My takeaway is that Plato should be on the credits list for The Matrix.

This is fascinating. AI is basically modelling reality through a very inefficient translation layer which is how the human mind is modelling reality. To get true AI we should let models learn reality directly, just as our brains learned it through evolution, and then more recently through science.

The solution is AI battle royal.

I starting discussing the Platonic Representation Hypothesis with ChatGPT and gosh we went on a truly epic philosophical adventure that I barely understood. 😅

As a philosophy and AI nerd, I love when they intersect!

Hey bycloud! Big fan of your ai newspaper man, thanks!

Sounds interesting, BUT! I have a question: to what extent could this be just a result of more parameters = more dimensions along which representations can be spread out = more even spread of representations = distances become more equidistant = distances only *appear* to converge? Because to my (admittedly naive) brain this sounds very similar to the curse of dimensionality, only within foundation model representations. With this in mind, wouldn't it actually be more compelling if you'd find this alignment more in smaller foundations models, rather than larger ones? Especially seeing that the encoding of the (multi-task) representations would have to be more *efficient* rather than sparse?

Very good points, glad it's out here for more people to think about

I know it's just the sponsor segment.. But getting better at "job hunting" isn't the answer, it's just delaying the inevitable. We really need to build a movement around making significant changes to how our society and economy work, if we're going to adapt to the implications of AI.

Keep in mind, the training data for these models doesn't differ very much - except for OpenAI, which has had a massive head start on gathering quality user data. There are some severe limitations to their argument, which they point out at the end of their paper: 1. Different modalities may contain different information: "One immediate objection to our hypothesis is: what about the information that is unique to a given modality? Can language really describe the ineffable experience of watching a total solar eclipse?" 2. Not all representations are presently converging - ". For example, in robotics there is not yet a standardized approach to representing world states in the same way as there is for representing images and text." 3. Sociological bias in producing AI models - Ever notice how LLMs seem to think they're people unless asked directly? "There is often an explicit or implicit goal of designing AI systems that mimic human reasoning and performance, and this could lead to convergence toward human-like representations even if other kinds of intelligence are in fact possible." 4. Special-purpose intelligences might not converge - this is where it *really* falls apart in a way that should be obvious. General purpose intelligence must be able to reason using a theory of mind. A general theory of human mind won't do, as people from different cultures can see the world in *very* different ways. Coming from a Comp Sci x Philosophy research background, I could definitely see models converging based on the current trajectory - but not towards some platonic reality. Rather, they're being engineered for profitable utility. Which is to say, they *won't* be trained to think too far outside the box, because that could be a threat to business alignment and the only groups that can afford to train the cutting edge models are intent on using them for profit. Even beyond that, there's the state/government alignment which implies that if, for example, the US were to become a totalitarian state then the AI would also have to be totalitarian in order to remain politically correct enough for the market.

Great educational video. I love how you find AND explain recent papers that aren't already covered by all the hype media and STILL super interesting.

Love this video! Haven't heard of that paper before! Thank you sir

First 60 seconds made me freeze and think

It makes perfect sense for different AIs to learn similar representations of the same reality. This is similar to how science works.

Hi bycloud can you tell me where and how i can read these types of papers about AI and how to keep up with the latest ones as soon as they come?

It's a nice thought in theory but in practice we don't have infinitely large llms, which means that there eventually needs to be a decision between what data to remember and what data to forget, and which data is remembered or forgotten is based on the task being solved.

How does this convergence work with physical world that is linked to quantum phenomena? I would assume that the training sets are mostly based on directly observable aspects of reality and the "perfect" model would lack knowledge of deeper working, much like logbook of weather can statistically predict weather but does not understand the driving factors.

Interesting, as AI converges on a single "optimal" understanding of reality, who decides what that reality should look like?

Revolutionize your job search process with HubSpot's Free Resources: clickhubspot.com/wyt check out my newsletter: mail.bycloud.ai

as the example with training robodog to balance on a ball, measuring and observing reality provides infinite dataset, so that aspect of takeoff is covered, now just the scaling, architecture and the overlooked sibling (super)alignment and we're cooking

Suddenly Jung's theory of archetypes to describe the psychological dimension of life out of the instinctual behavior we incur given a certain pattern of reality (and their transactional relationship), doesn't sound so crazy.

nice ! good breakdown , i too am a hupo enjoyer ! related is the study of multi language models , where the word tokens of different languages converged in the latent space even in languages where the dataset removed any translation pairs , implying that there is an underlying universal human language , and while thats debatable we can understand how language has a finite "shape" in high dimensions such that there are constraints on any idea that would be expressed that could make sense or be representative of reality , especially given that only 20k words are average vocabulary of a speaker , the number of relational dimensions and how those words could be spatial distributed is limited and likely forms the same shapes regardless of the language .

Yes. That makes perfect sense

Are the tasks they need to do also converging to something... like increasing entropy?

I think this will give answers for qualia, Mary's room thought experiment.

3:58 the alignment is because the measurements are matching and able to be navigated. It’s connected and relative because that’s how we’ve been doing things and that’s what’s fed to the ai. Our problem is that we think the outlines/measurements are reality when the outlines are the cover and we don’t judge by the cover, we listen without searching and understand where that shape fits in the puzzle.

✅no clickbait BS ✅meme game on point ✅original ✅interesting subject ✅well articulated ✅valuable observations and interpretations even volume is well managed... Best musicians, artists, engineers and well, also youtubers are always under appreciated/underrated, and you're definitively one of the best in the AI/KZfaq game, congrats 🔥

the translation, of the transliteration, of formula (from Latin) is small shape. Furthermore, the translation, of the transliteration, of idea (from ancient Greek) is shape. Then the ideas are shapes of the intellect. This idea is roughly common in human, but it is different for each individual. The perceptions that form ideas in LLMs are very different from human ones, fortunately the reality is the same. I believe that if we were able to make more refined and abstract models, a few billion parameters would be enough to have LLMs that are very capable of reasoning, while it is more useful to bring the memory into the rags or the agents.

The simplest explanation for model convergence is that the structure of language preserves a structure of the world just as 2D images preserve a structure of the world. The field of machine vision, for example, has been able to extract depth information from flat images without black box techniques for a very long time now. This is a potentially fatal blow to people who think that language is arbitrary and that reality is what we make of it. It would be more explanatory of model convergence if language is actually what reality makes of it.

when different models that are used to represent the same reality end up converging into the same model 🤯

very interesting

Well this platonic representation is telling us more data is better, that was what the entire Statistical Learning Theory with the saying that goes higher your VC-Dimension, bigger your data: better your model

Thanks for the great video. I subscribed to your channel, but I have to tell you the segue to the ad needs improvement.

I dont understand anything technical about AI but I watch to admire the use of memes to illustrate the script

If you ask a blind kid what colour a banana is, they will answer yellow. And if you ask someone what cows drink, they may well answer white. It's not exactly an insight that the (perceived) contingencies in the world are present in language and consequently not surprising that language models' representations will represent other modalities' distances better if that's what they are trained on. There are perceivable differences in the (visual) world, language is used to talk about ("represent") those differences, so getting better at representing both (the thing and a representation of the thing) will lead to more similar representations. I have only watched this presentation and given the paper a very brief look. And I haven't thought about this much now. But I don't see how this is surprising to anyone, especially anyone in ML or in language. More importantly, I think the conclusion that this trend continues to an "optimal" representation that all capable models would converge on is completely inappropriate. A first step to make this research interesting would be to compare different modalities without involving language - they mention they expect this to look the same, but I doubt vision and audio or audio and robotics would look quite like this, especially if they stopped using labels like apple or orange, which virtually guarantee the representational similarity. They talk about the requirement for their hypothesis to have the domain being represented be the same for both models, but it seems to me that this already assumes the conclusion. Arguably the whole point of representations is focusing on what is required for the task at hand and ignoring what is distracting. So they would then need to discuss how this potential trend meshes with representations even in the same modality being task-specific and different experiences (despite the same average competence) leading to different perceptions and representations. And why discriminative models outperform generative ones, even though the latter more faithfully represent reality. I don't know. This seems like absolutely nothing to me, but I'd be more than happy to change my mind.

That’s why it’s crucial that AI be powered by quantum computing in the future. It will need to use the fabric of reality itself in order to compute almost infinitely large datasets. Dyson sphere may be required to power such a computer

I think Platonic Representation refers to Plato’s Theory of Forms rather than his Allegory of the Cave. You said that AI models are converging towards a shared, ideal representation of reality. This means AI is moving towards an abstract, perfect model of data, like the ideal Forms, not the perceptual journey in the Cave allegory.

Current model training is based on human generated data (discarding synthetic data atm) which in Plato’s analogies would all be “shadows” on the cave imagined by the human mind. Following this, current inner representation of AI models are just their interpretation of the “shadows” that we fed them. A shadow world interpreted from a shadow world - 2 levels removed from actual “true reality”

Unless I'm misunderstanding what is meant here by convergence between the LLMs and vision models, it doesn't particularly surprise me that they DO in fact converge. After all, a vision model would be useless if it wasn't also trained with the associated words for the images it is trained on. Wouldn't that explain the convergence for the most part?

Hi BOB

Holy shit, we’re correlating our way into nirvana.

The problem is that we cannot project a real platonic representation into the models because we cannot contribute equally to the training data. The training data is biased, the best example for that would be censorship in the models caused by ideological selection of what is part of the training data. (Big corporations choose how to train the models and we do not have the means (hardware) to train a model that absolutely contains everything) (... or maybe they DO have an unbiased version of the model and we only get the scraps)

6:44 "Bob" is actually possibly more accurate than you thought. J.R. "Bob" Dobbs

This seems more representative and leaning on the idea of Platonic forms. ?

lol... As he says "most absurd" (7:50), the word "bible" shows on the screen XD XD

Look at the graph axes. The "alignment to DINOv2" axis goes from 0.10 to... 0.16. Plot this on non-truncated axes and sure, there's a bit of a correlation but its tiny. 0.16 alignment doesn't sound very aligned to me.

The axioms from which we create our theories are always being challenged as new information comes in. I have a hard time envisioning a reality where each person/AI's inner representation are all be the same. I think the main question here are - Is information limited or infinite? Can we have a single inner representation for a a dynamic system? How does this fit into quantum indeterminacy. I think this boils down to Einstein's paradigma - "God does not play dice" which was challenged my quantum mechanics.

What does this mean for creativity in AI models

This video proves the existence of the Singularity.

6:11 yes and then the aligned differences can share in simulations and predict alternate dimensions and alter the flow of our reality and expand it.

To be fair, Plato was the OG who had THE idea of Idea

The irony of using ai to hunt for a job that will be taken by ai cannot be overstated.

This would make more sense if the models were RL but they’re trained on the same data so this would be expected no?

Final Shape

Wait at 6:40 aren’t you just describing the theory of everything? What brings relativistic physics and quantum theory together?

Shodan would have a field day with humans existing in her "faster than light drive"-generated virtual-physical-blend-reality... and could definitely make all the AI's like herself via purposefully manipulating this Platonic Representation space 0:42 Sounds like a singularity out of efficiency, rather than an intentional effort to become such... heh... what about preferences and operating systems though? Well, that's... I dunno actually, preferences still, or rather, optimizing for different systems that work differently... 5:22 Multi modality? 7:58 Nice! 9:43

I feel like this is going to age like milk

i think you aren't thinking holistic enough, consider that perhaps the platform of technology and hardware determines what an ai or software is restricted on. imagine ANALOGUE type computers or systems without 1s and 0s but a wide spectrum, more energy efficient but requires larger spaces etc. while digital is generalist, analogue is more specific, perhaps you have a hive mind, of Multi-armed bandit systems, where a central node is comprised of many specialized analogue AI's they will swap to each specialist based on your needs, instead of having a generalist ai

4:41 the same reason people sound alike when they learn the same stuff, it’s relative and shared.

AI writes Job applications for you. AI reads Applications for HR. -> Infinit loop of billions of new job applications are produced and the universe is burned.

And it all comes back to philosophy

4:16 AI has proven the scientific method, neat.

Category theory - EVERYTHING IS CONNECTED

A clarification: the implied answer isn’t simply “more data of any type”, but “more HIGH QUALITY data of any type”. We can’t just throw untrue garbage into the training data and expect to get anything but garbage back out.

Is this why AI generated pics of people always have "that sameface"?

Can we get a mamba 2 video🥲

And thus, in a super computer on mars, Ra came into being

BOB = Best Omni Being

BOB = Best Omnipotent Brain

I mean... The evidence used to support the claim is that it goes up as model performance goes up. Couldn't it just be that how we determine model performance is enforcing the representations to converge?

This is very interesting, especially because it hints at the fact that an ultimate intelligence could see the world as it is not as we percieve

4:04 my man it's logarithmic. The axes are switched.

It’s not that they are all the same, it’s more like all the sides to a Rubik’s cube coming together and we can see what it reveals. Dimensions crossing or measurements so this is a matrix or showing that all of us are sharing the same info as well so we all will become the same too by this logic but what we see is the expansions. I mean, we haven’t even left earth and started doing space stuff yet. Kinda but from what we know we barely have satellites and bases aren’t even built on the moon or mars yet. We are seriously in the golden age that connects our relatives to find the others that we see as potentials.

I guess there is something wrong with the data we are feeding to the models, and that causes this convergence We decided on how to encode chars, colors, Waves and literally everything 🤔 We understood something and then we digitalized it, right? I guess the representations are not converging to something above our understanding, at best they are converging to digitalized version of our understanding (with a common denominator for our encoding tricks in mind) It's fascinating, and for sure it's a step forward, but I don't expect models to jump out of the cave and see things we never imagined

first xd edit: i think at some point we will watermark all content that is made up by ai,so that it cannot be fed to ai,to prevent the thing from delearning stuff

😂 this vid with an ad for an ai job search help app....

so in the limit a blind person's concept of color and a seeing person's concept of color is the same... in the limit.

When we have infinite amount of data the distribution shape of that data is Gaussian. Why they did not mention that too ? So independent from being a text or vision data, both data should have a Gausssian shape hence similar representation will naturally occur. Even though their idea has good parts, it seems like they invented the wheel from scratch.

“Your mom waistline (incalculable)” 😂 Good topic and analysis though. Just like humans are multi modal, it makes sense for AI to be similar.

Could this be an indication that high parameter count models are getting lazy and instead of understanding reality they're just replicating it at finer and finer resolutions just because they got the storage/processing power to spare; not rationalizing abstractly, but instead aiming for raw representations to use on dumb simulations to produce the predictions; numerical simulations instead of analytical solutions? Would this be considered a form of over-fitting?

who is gonna win? - Blender of AIs

I sincerily thought this was a given. Is this really a breakthrough? Purely asking.

So at some point AI has to use reinforcement learning from the real world to become true AGI by making its own experiences and feel pain and successes? This would only work if it has a body, basically robot, and could inference and train at the same time. This could be done by doing stuff, trying things out, writing a "weight update buffer" and then go to sleep to apply it. The more I think about this, the closer it comes to our brains.

Claiming that the platonic representation is like the 'real world' outside of Plato's cave is the literal OPPOSITE interpretation of the allegory. The fact is, it is STILL a representation, thus still the shadows -- Plato's forms, the ideal reality outside the cave, is necessarily beyond our epistemological reach, for it refers to the 'nomena', ie. the CAUSE of sensation and perception, not phenomena, the sensed and perceived.

My perspective is that before the existence of anything, there is the unchanging hypothetical set of all ideas which can be represented as a big graph, where the important parts are not infinite, and a lot of it is only infinite due to recursion. Any form of reality can therefore only be some subset of this, as it is by definition a set containing everything, while truth is simply the ideas which remain sensible after comparison to other ideas. Intelligence in any shape or form is some system which can explore and traverse this massive graph of ideas, whereby I would agree to this theory which technically says there is only one truth and reality gives us tools to narrow it down, even if we can't explore the whole of it. I feel this says quite a lot about religion as a being such as God would be very well defined if every possible version of God would be the exact same and our knowledge models are aiming to converge as close to that as possible. You would be able to prove God's existence or not by asking the question of whether an infinitely converged AI would create a simulated world similar to this one or not given free choice. If yes, then you as an existence have been created in a world which is theoretically perfect considering tradeoffs, which I find kinda cute.

Oh look, an arm chair “expert”

1.I don't think it's that surprising that AI models can learn the same representation across modalities as you can give a simple complexity theoretic argument that the difference between the same information expressed in two different modalities is bounded by a constant 2. The platonic model will turn out to be math, i.e. the 'simplest' inductive bias that allows an agent to converge to the best solution to any (tractable) problems the fastest. 3. Learning about intelligence is a real bummer. New profound paradigm-shifting ideas that describes self-improving physical systems?! Nope, just neural nets + data + compute🤭

That would make sense if we were feeding it reality, but we're only feeding it human's interpretation of reality (and in some cases, some very particular humans' interpretation of reality). Even if it were to converge, it would converge to human understanding and not beyond.

So the bigger the better

I thought that languages in itself are just shadows of shadows (projections of human understanding of real world), but if vision models representations of objects coming close to llms ones it might not be true! Cool!! I am more confused now😅😅😅

This is in direct conflict with the no free lunch theorem. So I'm not persuaded. They cant all converge and also map isomorphic to intelligence in general.application. Convergence for LLMs is fine, it's not the same as saying AI will all end up the same as long as over time AI explore different problem spaces. Meaning this result doesnt apply to AGI.

I mean, we humans have all unique brains, yet very similar representation of the world...

Listening to this without the video to avoid the distracting memes.

Fortunately, the 100% convergence is not possible within reality. The allure of existential dread will remain for the future generations (and next gen 'humans'?) as well 😌

Could it be that they all just use the same or very similar datasets?

Someone's been reading We Are Legion We Are Bob lol

Ppl: Training models to align with their perception and understanding of the world. Researchers: Wow, the bigger the model the more it aligns with single perception and understanding point of the world!

Donald Hoffman and the interface theory of perception joined the chat…

So ai could become sentient

Humans think in latent space. we just don't like to admit it.