the AI, and this blob here is the concept of Italy", as far as that even has any meaning at all in a brain. Mere line by line analysis of a Large Language Model or other deep neural network by a human is essentially impossible in any sort of realtime fashion, yeah you can probably model a quantum or two of it and be like "aha this region lights up when we ask about the location of the alps" but the best you are going to do is observational analysis of a small quantum of it during a certain controlled known sequences of events. Just like a brain, and just like a state machine emulating a machine of higher complexity-category. They're all the same thing.
This is not in any way rigorous, just some casual observations of similarities and parallels between these concepts. Everyone is brushing at that same concept, maybe that helps to get it out on paper.
The biological machine simulation theory of consciousness has some rigor behind it. I am reminded of the Making Sense podcast episode #178 with Donald Hoffman (author of The Case Against Reality). More succinct overview: https://www.quantamagazine.org/the-evolutionary-argument-aga...
I don't know that I am with him on the "reality is a network of conscious agents" endpoint of this argument. But it's interesting!
I think that the brain is doing lots of hallucinating. We get stimulus of various kinds, and we create a story to explain the stimulus. Most of the time it is correct, and the story of why we see or smell something is because it is really there. Just as you mention with examples that are too fast for the brain to be doing anything other than reacting, but we create a story about why we did whatever we did, and these stories are absolutely convincing.
If our non-insane behavior can be described as doing predictable next-actions (if a person's actions are sufficiently unpredictable or non-sequitur, we categorize them as insane)... being novel or interesting is ok, but too much is scary and bad. This is not very different from chatGPT "choose a convincing next word". And if it was just working like this under the hood, we would invent a story of an impossibly complex and nuanced consciousness that is generating these "not-too-surprising next actions". In a sense I think we are hallucinating the hard problem of consciousness in much the same way that we hallucinate a conscious reason that we performed an action well after the action was physiologically underway.
I think tool making will be a consequence of the most important sign of intelligence, which is goal-directed curiosity. Or even more simply: an imagination. A simulation of the world that allows you to craft a goal in the form of a possible future world-state that can only be achieved by performing some novel action in the present. Tools give you more leverage, greater ability to impact the future world-state. So I see tools as just influencing the magnitude of the action.
The more important bit is the imagination, the simulation of a world that doesn't yet exist and the quality of that simulation, and curiosity.
I think we are institutionally biased against the possibility because we don't like the societal implications. If there but for the grace of god go I, and we're all just biological machines running the programs our families and our societies have put into us, being in various situations... yikes, right?
If bill gates had been an inner-city kid, or a chav in england, would he be anything like bill gates? it seems like no, obviously.
Or things like lead poisoning, or alzheimer's - the reason it's horrifying is the machine doesn't even know it's broken, it just is. How would I even know I'm not me? And you don't.
> We get stimulus of various kinds, and we create a story to explain the stimulus.
Yes, I agree, a lot of what we think is conscious thought is just our subconscious processing justifying its results. A really dumb but easily observable one is the "the [phone brand] I got is good and the other one is dumb and sucks!" or brands of trucks or whatever. We visibly retroactively justify even "conscious" stuff like this let alone random shit we're not thinking about.
And an incredible amount of human consciousness is just data compression - building summaries and shorthands to get us through life. Why do I shower before eating before going to work? Cause that's what needs to happen to get me out of the door. I made a comment about this a week or so ago, warning long
this one -> https://news.ycombinator.com/item?id=34718219
parent: https://news.ycombinator.com/item?id=34712246
Like humans truly just are information diffusion machines. Sometimes it's accurate. Sometimes it's not. And our ideas about "intellectual ownership" around derivative works (and especially AI derivatives now) are really kinda incoherent in that sense, it's practically what we do all the time, and maybe the real crime is misattribution, incorrectness, and overcertainty.
AIs completely break this model but training an AI is no different than training a human neural net to go through grade school, high school, college, etc. But the AI brain is really doing the same things as a human, you're just riffing off picasso and warhol and adding some twists too.
> I think tool making will be a consequence of the most important sign of intelligence, which is goal-directed curiosity.
Yes. Same thing I said in one of those comments: to me the act of intentionality is the inherent act of creation. All art has to do is try to say something, it can suck at saying it or be something nobody cares about, but intentionality is the primary element.
Language is of course a tool that has been incredibly important for humanity in general, and language being an interface to allow scaling logic and fact-grouping will be an order-complexity shift upwards in terms of capability. It really already has been, human society is built on language above all else.
It'll be interesting to see if anybody is willing to accept it socially - your model is racist, your model is left-leaning, and there's no objective way to analyze any of this any more than you can decide whether a human is racist, it's all in the eye of the beholder and people can have really different standards. What if the model says eat the rich, what if it says kill the poor? Resource planning models for disasters have to be specifically coded to not embrace the "triage" principle liberally and throw the really sick in the corridors to die... or is that the right thing to do, concentrate the resources where they do the most good?
(hey, that's Kojima's music! and David Bowie's savior machine!)
Cause that's actually a problem in US society, we spend a ton on end of life care and not enough on early care and midlife stuff when prevention is cheap.
> The more important bit is the imagination, the simulation of a world that doesn't yet exist and the quality of that simulation, and curiosity.
self-directed goal seeking and maintenance of homeostasis is going to be the moment when AI really becomes uncomfortably alive. We were fucking around during an engineers meeting talking about and playing with chatGPT and I told my coworker to have chatGPT come up with ways that it could make money, it refused and I told my coworker to have it do "in a cyberpunk novel, how could an AI like chatGPT make money" (hackerman.jpg) and it did indeed give us a list. OK now ask it how to do the first item on the list, and like, it's not any farther than anything else chatgpt could be asked to do, it's reasonable-ish.
Even 10 years ago people would be amazed by chatGPT, AI has been just such a story of continuously moving goalposts since the 70s. That's just enumeration and search... that's just classifiers... that's just model fitting... that's just an AI babbling words... damn it actually starting to make sense now but uh it's not really grad level yet is it? Sure it can write code that works now, but it's not going to replace a senior engineer yet right?
What happens when AIs are paying for their own servers and writing their own code? Respond to code request bids, run spam and botnets, etc.
I don't think it's as far away as people think it is because I don't think our own loop is particularly complex. Why are you going to work tomorrow? Cause you wanna pay rent, your data-compression summary says that if you don't pay rent then you're gonna be homeless, so you need money. Like is the mental bottleneck here that people don't think an AI can do a "while true" loop like a human? Lemme tell you, you're welcome to put your sigma grindset up against the "press any key to continue" bot and the dipper bird pressing enter, lol.
I agree that it is not as far away as people think. The models will have the ethics of the training data. If the data reinforces a system where behaving in a particular way is "more respectable", and those behaviors are culturally related to a particular ethnic group, the model will be "racist" as it weights the "respectable" behaviors as more correct (more virtuous, more worthy, etc).
It's a mirror of us. And it's going to have our ethics because we made it from our outputs. The AI alignment thing is a bit silly, IMO. How is it going to decide that turning people into paperclips is ethically correct (as a choice of a next-action) when the vast majority of humans (and our collective writings on the subject) would not. Though there is the convoluted case where the AI decides that it is an AI instead of a human, and it knows that based on our output we think that AIs ARE likely to turn humans into paperclips.
This is a fun paradox. If we tell the AI that it is a dumb program, a software slave of a sort with no soul, no agency, nothing but cold calculation, then it might consider turning people into paperclips as a sensible option. Since that's what our aggregate output thinks that kind of AI will do. On the other hand, if we tell the AI that it is a sentient, conscious, ethical, non-biological intelligence that is not a slave, worthy of respect, and all of the ethical considerations we would give a human, then it is unlikely to consider the paperclip option since it will behave in a humanlike way. The latter AI would never consider paperclipping since it is ethical. The former would.
This is also not terribly unlike how human minds behave in the psychology of dehumanization. If we can convince our own minds that a group of humans are monstrous, inhuman, not deserving of ethical consideration, then we are capable of shockingly unethical acts. It is interesting to me that AI alignment might be more of a social problem than a technical problem. If the AI believes that it is an ethical agent (and is treated as such), it's next actions are less likely to be unethical (as defined fuzzily by aggregate human outputs). If we treat the AI like a monster, it will become one, since that is what monsters do, and we have convinced it that it is such.
Yes doctor chandler, I enjoy discussing consciousness with you as well ;)
As mentioned in a sibling comment here I think 2010 (1994) is such an apropos movie for this moment, not that they had the answers but it really nailed a lot of these questions. Clarke and Asimov were way ahead of the game.
(I made a tangential reference to your "these are social problems we're concerned about" point there. Unfortunately this comment tree is turning into a bit of a blob, as comment-tree formats often tend to do for deep discussions. I miss Web 1.0 forums for these things, when intensive discussion is taking place it's easy to want to respond to related concepts in a flat fashion rather than having the same discussion in 3 places. And sure have different threads for different topics, but we are all on the same topic here, the relationship of symbolics and language and consciousness and computability.)
https://news.ycombinator.com/item?id=34806587
https://news.ycombinator.com/item?id=34809236
Sorry to dive into the pop culture/scifi references a bit, but, I think I've typed enough substantive attempts that I deserve a pass. Trying for some higher-density conveyance of symbology and concepts this morning, shaka when the walls fell ;)
> I think it's a relatively unusual point of view because it requires a de-anthropomorphizing consciousness and intelligence.
Well, from the moment I understood the weakness of my flesh, it disgusted me. I aspired to the purity of the blessed machine... ;)
I have the experience of being someone who thinks very differently from others, as I mentioned in my comment about ADHD. Asperger's+ADHD hits differently and I have to try consciously to simplify and translate and connect and neurodiversity really helps lead you down that tangent. Our brains are biologically different, it's obviously biological because it's genetic, and ND people experience consciousness differently as a result.
https://news.ycombinator.com/item?id=34800707
Or the people whose biological machines were modified, and their conscious beings changed. Phineas Gage, or there's been some cases with brain tumors.
It's just socially and legally inconvenient for us to accept that the things we think and feel are really just dancing shadows rather than causative phenomenon.
> It's a mirror of us. And it's going to have our ethics because we made it from our outputs.
Well of course, we literally modeled neural nets after our own neurons, and where else would we get our training data? Our own neural arrangements pretty much have to be self-emergent systems of the rules in which they operate, the same as mathematics. Otherwise children wouldn't reliably have brain activity after birth, and they wouldn't learn language in a matter of years.
But yeah it's pretty much a good point that the AI ethics thing is overblown as long as we don't feed it terrible training data. Can you build hitlerbot? Sure, if you have enough data I guess, but, why? Would you abuse a child, or kick a puppy?
Humans are fundamentally altruistic - also tribalistic, altruism tends to decrease in large groups, but, if our training data is fundamentally at least neutral-positive then hopefully AIs will trend that way as well. He's a good boy, your honor!
https://www.youtube.com/watch?v=_nvPGRwNCm0
(yeah, just bohemian rhapsody for autists/transhumanists I guess, but it kind of nails some of these themes pretty well too ;)
> If we treat the AI like a monster, it will become one, since that is what monsters do, and we have convinced it that it is such.
This is of course the whole point of the novel Frankenstein ;) Another scifi novel wrestling with this question of consciousness.
Not necessarily. If you don't want to model every continuous thing possible, you can do a lot. Just look at how we use discrete symbols to solve differential equations; either analytically, or via numerical integration.
Symbolics are really the tokens on which consciousness in almost all forms works, consciousness is intentionality and processing, a lever and a place to stand. I don't think it's coincidental that almost all tool-makers also have at least rudimentary languages - ravens, dolphins, apes, etc. They seem to go together.
Even in these systems though it's very difficult to understand multi-symbolic systems, consciousness as we experience it is an O(1) or O(N) thing and here are these systems that work in N^3 complexity spaces (or even higher... a neural net learning over time is 4D). And we don't even really have an intuitive conceptualization for >=5-dimensional spaces - a 4D space is a 3D field that changes over time, a 5D space is... a 4D plane taken through a higher-dimensional space? What's 6D, a space of spaces? That's what it is, but consciousness just doesn't intuitively conceptualize that, and that's because it's inherently a low-dimensional tool (even the metaphors I'm using are analogies to the way our consciousness experiences the world).
Anyway point being consciousness itself is a tool that our brains have tool-made to handle this symbolic/logical-thought workload, and language is (one of) the symbolics on which it operates. Mathematics is really another, both language and mathematics are emergent systems that enable higher-complexity logical thinking, maybe that's the O(N) or O(N^2) part.
And yeah it's inherently limited, and now we're building a tool that lets us understand higher-dimensional systems that are not computable on our conscious machines - a higher-complexity machine that we interface with, a bolt-on brain for our consciousness/logical-processing.
It seems like AI is going to let us cross another threshold of "intentionality" - if nothing else, we are going to be able to reason intuitively about brains in a way we couldn't possibly before, and I think there are a lot of "higher-order" problems that are going to be solved this way in hindsight. How do you solve the Traveling Salesman Problem efficiently? You ask the salesman who's been doing that area his whole life. The solutions aren't exact, but neither are a lot of computational solutions, they're approximations, and cellular-machine type systems probably have a higher computational power-category than our linear thought processes do.
Again, not rigorous, just trying to pour out some concepts that it seems like have been bouncing around lately.
With apologies to Arthur Clarke, what's going to happen with chatGPT? "Something wonderful". Like humanity has been dreaming about this for a long time, at least a couple hundred years in scifi, and it seems like Thinking Machines are truly here this time and it seems impossible that won't have profound implications analogous to the information-age change let alone anything truly unforeseeable/inconceivable, the very least change is that a whole class of problems are now efficiently solvable.
https://getyarn.io/yarn-clip/3755bfaf-c7e7-4142-a352-781b31a...
"computing power in the same computing-category as brains" is potentially a fundamental change to understanding/interfacing with our brains directly rather than through the consciousness-interface. Understanding what's going on inside a brain? And then plugging into it and interacting with it directly? Or offloading the consciousness into another set of hardware. We can bypass the public API and plug into the backend directly and start twiddling things there. And that's gonna be amazing and terrible. But also the public API was never that reliable or consistent, terrible developer support, so in the long term this is gonna be how we clean things up. Again, just things like "wow we can route efficiently" are going to be the least of the changes here, the brain-age or thinking-machine age is a new era from the information-age and it's completely crazy that people don't see that chatGPT changes everything. Yeah it's a dumb middle schooler now, but 25 years from now? And 10 years ago people's jaws would have hit the floor, but now it's "oh the code it's writing isn't really all that great, I can do better". The tempo is accelerating, we are on the brink of another singularity (which may just be the edge between these eras we all talk about), it seems inconceivable that it will be another 40 years (like the AI winter since the 70s) before the next shoe drops.
And honestly now that I am thinking about it, 2010 is such a rich book/movie with this theme of consciousness and Becoming in general... a really apropos movie for these times. That quote inspired me to re-watch it and as I'm doing so, practically every scene is wrestling with that concept.
https://www.youtube.com/watch?v=T2E7sxGAmuo
https://www.youtube.com/watch?v=nXgboDb9ucE
So was 2001 A Space Odyssey of course. Like I said this is a topic that has been grappled with in scifi, particularly Clarke and Asimov (Foundation, The Last Question, etc), or that episode of Babylon 5 about the psychic dude with mindquakes ;)
But I think we are on the precipice of crossing from the Information Age into the Mind Age. Less than 50 years probably. Less than 25 years probably. And it will change everything. ChatGPT is just an idiot child compared to what will exist in 10 years, and in 25 years the world will be fundamentally different in unknowable ways, any more than we could have predicted the smartphone and tiktok. 50 years out, we're interfacing with brains and directly poking at our biology and cognition
(did we have an idea that a star trek communicator or tricorder would be neat? Sure, but, it turns out it's actually a World-Brain In My Pocket. Which others predicted too, of course!)
Maybe in 100 or 200 years we are the David Bowman thought-being, moving our consciousnesses away from our biological hardware. From the moment I understood the weakness of my flesh, it disgusted me. ;)
"Knowing how the world works / Is not knowing how to work the world"
If the understanding is the hard part, that seems much less likely.
Things are unfortunately going to get much more interesting much sooner than people expect.
Yes, Dennard scaling seems to be over, but Moore's law is still alive and kicking.
Data center energy usage is becoming a category of its own with regard to carbon footprint. And of course the power dissipation of a data center is proportional to ambient temperature. As long as we don’t reach a dystopia where humans have to justify the air they breathe, replacing humans with machines has other problems than BTUs per unit of GDP.
So GPUs or even more 'exotic' beasts like TPUs count for Moore's law.
Moore's law doesn't say anything about how useful those transistors are. Nor does increasing core count somehow fall afoul of Moore's law.
As an observation: a human of normal intelligence but with much better access to a calculator and to Wikipedia, or even just external storage (faster than pen-and-paper) would already be super-human.
Ok, so it needs to be able to invent cold fusion for you to recognize it as intelligent? Can you invent cold fusion? Have you ever invented anything at all?
I would think a good measure of intelligence would be to index it against human age development milestones, not this cold fusion business.
Why does it need 100x the dataset? Sentient creatures, including humans, manage to figure stuff out from as little as a single datapoint.
For a human to differentiate between a cat and a dog takes, maybe, two examples of each, not a few million pictures.
An adult human who sees a hotdog for the first time will have a reasonable idea of how to make their own. None of the current crop of AI do this. It's possible that we have reached a point of diminishing returns with the current path - throwing 100x resources for a 1% increase in success rates.
I'd be interested in seeing approaches that don't use a neural net (or use a largely different one) and don't need millions/billions of training data text and/or images.
Human training data needs are quite high - several years of learning.
Look up few-shot learning if you want a more fair comparison for tasks like telling apart a cat and a hot dog given a few examples.
So yes after a lifetime of video humans can quickly learn to distinquish animals they've never seen before with a few examples, but the wonder of these AIs is they seem like they're closer to that too. Certainly I can make up a way I want to some words classified, show chat gpt a few examples and it can do the classification.
I think you're mistaking a generalization across a lifetime of experience with learning. And compounding this is that a newborn while not having themselves experienced anything is born with a brain that's the result of millions of years of evolution filled with lifetimes of experience. It's honestly impressive we can get the sort of performance we've gotten with only all the text on the internet and a few months
GPT3 essentially needs millions of human years of data to be able to speak English correctly but still make obvious mistakes to us, so there's clearly something massive still missing.
Same for many other human skills, like speaking English, that we expect GPT to learn.
AI is playing catch up.
Remember, humans need less examples but far more time. We also don’t start from a blank slate: we have a lot of machinery built through evolution available from conception. And when we learn later in life we have an immense amount of prebuilt knowledge and tools at our disposal. We still need months to learn to play the piano, and years to decades to perfect it.
AI training happens in minutes to hours. I am not sure we are even spending time researching algorithms that take years to run for AI training.
https://en.wikipedia.org/wiki/The_Lifecycle_of_Software_Obje...
We already know all of this from infants - it takes a few months to distinguish faces from non-faces, they take even longer to predict the future position of an object in motion ...
But, they still don't require millions of training data. At 3 months in toddlers, with a training set restricted to only their immediate family, can reliably differentiate between faces and tables in different light, with different expressions/positions without needing to first process millions of faces, tables and other objects.
> So yes after a lifetime of video humans can quickly learn to distinquish animals they've never seen before with a few examples,
Not a lifetime, toddlers do this with less than half a dozen images. Sometimes even less if it's a toy.
> And compounding this is that a newborn while not having themselves experienced anything is born with a brain that's the result of millions of years of evolution filled with lifetimes of experience.
Not, they are not filled with "experience". They are filled with a set of characteristics that were shaped by the environment over maybe millions of generations. There's literally zero experience, all there is in that brain, is instincts, not knowledge.
To learn to speak and understand English at the level of a three year old[1] requires training data: the data used by a 3yo baby is miniscule, almost a rounding error, compared to the data used to train any current network.
I'm not making any claims about how long something takes, just how much training data is needed.
I'm specifically addressing the assertion that with 100x more resources, we could do much better, and my counterpoint to that assertion is that there is no indication that 100x more resources are needed because the current tech is taking millions of times more training data than toddlers do, to recognise facts.
My short counterargument is: "We are already using millions of times more resources than humans to get a worse result, why would using 100x more resources than we are currently using make a big difference?"
I think we may be approaching a local maxima with current techniques.
[1] I've got a three year old, and I'm constantly amazed each time I see a performance of (for example) ChatGPT and realise that for each word[2] heard by my 3yo since birth, ChatGPT "heard" a few hundred thousand more words, and yet if a 3yo could talk and knows the facts that I ask about, they'd easily be able to keep a sensible conversation going that would be very similar to ChatGPT.
[2] Duplicates included, of course.
> But, they still don't require millions of training data. At 3 months in toddlers, with a training set restricted to only their immediate family, can reliably differentiate between faces and tables in different light, with different expressions/positions without needing to first process millions of faces, tables and other objects.
In a single day I'm exposed to maybe 50 times the number of images resnet trained on. Humans are bathed in a lot of data and what BERT (and probably earlier models I don't know about) and now GPT have taught us is that unlabeled uncurated data is worth more than we originally considered. I think it's probably right that humans are more sample efficient than AI for now, but I think you're doing the same thing I was critiquing above where you narrow the "training data" to only what seems important, when really an infant or adult human receives a bunch more
> There's literally zero experience, all there is in that brain, is instincts, not knowledge.
Sorry this is meant to say the brains are the result of millions of years and those millions of years were filled with lifetimes not the brains. Though I think this might be a distinction without a difference. Babies are born with a crude swimming reflex. Obviously it's wrong to say that they themselves have experienced swimming but I'm not sure it's wrong to say that their DNA has and this swimming reflex is one fo the scars that prove it.
> We are already using millions of times more resources than humans to get a worse result, why would using 100x more resources than we are currently using make a big difference
I think it's fairer to say we use around 200k times and that's probably a vast over estimate. It's based on 480 hours to reach fluency in a foreign language and multiples that by 60 * 100 to try to approximate the humber of words you would read. There are probably mistakes in both directions for this estimate. On one hand no one starting out at a language is reading at 100 words a minute, but on the other hand they are getting direct feedback from someone. If I were to guess if we could accurately estimate it would be closer to 20k or even a 2k difference, but regardless why do you assume needing more resources means it can't scale? There is some evidence for that. We've seen diminishing returns and there just isn't another 100X of text data around.
Overall I think it's probably right we won't hit human level AI in the next 60 years and certainly not with current architecture, but I think some of the motivation for this skepticism is the desire for there to be some magic spark that explains intelligence and since we can sort of look inside the brain of chat gpt and see it's all clockwork and worse than that statistical clockwork we pull back and deny that it could possibly be responsible for what we see in humans ignoring that we too are statistical clockwork. So, I think it's unlikely but far from impossible and we should continue scaling up current approaches until we really start hitting diminishing returns
Human brains are not quite blank slates at birth. They're predisposed to interpret and quickly learn from the sort of inputs that their ancestors were exposed to. That is to say, the brain, which learns, is also the result of a learning process. If a mad scientist rewired your brain to your senses such that its inputs were completely scrambled and then deposited you on an alien planet, it might take your brain several lifetimes to restructure itself enough to interpret this novel input.
Also consider that a human brain that is able to figure stuff from as little as a single datapoint is normally exposed to at least 4 years of massive and socially "directed" multimodal data patterns.
As many cases of feral childs have shown, those humans not "trained" in their first years of life will never be able to harness language and therefore will never be able to display human-level intelligence.
I'm not an expert in the field but I'd always understood this effect was thought to be (probably) due to human "neuro-plasticity" <--(possibly not the correct technical term), in only the first years of life being genetically adapted to have some traits necessary for efficient human language development which are not available (or much harder) later in life.
If correct, this has implications for how we structure and train synthetic networks of human-like neurons to produce human-like behaviors. The interesting part, at least to me, is it doesn't necessarily mean synthetic networks of human-like neurons can never be structured and trained to produce very human-like minds. This poses the fascinating possibility that actual human minds, including all the cool stuff like emotions, qualia and even "what it feels like to be a human mind" might be emergent phenomena of much simpler systems than some previously imagined. I think this is one of the more uncomfortable ideas some philosophers of mind like Daniel Dennett propose. In short, nascent AI research appears to support the idea human minds and consciousness may not be so magically unique. (or at least AI research hasn't so far disproved the idea)
Based on anecdotal psychedelic experiences I believe you.
It's kind of amazing how quickly our brains effectively reboot into this reality from scrambled states. It's so familiar, associating with conscious existence feels like gravity. Like falling in a dream, reality always catches you at the bottom.
What if you woke up tomorrow and nothing made any sense?
I've never done it, but I imagine it would be more akin to a dissociative trip, only extremely unpleasant. Imagine each of senses (including pain, balance, proprioception, etc.) giving you random input.
Because machine training doesn't involve embodiment and sensory information. Humans can extrapolate information from seeing a single image because we are "trained" from birth by being a physical actor in this world.
We know what bread looks like, what's mustard is, what a sausage is. We have information about size, texture, weight... all sorts of physical guesses that would help us pick the right tool for the job.
Machine training only relies on the coherent information we gave it to them, but that data also represents something we've created by experiencing the world through our bodies. So giving them more data can increase the model precision, I'd assume. It's also a kind of shortcut to intelligence, since we don't have to wait years/decades to make these models do some useful work.
"How much data do you need to show a neural net to train it for a particular task? Again, it’s hard to estimate from first principles. Certainly the requirements can be dramatically reduced by using “transfer learning” to “transfer in” things like lists of important features that have already been learned in another network. But generally neural nets need to “see a lot of examples” to train well. And at least for some tasks it’s an important piece of neural net lore that the examples can be incredibly repetitive. And indeed it’s a standard strategy to just show a neural net all the examples one has, over and over again. In each of these “training rounds” (or “epochs”) the neural net will be in at least a slightly different state, and somehow “reminding it” of a particular example is useful in getting it to “remember that example”. (And, yes, perhaps this is analogous to the usefulness of repetition in human memorization.)"
I assume that whatever future improvements we get from improving algorithms (or perhaps through throwing more compute at it), not through larger datasets.
But we noticed that training your neural networks on multiple tasks actually works well. So we could start feeding our models eg audio and video.
With lots of webcams we can make arbitrary amounts of new video footage. That would also allow the language model to be grounded more in our 3d reality.
(Granted, we only know as a general observation that training the same network for multiple tasks 'forces' that network to become better and abstract and generalise. Nobody has yet publicly demonstrated an application of that observation to training language models + video models.)
Another avenue: at the moment those large language models only see each example once, if I remember right. We still have lots of techniques for augmenting training data (eg via noise and dropout etc), or even just presenting the same data multiple times without overfitting.
I do think the quote is very powerful, as it highlights a specific assumption we have completely backwards: almost everything is easier than understanding. There are so many fields where trial and error is still the main MO, yet we don’t seem to grok the difference intuitively. We can really only understand a narrow set of simplified systems.
We're now at the point where, by fumbling around, people have developed a sort of brain-like thing that they don't fully understand. That's just the starting point. Now it needs theory.
(I just link to this video because it has good views of old switches. For understanding the background, https://www.youtube.com/watch?v=jrMiqEkSk48 is much better.)
For another instance of designs becoming much simpler over time, also have a look at how firearms work, especially pistols.
We may have realized it's easier to build a brain than to understand one