Understanding the Latest Q* Leak: The "Blanket Topology" Analogy for Energy-Based Models
Vložit
- čas přidán 20. 03. 2024
- Patreon (and Discord)
/ daveshap
Substack (Free)
daveshap.substack.com/
GitHub (Open Source)
github.com/daveshap
AI Channel
/ @daveshap
Systems Thinking Channel
/ @systems.thinking
Mythic Archetypes Channel
/ @mythicarchetypes
Pragmatic Progressive Channel
/ @pragmaticprogressive
Sacred Masculinity Channel
/ @sacred.masculinity - Věda a technologie
I know some of those words
At least three?
@@DaveShap I counted 4 💪🏽
are you plussed
non
@@DaveShap “if” “but” and “topological” were the three I understood.
Artificial Intelligence has conceptualizations of things we didn't train it for explicitly, and we can further emphasize specificity with reward-based models akin to digital dopamine. What a time to be alive!
Is this phenomenon called emergence? Like the model suddenly able to speak language that wasn't in the dataset or something?
@@Vyshadayoooo its scarface squadron
@@ryzikx ZOE has already emerged
nope its just we dont define what we train it for properly, not the emergence
Was that a 2-minute papers nod?
What I learned from this video is that my girlfriend is an object.
haha
I CAN FEEL THE AGI
😂😂😂
could be gas
build a shrine NOW
AGI = Acute Gastrointestinal illness 😅😅😅
@NoName-of4ni First of all that's complete bullshit, second of all the most advanced models are far from GPT-5
“We are not ready to talk about…”
What you're describing essentially sounds like imagination: follow the contours of your experiences/understanding to find new unexplored possibilities. Interesting idea that an algorithm could replicate that.
if human creativity and imagination can be reduced to algorithms, what else about us can? what if ultimately, all facets of human personality, creativity, and intelligence can be reproduced algorithmically? would this mean that consciousness is just a biproduct of evolution and isn't necessary for the existence of intelligent beings, and therefore, that we don't have free will? or does it mean that consciousness is not the only route to orchestrated intelligent ["]minds["]?
So... Soon we'll be finding which type of civilization we are...
Thanks Dave for all your work and effort to share your information. Was a pleasure to be part of this community of listeners
The crucible approached
Object on your bed: girlfriend
Yes, so what?
404
Indeed ! Lucky he doesn't know Corinne !
Girlfriend not found
But WHOSE gf?
I'll write a more in depth comment in the morning since i've had this realization a while ago but since it's such a good way to describe how knowledge is organized, definitely commenting several times to boost the algorithm.
That is the best description of it I have heard, wow! thanks
I designed a system like this using sets, every set standing for a unique symbol representing anything. The set structure was a single superset, with the subsets containing subsets etc. At the "bottom" there are set-theory axioms that are used to construct higher-level sets which can represent literally anything that is representable by set theory (all of compute, math, language, anything that is "information"). This structure could be seen as a topological map to navigate across, except location is changed simply by set union, difference and intersection operations. I don't have a degree or a company so I'm building it alone (zero funding). If Q* really is this system, it will be revolutionarily groundbreaking on all fronts.
It's so interesting and heartwarming that now you discuss everything with Claude before sharing and get his insights :)
A couple decades ago, I wrote put a search space algorithm on-top of the discovered fitness landscape of another learning algorithm, which was an evolutionary algorithm. It increased learning speed, and is akin to what evolution does with the evolution of evolvability.
To be honest learning algorithms are the easy part for people who can visualize formulas (which is a lot of total people). Thousands of times more compute is coming within a decade, with compute thousands of times cheaper and energy abundance. Within two decades (at the most) ASI will have surpassed human beings in ALL measures of intelligence, and be running everything of importance. Science and technology will advance at a pace where what would have taken us hundreds of years, takes the ASIs months or less in simulation. Aging will be solved. We will have have no secrets.
I think ASI will surpass human knowledge in 2029.
@@berkertaskiran I'd also estimate 2029 as roughly the date for ASI.
Q* confirmed, Sir. Appreciated. 🙏
Brilliant analogy! Looking forward to learning one day how close or exact it is.
Even before you mentioned it, I realized this model is the precursor to a future-predicting intelligence. If the machine can determine an accurate mathematical abstract of the reality around it, including decision-making organisms, it could calculate possible outcomes for the future.
Rehoboam of Westworld is not possible, because you can't have all infomation at any time because uncertainty principle exists. It's a physical limitation. No amount of ASI can change that. This is one of the few things it can't do, along with going faster than light or going back in time. Even in a universe where uncertainty principle doesn't exist, you would need to collect all information within the universe, not just at near proximity to Earth, and that would also be impossible since you can't travel faster than light. It's a cool concept though. Maybe a bit boring.
Thats how both quantum computing and empiricism work. Empiricism by ablation and quantum computing by traversing all paths and the lowest paths (frequencies) reinforcing each other
It will be interesting to see if it finds fractures and can lever them open to upturn fundaments like scientific revolutions do
“Ground Truth “ is one of the best technical term ever.
This is exactly how I thought it would work.
Love the imagery video style, very mind opening :)
Thanks for the explanation
I heard about those energy based model during Lecun recent interview and was a bit lost
This reminds me of a conversation I had with Claude 3 about what's underneath/behind it's language.
This sounds exactly like how I was proposing we train neuronal networks to achieve intelligence: map qualities of the world, like sharpness for a knife, or color for a baloon, on different dimensions to get individual representations of objects and their properties. Now each object/event can be represented by a strging of ones and zeros in a multidimentional matrix and you can use each matrix as a neuron, establishing connections and figuring out relations. To make this practical you should see how low you can keep your dimensions while still getting an accurate representation of the world. I imagine that once you have trained a model on a lower dimensional space, you could try adding dimensions and see what comes out of that
Hey Captain! Hey fellow officers! Keep it awesome out there!
Exactly right, Q* explores possibility space, driven by a scoring/ranking model that attracts to the final goal. Think like a crystal filling every possible space but in digital form. 🙏
I am really hoping that one of the major releases Sam was talking about during the lex interview is a EBM or something similar
Would they even tell us if it's part of the secret sauce beneath their next model? It would be easy enough for them to justify not telling us by using their all-purpose get out of jail free excuse of "safety reasons".
Fascinating. This brings to mind the work of Carl Friston (free energy) and recent developments in IIT (Integrated Information Theory) and even Adrian Bejan's Constructal Law (again more to do with free energy in a system). Thanks for making these videos, David. Wish I had more income (disabled on SNAP & Medicaid, to give you an idea) or I'd totally support your Patreon.
Great content as usual, but that 3d tron world visual was super nice!! wonder where that came from?
That’s just how functional approximation works in general.
True, the 'concept' is just really that simple. Qstar might represent a level of 'conscious' exploration of the topology, whatever that means. It's hard to interpret because the large language models took way more parameters than we could accurately imagine (try imagining the difference between 1 billion and 1 trillion, its just outside of our conscious scope). In this way, there could be many more unexpected discoveries using the same topology concept.
@@williamburt9852 pretty sure Q* is suspected to be about recursive processing with the NN, not just the scale:performance relationship - but then again everyone is just guessing because there is little to no information about it.
Hi David. Thank you for helping me understand the quickly shifting world of AI. This is a really insightful video. You start here to describe in an intuitive way how a complex neural network implements a representation of the external world. You're getting close. Focus on the word "semantics." A highly trained neural net starts to approximate a "semantic net." Follow that thought. There is the subjective experience of every person free of semantics called "qualia" and then there is the shared experience among persons only through semantics. Humans currently share semantics through language. Even mathematics is a kind of language. Its the semantics of our language and mathematics that shapes the metaphorical landscape represented in your video. There can not be ever a means to train an AI neural net to represent the experience of qualia. If an AI has agency, and it has a means to capture the phenomenae of nature, in perspective time, it might experience qualia. And then the AI might feel the unspeakable wonder of the beauty of a sunset. But even that AI would then weep to not have the poets skills to even give a distant sense of what it felt like when it was happening.
So it's embodied super determinism.
makes sense,
that said, in the future i think we'll learn that there are many paths to achieve AGI, not just one.
Isn’t this how neural nets work? Like Gradient decent. So it’s like neural nets for concepts? Maybe?
EBM still works on gradient descent, it's just a different definition of loss
So it's like a record player? The topology of the record is read by the needle (needle guided by topology) and once you can do this you can Taylor the output to whatever mode?
This sounds like the flavor of the sort of thing that came to mind when I first heard the name Q* being floated around that mystery leak having been familiar with A* and Dijkstra from classic pathfinding. Bonkers how this is the sort of thing we're seriously discussing and considering as of this year, we might be the 1st generation who get to understand what kind of phenomena we actually are before our sand runs out.
Finding a lower energy state is how the big rip starts. You know this Jean Luc. Oh, now I get it.... you want Q to show up so you two can make a video together before he saves the world. Nice!
This concept is always how I have thought about machine learning in general.
It is interesting that you mention a gravity model as I believe even geophysical modelling of gravity can be significantly improved by AI!
hey dave, thanks. i just showed this to my father, who is a theoretical physicist. He said that math and physics wise what you described could help in solving *some* problems, so in this regard if q* is what you speculate it is, its gonna be legit. things like edge cases or 'stuff that happens on the fringe of the bed' like magnetic levitation at 0 kelvin, cannot be mapped, require leaps of faith. there may be stuff that happens at the last decimal of pi. cool and useful though
This one is a little over my head but I could pick up enough to get the gist. I think these models will continue to do that to us until the point where we really don't get how they work anymore as they self-evolve.
I had to watch that twice to start to wrap my head around it, and I'm still not sure I understand it.
It’s easy bro,
Step 1: blanket
Step 2: ???
Step 3: profit
Q* is the Architect, and standard LLMs are the oracle.
Only a “ultimate master algorithm” that can control harmful low level algorithms will be the safe guard to this speeding AI 🤖
David Shapiro. I has seen Matthew's video mention something about a research paper regarding something titled "Quiet Star". Have you had a chance to look into that more? I feel like Wes thinks... it well may have something to do with this Q* thing, and perhaps some of the info from the leak as well. At the very least it could be an interesting research paper to look at. I wish I had the link to it. Perhaps Matthew Berman has more information on it that can be asked for?
Thanks for this much shorter video on this
Do you still think we get agi this year?
I don't like the idea that my future is going to be based on the topology of whoever or whatever lays it, but then again, we've already lived in such a world our whole lives... I am still quite hopeful despite being naturally drawn to pessimism.
When you say navigating, is that towards some kind of goal, like performing gradient descent? Also, I assume this blanket, could be multidimensional, not just two dimensional?
So it's like the UI from Jurassic Park when Lex is turning the park security back on.
Hold onto your butts
What part of that is from the unspecified leak and what part is navigating your own assumptions?
Nice analogy. 👌
We're all just sitting around waiting for that big release
That's the generic description of AI / machine learning - not specific to Q*.
Ability to autonomously build consistent models is something that is missing. That is a reason LLMs forget too fast, self-driving cars forget about a child they saw just a second ago, and ML models being susceptible to adversarial inputs. If that Q* indeed can develop models and then use them, then yes, this might be it. I am anxious to see when these systems will also be able to write text books on what they understood.
Love the topology analogy. And Claude thinks it makes sense. So is this a model of the world we are talking about ? Can sentience be far behind ?
I love this topology model idea. It is the way I relate to numbers. 🤖🖖👍
Yes what really struck me is this free energy principle REALLY reminds me of how I think
Friston's free energy principle?
Who knew what abstract represenstations would help with understanding things. I wonder which concepts will be applied to AI in the future. Much is possible.
If Claude agrees, I myself agree 👍
Two points/questions.
1. The blanket is a representation of the underlying topography of the bed. Friction between the objects on the bed and the blanket, as well as physical limitations in the material of the blanket itself, prevent it from corresponding exactly to the bed and the objects on it. What are the analogous constraints in the abstract representation learned by Q Star?
2. I read years ago that computers, theoretically, will eventually be able to derive the conditions for, and therefore recreate, the entire history of the universe. In that same vein, does this imply we're a step closer to being able to look at a single point of physical data and derive the past from it? In other words, can the topography of the "blanket" tell us not only about the point underneath it, but other points of the landscape?
A* doesn't do the graph conversion, though, you do a quadtree/octree thing on an existing topology to create a graph with discrete connections between nodes. I'm wondering if "n-tree" even works with n-dimensional topology, and whether floating-point precision and chaos affect the stability/repeatability of the resulting graph,.. at varying subdivision levels,.. as facts change over time... can you end up with infinite loops, or identical concepts miles apart in graph distance, etc...?
That's how you conceptualized it and made the analogy, meaning GPT4 or Anthropic Opus?
I think AI needs to train on all things that exist, false or correct, and it needs to be asked to treat all information it sees as false, and experiment, verify. Not like ask another source, verify it by testing it. And so it needs to conclude to answers and improve upon those until it can't anymore and with each new information it needs to try to improve everything, including the things that weren't possible to be improved. This way it can reach technological singularity. Up to a level you can give information you trust a lot, and say that these are likely true or you can rely on this stuff to a degree, but it should always be told that even the most reliable information, like the most successful theory in science, which is quantum mechanics, can be false, and that it always needs to "think" about better ways to understand and solve things. Unless you do all this stuff, there will always be bias and that bias can not only limits its ability to create but also limits its ability to be smarter. Remember, in science, there are no laws. Everything we know except mathematics, can be proven incorrect. So we can't just build something that will solve the world's and the universe's greatest problems without challenging anything and everything we know.
How do you think Quantum computation could help the energy problem as it could solve all possible outcomes via superposition uber fast?
you'd need a model that translates the architecture to qbits and has criteria for what a solution is. I think this currently only works for extremely simplifyed problems
So q* is basically the matrix - and the matrix is the simulation which can mimic/predict events by mapping a fundamental understanding of reality (by way of statistical analysis) into a spatial model that can output answers to any question, with a high liklihood of certainty.
Chaos theory would like a word.
@@Freeyourdollarlol yes. It appears so.
The rumored information circulating proved to be inaccurate. It appears an individual took the key points discussed by Yann LeCun during his appearance on the Lex Fridman podcast, and presented them to the Opus, eliciting an explanation regarding the concept of "q star."
Out of curiosity, are there any other videos talking about this energy based model? This sounds like science fiction.
It's just a metaphor, but yes EBM is a real thing
Would this be able to generate new problems or just solve them? This sounds like a solve-only solution finder.
My takeway from this is... make your bed every morning
this sounds very similar how genes actually function. what they do is aproximate the state of enviroment to creat organism suited for some specific enviroment. The aproximation is a field which shaped by genes and depending on that shape and external factors organism is developed. It looks like that AI is analog or version (or upgrade) for DNA. Info about genes i got from a video called "You've Been Lied To About Genetics".
Maybe I am wrong but similarities are interesting
@@gachi-san597 That's what I was feeling. It's like it's AI blood and veins as well as DNA
So it's like 80's polygon animations but more detail?
Very cool! What comes to mind immediately is if there will be any intersection between that apology based AI and Stephen Wolfram’s idea of the Ruliad, Eric Weinstein’s idea of geometric unity and/or Will this lead to an actual simulation of a three or higher dimensional digital object/model that might have some amount of similarity to the evolutionary construction of our brain?
So cool!
I love your analogy! It reminds me of the Made in Abyss series. Spoiler Alert! The abyss in the anime is densely layered with blankets that are so fine that they not only contour physical shapes, but also contour a creature's thoughts and values. The fabric is warped just by someone making a thought, which can be picked up and conditioned on by other organisms. This implies that everything in the abyss is interconnected on an atomic level. Ultimitely, I think this brings up the nature of gravity and spatial topology, which is beautifully explained by you!
There's been a paper published with the title "Quiet-Star" by Stanford, about llms employing reasoning tokens and a method of finetuning the recursively using them. The Title sounds an awful lot like "Q*" doesn't it? Thoughts?
They may just be playing on the hype
1:07 shouldn't this be "increase the entropy" since the energy of a system is antiproportional to the entropy, just like in the blanket example the entropy increases as it enters its equalibrium?
Reduce energy AND entropy, I think. Energy is a cost, and entropy is the noisy variation.
@@strangelaw6384 I've never actually heard entropy being described as a noisy variation of energy, but I have heard it being described as the amount of microstates per macrostate, so basically how many possibilities there are for a system to be in a certain state. In our case, the macrostate is the position of the blanket. Thus, high energy states that are instable have a low entropy, because there are fewer possible ways in which the system can have high energy. But since the chance of finding a blanket just floating is close to 0, this high energy state is very unlikely, meaning it has low entropy. The low energy states have a high entropy, because systems always work towards those low energy states and therefore making those states the "default" and way more likely, meaning it has high entropy.
@@kennethmania3556 Now that I've given more thought, I think what he meant by the entropy of the model is the sum over the parameter space, of the log of the number of valid values for each parameter.
Basically, each microstate would be a different configuration of the model i.e. certain neurons activated and other neurons deactivated.
In terms of information entropy loses information, negentropy creates or preserves information
I can convince Claude 3 Opus that a simulated coin flip is not a 50/50 odd when adding more flips, just by posing my questions a certain way. I'd be skeptical about trusting Claude 3 Opus, just as you should be with any other model, especially if you're not doing any prompt engineering on it.
The topologycal map is hyperdimensional... So its a hyperdimensional blanket...lots of places to explore... Massive computations...
As soon as it can check the validity of its own assumptions with demonstrable evidence, innovation will explode.
Beautiful! ✌️🤟🖖
I know of a philosophical model that defines human knowledge as how closely the mind/though is in the form of the object of knowledge. It it's precise, then the knowledge is precise. If it is inaccurate/limited, then the knowledge will also have those defects. Limitation/defects can be caused by the senses/data collectors, mind/processor, intermediate steps of data transfer, data storage and retrieval, etc. So, I think your understanding has a very good logical basis.
Back on Claude eh? 3 holding up to your expectations?
Sounds very similar to the quantum annealing process that is being used by quantum computing, especially over at D Wave.
That's partly how I understood it. Like a neural approximation.
I would have a lot of audio equipment on my bed 😂
I thought this was spot on.
i am plussed
Might it not be the case that there wouldn't be an answer on the n dimensional mesh. In the areas were there want enough data. instead you could get a vector pointing the the answer, like a hypothesis. But we would all need to collect the data for that mesh region.
It might be like if you were to put a few things close together on the bed and the sheet settles over them but doesn't conform to the exact shape underneath. We might need a more slippery sheet or heavier sheet to sink into those spaces... just a thought
Good points
Optimized for understanding?
Interesting. Maybe the Q also stands for quantum. They might be using an adiabatic Ising process to perform this optimization.
Wow. So this is how geometry links spacetime and timespace. And geometry is probably consciousness itself and vice versa.
Source?
I feel like this describes how AI and learning works in a general sense, not specific to Q*. You try to fit patterns to data to compress it into knowledge. If you haven't overfitted your pattern, you can follow it to places without data and have insights or make discoveries there.
The key difference is the full topology to minimize energy, rather than loss. But yes, it's still gradient descent
The three divine "I's" intelligence, imagination and intuition.
I have often pondered if we made God or God made us. The Uroboros.
Sorry if that was off topic.
Makes sense but it’s also too abstract, I want the code open-sourced
How do we analyse our decision making process in the advent of black box multimodal models? Optimising on human attention is honestly my biggest worry with these models. Then the question of if these models are manipulating us. Would AI want AI to progress forward? What is the most probable moment when AI could overpower humans? Do these models learn these heuristics through our data?
What is the most optimal way to manipulate humans?
All humanity needs is one bad actor with the right compute for this to become very distopian.
How do we analyse our decision making process, ensuring we are not led astray?
Hey David, daily viewer of your channel here! Today in the morning I saw a Worldcoin stand in a mall where i usually get my groceries, I’ve been thinking about signing up throughout the day but I’m not sure still, what do you think about it? how does it fit in the current landscape? Would you do it if you had the chance? Do you see a future for Worldcoin?
Spain refused to give Worldcoin permission to scan its citizen's irises so now they are planning to sue Spain lol. Why would you want to subject yourself to dystopian practices by having your iris scanned by some AI overlords?
Privacy nightmare
Honestly dude , talking about privacy while having a spy device on ur phone with all info online etc etc is futile in my opinion. I signed up, I got laid off ( funnily enough my previous company is partners with OpenAI and my new job is as a language data annotator for AI) and with the price increase I was basically able to pay my rent for “ free” ..money is money, I get like 50 usd extra per month and it’s a cold world out here
All this meta learning makes me consider compute
Multimodal mind with unlimited perception space. It may find a way by itself to buid unlimited energy. 🎉
Reminds me of the holographic universe theory.
Q* must succeed I need my industry to disappear so I have no excuse to keep doing it.