ffwd

I agree and I think if there are any hazards to modern information landscape it's kind of 2 fold (that I can think of) 1) the brain likes simple generalized information/models more than specific information and 2) the brain likes information that confirms our existing beliefs more than staying in a gray area where it's not sure what information is true or not.

Regarding first one I think one of the main challenges will be finding which general explanations are true or not. This can be anything from cultural things like gender discussions to conspiracy theories to scientific theories. If there are generalized models that explain a lot of phenomena, the brain likes those, but at the same time they can also take the place for when empirical evidence should have been gathered (that doesn't fit with the general model), and so in that sense they can be hazardous. And models that already have stood up to a lot of empirical data are especially hazardous because they have already proven themselves to some extent which means there's even less incentive to gather empirical data. Or at least that's my current (maybe hazardous) general view.

I also don't think we should give up on generalized models because they are so powerful and useful, but it's definitely a challenge in todays climate because they get so many clicks and they are very satisfying to know.

There are two types of gradient though, conceptually. If consciousness is some state of matter that is unknown still, and each neuron, for example, contains "one bit" of consciousness, then the gradient is that as you add more neurons, you add more complexity to the consciousness, but you do not change the fundamental experience of consciousness. You add more content but not more experience in itself.

If on the other hand consciousness is this emergent phenomenon that depends on neurons and their connections, then the gradient (and thus the experience) would be far more diverse and there would be a lot of different ways consciousness could "feel".

The problem I have is that for example, as far as my brain can remember, stimuli has looked the exact same all throughout my life. If I saw my a tree when I was 10, and I saw the same tree now, the conscious "qualia" of this would look exactly the same. To me this is a mystery, that the connections in the brain do not change the experience of qualia at all. Red looks like red no matter what the neuronal state of your brain is. I don't have an answer to this but just something I've been thinking about.

I'm not one to whine about downvotes but I just have to say, it's a bad feeling when I can't even respond to the negative feedback because there is no accompanying comment. Did I misinterpret something? Did you? Who will ever know when there is no information. :L

Just speculating but I think attention enables differentiation of semantic concepts for a word or sentence within a particular context. Like for any total set of training data you have a lesser number of semantic concepts (like let's say you have 10000 words, then it might contain 2000 semantic concepts, and those concepts are defined by the sentence structure and surrounding words, which is why they have a particular meaning), and then attention allows to differentiate those different contexts at different levels (words/etc). Also the fact you can do this attention at runtime/inference means you can generate the context from the prompt, which enables the flexibility of variable prompt/variable output but you lose the precision of giving an exact prompt and getting an exact answer

Thanks, I can't write a full reply now (need to think) but for some reason my intuition was, let's say you have a constant punishment signal, and a timer, and if it doesn't solve whatever problem by the time the timer goes down, then it has to find the optimal action set, and if it reaches a goal earlier than the timer then it has to weigh that solution stronger? Like at least how I see with organisms it's about finding the optimal use of the limbs in order to solve an ongoing problem/goal state, and if you accumulate specific actions (instead of one network that optimizes one "space"), and different types of goal states, then it has to find the optimal set of actions to reach the different goals, it just seemed more efficient. But this is off the cuff a bit right now.

Edit: i think what i'm thinking of are two "global" numbers. A "closer to the goal" number (higher when closer), and a countdown timer, and then it has to maximize those within the above setting

I don't know much about RL but I was wondering has anyone tried the opposite? Like have a fixed set of actions, and a fixed/ranged movement speed, then a punishment every time it doesn't reach the goal? Or does this not work?

I think there is another aspect to human thinking other than system 1/system 2 though, which is the abstract world model humans have. system 1 / 2 is more like the process, while the world model is the actual data being 'processed'.

And I think basically, humans have a much simplified, 'low dimensional' world model that consists of a set of objects (let's call them patterns), and then a "list" of essential properties that those objects have, that leads to a constraint on how each object can behave in the world model as a whole.

And this is sort of hierarchical or at least, we can zoom in and out in detail depending on the level of knowledge we have about a particular pattern.

So problem 1 is: It's not clear to me that text or any sort of data would contain all the necessary constraints so that any particular prompt would result in a world model that exactly takes into account the constrains of each object and 2) Even if was, I'm not sure the process of step by step thinking (system1/2) about each object and computing world states could occur in current architectures. This is especially important for computing a set of objects, then abstracting the result, then doing another round of computing with that result, or something like this.

I'm not hard set on this but this is my current thinking.

Regarding the dejargonizer - Just be careful of hallucinations! I did a similar gpt prompt where i asked for a simple basics for some complex topic, and sometimes there would be incredibly subtle hallucinations like even on a word basis, and so I had to stop using it. I'm not sure how well yours works or if it's much better now, but just something to be aware of if you're not familiar with the topic you query about

I have a question which I don't know the answer to:

With those structured numbers will the LLMs be 100% accurate on new prompts or will they just be better than chance (even significantly better than chance)?

Because this is one thing, it has to learn the structure and then create probabilities based on the data, but does that mean it's actually learning the underlying algorithm for addition for example or is it just getting better probabilities because of a narrowing of them? If it can indeed learn underlying algorithms like this that's super interesting. The reason also this is in an issue if it _can't_ learn those, you can never trust the answer unless you check it, but that's sort of a sidepoint.

Yet another post about LLMs (but give it a chance!) English is not my first language, I know it is a little rough in places, but I did my best.

I'm just curious, how polysemantic is the human brain with each neuron? Cause it feels to me, what you really want, and what the human brain might have, is a high-information (feature based / conceptual based / macro pattern based) monosemantic neural network, and where there is polysemantic neurons, they share similar or the same information in the feature it is a part of (leading to space efficiency? as well as computational efficiency). Whereas in transofmrer models like this, it's as if you're superimposing a million human brains on top of the same network, and then averaging out somehow all the features in the training set into unique neurons (leading naturally to a much larger "brain"). And also they mention in the paper that monosemantic neurons in the network don't work well, but my intuition would be because they are way too "high precision" and they aren't encoding enough information at the feature-level. Features are imo low dimensional, and then a monosemantic high dimensional neuron would the encode way too little information or something. But this is based on my lack of knowledge of the human brain so maybe there are way more similarities than I'm aware of...

Well I can't say any of this for sure but I want to say upfront, I think llm's can in theory do a lot (not sure if most) of the computations a human can do, but it's important to realize, imo, it's not actually stepping through the steps in the way humans are. When we give complicated step by step prompts and so on, it only means it's creating new constraints for what the probability of the next token is (from what exists in the data/model). If the data/model doesn't contain the data needed to produce the desired result, or the data that it was trained does not have examples that can generalize (but not be specific to) the desired result then it can't produce it.

That's the difference between humans and llm's imo. We can generalize any "computation" we have to any other "desired output" we want, by thinking about it, while llm's aren't at least not now, so general that they use low level representations of all the 'objects' we can prompt about. Like humans can reason about the objects and things in our mind almost infinitely and recursively while also retaining all the physical realities and facts of those objects, while an llm is limited in this regard. Doesn't mean it can't in theory, there is some weird generalization going on as far as I can tell, but it feels like it's going to need a lot more data or something to do it.

To me predicting the next token is obviously not how humans think.

If I ask you to envision a green triangle and a red square next to each other, and then swap the shapes but keep the colors in the same locations, and answer what color is the triangle now, you say the triangle is red, but you do so because you envisioned the triangle swapping places and did the mental steps etc.

An LLM if even answering correctly, is statistically answering based on billions of lines of text + rlhf and all of this, I highly doubt there is a mental model of the world, but rather a large set of constraints in the probabilities which leads to the resulting answer. The reasoning ability is a secondary effect of the probabilities which is why it's hard to make it so every probability is correct for every answer I think.

And regarding OP about hallucinating vs confabulating. To me hallucinating is a fine word for it, because it is filling in a gap or there aren't enough constraints in the model/data/tuning to account for that specific answer that it gave that was incorrect. Hence it "hallucinates" something in the gap. The real power of LLM's is that it seems to accumulate these 'constraints' (generalization), so that with the right model, it should be able to answer more and more prompts correctly, which is kind of amazing.

Confabulation works too but is a little more high level IMO.

I understand the argument but I disagree and would phrase it the other way around. We don't have evidence of conscious experience because we don't have evidence for a division between conscious experience and "existence/physical stuff" in the first place. This division is created by our brains and the model we have of the world but in reality the only evidence we have is that something exists.

Therefore, what we call subjective experience is evidence for the existence of what we experience, but until we can empirically show that there is a difference between what we experience and what reality is, the only thing we can say is that it exists. Also, that's the problem with the hard problem of consciousness - it assumes a division between the two, and we have no evidence of such a division. And lastly if it's not clear i am not a panpsychist either because that assumes a division on some level again. We don't know what existence itself is and how it functions really, at least not the brains function within existence.

In 2022 it was definitely hunger as I stopped eating high calorie stuff completely. But that level of eating was unsustainable and only possible because of the health issue I was forced to. I was basically thinking about food 24/7.

However after I started eating less healthy food it definitely can be. That's the issue I have atm, giving up food I enjoy vs feeling high calorie withdrawal (at least relatively, I eat a lot less now than before).

I actually sleep better when not eating before sleep. Sort of a bigger point about this - last year when I was losing a lot of weight I was eating around 1500 calories per day, and I felt pretty amazing ironically. Felt a lot more energetic and lighter, but I basically ate no fast food or any food I enjoy, and it was only this year when I started to be able to eat my favorite food I started to realize I feel a lot worse (basically a little more bloated, heavier, not as energetic) when eating it (things like pizza, burgers, etc). I had to stop with chips and chocolate completely though.

But my point is it feels like that's sort of the issue I'm dealing with right now is not eating much at all reduced my appetite to almost nothing, but I basically had to stop eating any food I enjoy (like pizza, burgers, etc), so I had to choose between food not being a part of my life much (and get stable blood sugar), or eat unhealthy in periods and then feel worse. I haven't come up with a perfect solution but I felt like I couldn't give up the food so...

Edit: Also by the way, regarding sleeping hungry directly, my motivation was that I knew I was losing weight or "accomplishing the goal" when doing it, so mentally I felt good about feeling hungry. I couldn't wait till the next day came around and I had slept and knocked off those hours up until 9 am.

I didn't know that's what it was called but yes, it appears it is.

I made a simple rule, but i'm not sure it will work for everyone, but I lost 77 lbs in 2022 due to a health issue and it seems to work.

Basically I have to be hungry from 6 pm to 9 am (and not eat), for at least a week, preferably 2 weeks. If I feel hungry for a period, i lose weight or at worst stay at current weight. Usually i lose a few pounds. Then I can have periods where I can eat more and gain a few, then I just go back to not eating as much and it goes down again. The key thing is to just feel that hunger feeling relatively often and it's usually a good sign.

I get used to the hunger feeling after awhile, so it's not that big of an issue after that.

I'm going to be contrarian-ish and say that LLM's / ai will never gain the creativity, flexibility and ability to adopt to novel stimuli nor create the new novel thing as much as a skilled human. That's why artists will also survive.

There will be a few years of tumult and experimentation with AI's but humans (the best ones in particular areas) will always be one step ahead imo.

Only jobs that need very little or no adaptation or creativity can be automated forever I think.

That is until, or if, there is an AI with an actual "life life" brain with new capabilities currently not existing

I have one slight addendum that didn't fit in the main post character limit (sorry I know this is long) :P:

And then my final point is that (and I'm not sure about this exactly), but any action that current AI's do, are essentially only the computer code that some programmer wrote to execute at that time. For example if we have an AI it can classify images, someone wrote a classify() function, and right now, the AI cannot do anything more or less than run exactly that function. All the knowledge and actual activations in the deep learning net are essentially "dead execution", it doesn't matter nor change the classify() function at all.

But that's not how humans work, humans create novel functions other than classify() and we also have a biological imperative to do so, as well as a capability. Right now none of the information in the AI neural net can actually be utilized to create alternatives to classify() nor does it have an imperative nor technical ability do so, so the neural net remains "dead".

And to do so you would need to make different parts of the neural net actually count in the creation of the new output function and also to actual physical actuators in the world, but even then we are back to how would it actually synthesize disparate pieces of information in a coherent way as said above.

This is my experience with Win 10 as well. In fact I dare say, Win 10 in the past 5 years is one of the best desktop OS's of all time. No crashes, perfect customizability to turn off things, fast and responsive.

However that said, I'm extremely worried about Win 11, and when I saw all the things they changed in the taskbar (no text labels, no option to never combine icons), the right click menu and so on, I was very disappointed. I did get the send to menu back with a thing and I know you can use winaerotweaker to get all the old stuff back, but a part of me wants to be on the native OS stuff whenever possible and not change core OS with third party apps. But I'm also worried about what they might further do in the future and in 5-10 years, incompatibility in programs/etc with Win 10 might make using it very bothersome so you have to update at some point.

The reason I don't switch is because a 0.5L coke has around 210 calories. You're only supposed to consume around ~2000 calories per day, so even with 4 cokes in a day that's almost half of calorie intake. So not an option for me.

It was poorly communicated in my post but what I was referring to there were earlier programs like in 2015 and not SD and newer ones. If you put in an image of a landscape it could fill out the landscape with eyes and elbows all over the generated image because it had no information or context for what an eye was or where it should go.

But now you get SD, dalle and others which add more information not just by scaling, but also by mapping sentences/words to pre-existing images that already have cohesion. That way when you write in sentences to the text prompt, the model has more semantic information about what an eye is, but (IMO) only _indirectly_ because it will map a sentence to images that match that sentence. The question is always what information is actually contained in the training set and what is missing from it and when it creates an image where is the information from etc.

In some ways, that means I think that meaning to us as humans, is different from scaling which is almost like pixel resolution except resolution of patterns and differentiation of patterns. Meaning in this sense is things like creating a doorway with no actual door, but still the doorway itself looks super realistic is rendered. You can fix it by scaling and increasing the differentiation of patterns I guess, but you can never fix all instances completely with scaling. That's why in some ways I think meaning is sort of orthogonal to scale, however on a philosophical level, they should converge but that's for another topic.

I may have missed something in my thoughts here because this is sort of difficult to talk about without writing a book eventually.

Actually I should have mentioned this in the original post but I think the "3 arms" thing is kind of a bad example come to think of it. I think in general at least with SD, if's very unlikely to create 3 arms or or 8 arms if you for example ask for a person. Mostly it looks like a person because the text prompt maps to training data of persons, and so they will generally look like people with 2 arms.

However, where it struggles I find is with finer details, and also _placement_ of things like arms, eyes, and relationships between them. This I think is because it only has a general idea of the shape of persons but no data for the exact specifics like where the arms, legs, eyes and so on should be placed in a very realistic anatomical way, and this is where I think the challenge is - the gap between a general pattern of a person and an extremely specific but also general one where it can modify it and transform it like a real human artist can. I'm not sure that's in the data exactly

I get your point, but I also think it depends on what you mean by oversimplification. Of course there is _a lot_ of stuff going on and things like SD capture all kinds of information, not just what I described, however, any way you want to describe it, capturing all the "constraints" and real life knowledge to perfectly create realistic images with all the details and all the higher abstractions correctly is not anywhere close I think. Also it's not only to always get 2 arms, it's to - at the same time - also get 2 ears, 2 eyes, perfect pupils, perfect fingers, perfect trees, perfect chairs, all simultaneously (if it is to be used at least in the mainstream) - etc you get my point.

I also don't think there's anything wrong with the model architectures in themselves or the data, nor that it is impossible, only that it is hard and as you say I think it needs a lot of data and clever engineering to fix mistakes. It may even be possible to fix most mistakes, over time, which would be pretty impressive imo, but the absolute limits of what a model can produce/"contain" with our hardware is kind of an open question though interesting.

Not sure if I'm missing a subtle nuance in your point but to me those "artifacts" are completely expected. Those artifacts like 3 arms are the patterns / outputs in the model, but since it doesn't have a fundamental understanding of the patterns/objects like arms, it just blends many images of arms together and create things like 3 arms. Also why there are so many eyes, arms, legs and other things in other generative programs. It just spits out the training set in random configurations (ish).

I suspect also the reason the images look OK at a glance is because the images as a whole also represent patterns in the model so they actually come from "real life" / artist created images and thus have some sense of cohesion. But making the AI have all the right patterns so it never makes a mistake at all scales of the image while also being able to combine the pattern with real understanding of what they are conceptually is the real trick but until then it will be a "salad bowl collage" thing at random intervals.

The closest thing to the brain it looks like to me is simply the hierarchical nature of it which seems similar to v1/v2/the vision system in humans but I've only been told that, I'm no neuroscientist.

I slapped together a really quick script to get oldest videos via Youtube API. It goes through all the videos from a playlist ID and then writes them to a html file with oldest first. Looks like this: https://i.imgur.com/ZxCgnBY.png Not perfect but works ok

https://github.com/n0x5/scripts/blob/master/youtube_api_uplo...

Need an API key and also the Uploads playlist ID from the channel which can use this script to get: https://github.com/n0x5/scripts/blob/master/youtube_get_uplo...

It's a little convoluted I might implement get uploads id right into the script but the issue is it doesn't always return the correct channel when searching by name, in which case one has to manually go through the results and find the right channel id which would take longer to code to actually work all the time.

I have a hypothesis about this type of feeling but it is work in progress and probably not finished.

I think one primary reason why we feel this loneliness is because we fear the situations where we might need other people, in particular situations where we do not have a computer or we are sick or otherwise unable to live the way we usually do on the computer (speaking for HN crowd specifically) ;P

What i mean by this is for example I can imagine myself getting sick or getting old to the point where I can't do the things I am passionate about, so I end up sitting alone somewhere just staring out the window or something, with no friends or family, and that is a pretty lonely thought.

The problem in the modern world is that we can for the most part actually live alone and do the things that we are interested in, and these situations where we actually need other people in person become an abstract idea rather than a concrete need, but historically and probably evolutionary we did need them. I think in prior decades it was a lot harder to entertain yourself and be alone as much as we are today. We can get a hold of food, social interactions online and accomplish goals in a way that was never possible without in person interaction in the past and this confuses our brain which is not used to this arrangement. Edit: Just to make it clear, I think we oscillate between feeling complete happiness and contentment from the things we do alone and enjoy (like coding/reading/whatever you do), and the feeling of being afraid of the moments we might need people. It is this oscillation pattern that confuses our brain because we don't know if what we have is enough or not. We want to "cover" all possible scenarios we can imagine.

I also want to mention the feeling we've all probably had when we do something on our own that was a challenge and we succeeded - it gives a strong feeling of relief when we actually accomplish this. I think, if we can get that feeling for ALL life situations we would no longer feel loneliness since we would be self sufficient in all situations. But up until we are convinced of this, we will always be afraid of the abstract imaginary situation where we will need other people but will be alone I think.

So i think there are two possible solutions -

1) Either convince yourself through experience or thought that there is no situation you cannot deal with alone and it will be fine

or

2) Do the hard work of making and keeping friends which means sacrificing your own time and thoughts etc to them. This is like studying for a test in a boring subject there is no way around it. Don't get me wrong I don't think other peopel are boring I just mean there is work needed there that usually goes beyond the comfort / convenience level of what you're used to.

Tbh I am far from good at this, I am mostly alone nowadays but I do enjoy talking to some select few people, but I try to keep the fear mentioned above in check and also be content that I can handle the situation I am in and future ones.

I suppose so... Not sure about the second part but for the first part I think that comes from the idea that I don't think we've ever empirically or even subjectively observed the universe in any other state than "now". We've never personally nor in a scientific experiment observed any physical system being in more than one state, and that one state has to by definiton be in the "now" (Unless I am mistaken). So the past and the future states are essentially nowhere to be found in empiricism ?

As for the "here" part I am not sure how to respond to that but it sounds right - any observer has to have contact in some direct way (whether with a scientific instrument or not) to the thing physical system he is observing and as such claiming anything about anything indirect or far away is very hard if not impossible so you get the "here".

I feel like I must be missing something but in the part about the relativity of simultaneity [0] I feel like this is more about the relativity of the observers and the limitations of speed of light than it is about the ontology of time itself?

The definition of "now" as when the light hits the mirror on each side is a "constructed" now, a false now if you may. It takes the light longer on the right side than the left side because she is moving through space, all that tells me is that there is no way to communicate a "now" between different observers because there can't be instantaneous communication between observers, not that time itself is relative or physical stuff itself doesn't have a constant time?

She says in the video that we have to find a way to operationalize how to find a now, to measure it empirically but maybe that's not possible without running a complete simulation after the fact. But to me it seems not surprising that without instantaneous communication there is no way to create a "now" empirically but that doesn't mean it extends to the ontology of physical stuff itself or time itself. There would still be a time in there (you should be able to see it in a simulation?)

Also for example, if you use sound instead of light, if I'm far away from a sound source (farther than the speed of sound), and another observer is closer to it, we would not be able to use the sound source as a source "now", because I would hear it later. Is this the same type of situation or different? I feel like I must be missing something fundamental here but I'm sort of going out on a limb so please don't slaughter me :P

[0] https://youtu.be/GwzN5YwMzv0?t=542