thoughts

Looking into this...

The more I look into this, the more I think that neural networks and deep learning technology is the most important thing we need to understand.

The concept/structure is relatively simple. Neural networks are things that break up an input into small single packets/units, assigning those parts to a "neuron", and then transforming those inputs across multiple neurons until we get some output that we deem to be useful. For example, a digital image of a cat gets each pixel assigned to its own neuron in an initial layer of neurons. Then, a neuron in a next layer takes inputs from the neurons before it, multiplies it by an adjustable number, adds them up, and then transforms the result by some mathematical function. This process repeats many more times through more layers until it comes out as some output. We then examine the output and adjust certain numbers associated with these mathematical functions (weights and biases) until we get something we deem good.

This mechanism is a fundamental departure from how we typically explore things. Usually, we go at something with hardcoding certain rules, then keep iterating those rules until we get something good. Neural networks and "deep learning" starts with data first, then a certain goal we deem to be good, then we give feedback to it until it produces that good thing we wanted. Neural networks and deep learning are essentially a pattern matching machine.

This seems great, but I get the feeling that the more we push in the direction of ignoring the processing and how it works to just get better outputs, the more we lose how valuable these technologies can be for us.

There's been a huge push in almost every academic field to forego fundamental understanding for better outcomes. This is good in areas where the practical things we can do with the technologies greatly benefits our day-to-day experiences. If our current understanding of quantum mechanics gets us to a point where we're making computer chips, then who cares what the philosophical underpinnings of quantum mechanics are or what it says about how we should interpret the reality we experience. Large language models (LLMs), a technology built on neural networks and deep learning, make work easier since they give output in a form we can literally understand (language), and it seems to produce genuinely great things that help us work. But true value always comes from an understanding of the underlying concepts; from reducing the chaotic complexity of things into something simple.

As an example, your ability to communicate ideas and concepts gets better only if you get a better understanding of what communication is. The message becomes as important as the mechanism behind getting the message across – the value comes from the message's ability to be compressed (sent) and uncompressed (received) without losing what makes it useful. If we went at communication with an attitude of "let's just simulate 6,053,183,229,301 conversations until we get the one that seems the best", then sure, we'd get some pretty good explanations, rhetoric, and maybe even come up with some new techniques for speech. But, we lose all fundamental understanding of the thing itself, so no matter how hard we try we could never hope to find those laws and structures that make us truly capable of mastering understanding and thus communication.

I think the wholesale adoption of this method of "run it a bunch till we get something good" for LLMs and neural networks is based in our relatively new utilitarian focused society. If we can't do something useful with it then we don't assign much meaning or value to it at all. But so much of the greatest movements and discoveries in the world were made from trying to just understand what's happening.

It's not clear to us now how understanding everything that happens between the input and the output of neural networks can help our lives, but I can bet my life that something more life changing will come from it than building the greatest AGI. The AGI will benefit from all the work we put into it, but we would get no closer to understanding anything fundamental about the world.

The future you were promised does not exist.

Further below this I've put a transcript of a conversation I had with one of these machines built on neural network/deep learning technologies, a Large Language Model (LLM) that is called Claude Opus 4.8. While it cannot be said whether or not we can have an actual dialogue or conversation with these tools, they are astoundingly useful in many ways, and as a technology they are genuinely innovative and do what technology should do, i.e. be a tool that helps people better manifest ideas they have.

I've used Claude a lot to work on personal projects, research, do university level physics problems, write YouTube tags for the algorithm, and a lot more. (I also try to keep a markdown formatted file vault that stores all sessions and memory for it, leaving room in the future to transition my workflow to a local model when that becomes feasible.) One of the things I refuse to use it for, however, is anything related to my personal art, whether that be writings, music, or videos. I think this refusal from me and many others is indicative of a fundamental problem with these LLM technologies; a problem with the philosophy behind how these tools are designed to be used by those creating them.

These tools have been hijacked by incentive structures that are not aligned with what is good for a people. That word good is doing a lot of heavy lifting, and is ultimately something subjective (though it's definitely not defined as whatever makes the GDP and capital of a society go up). Typically, what is good needs to be posited by a person, discovered through a mix of intuition and reason, and then agreed upon by a consensus of people. This idea of good should go through rigorous tests and analyses before ever getting to the stage where it determines the goals and objectives a collective group moves towards. Unfortunately, deep learning technologies progressed so quickly technologically, and in an environment so opposed to questioning the philosophy of things, that it has gone in a direction that incentives attention and money above everything else. Generally, these are decent metrics to base goodness off of; if someone offers something of genuine value that other people deem as good, then we should get as many eyes on it and reward them with something for their effort and time. Money is exactly that reward mechanism, so good things should be rewarded with a lot of money. I imagine you're already screaming at your screen about how this isn't how things play out in reality. Things that make a lot of money today are often derivative (see the huge market of remakes in cinema and games these past few years), appeal to the average user, and can be artificially inflated through bad actors manipulating markets and circumventing rules and regulations. Massive tech companies like Meta, Google, OpenAI, and Anthropic are not strangers to this either, and unfortunately the environment of the world today and the economic incentives in place do not incentivize or reward companies to act in the best interest of people.

There's a big problem with everything I discussed so far, and unsurprisingly it's a problem at the foundation of it all – what do we even consider to be good? Not only have modern technological advances exploited our reward structures, disregarding what is good for us, they haven't even presented a philosophy of what the good of society should even be. Some have gotten close to it. Sam Altman expressed that the future of society is one where intelligence is commodified and you pay for intelligence just as you do for other utilities like power and water. Elon Musk expressed that the good for humanity is building AI that seeks truth maximally and our species becoming interplanetary (how genuine his motives are, and whether he's the right person to be championing these pushes for society are a separate question entirely.) I'm sure every other CEO in Silicon Valley has their own thesis on what would be good for society, but the biggest thing is that no one fundamentally agrees on what artificial intelligence technologies like LLMs should be used for.

We could let the one LLM that rises above them all, the company to make the first Artificial General Intelligence, lead the charge on what these technologies should be used for and how their applications better the world. This sounds reasonable, but the problem is that becoming the one company that rises above them all isn't exactly a fair game that encourages acting for the greater good. That race heavily encourages acting hastily and manipulating psychological desires among other strategies. So... we've built a society where the ones at the top get to decide what is good for us, and we've designed the way to the top in such a way where they have to employ incredibly hostile and negative techniques. The only thing you can hope for in this kind of playing field is for an overly philanthropic and saintlike figure who just so happens to have the perfect amalgam of taste and intelligence... :-) ... Basically we're looking for the second coming of Jesus Christ Himself.

How do we escape the matrix?

For these technologies, as they exist today, it's up to the user to use them in responsible ways that are good for them and benefit them. Giving the user this responsibility is usually not a problem and is somewhat ideal. After all, you don't want to create technologies that people have a monopolistic control over and can decide what you can and cannot do with it. But, our society isn't at the point where we can trust everyone to have such powerful and capable tools, using them only for their benefit when it's so easy to use for their downfall. We've seen this before with the advent of cheap and calorically dense foods; it's never been easier and more encouraged to be lazy, lethargic, and unhealthy. In the same way, with the advent of LLMs it's never been easier and more encouraged to be lazy, uninspired, and slowly atrophy your brain's ability to critically think. If someone is incredibly self-disciplined, cautious, and wise, they can use these technologies for their benefit. But when the people in charge of these technologies keep making the worst aspects of these technologies so much easier, so much more prevalent, and make them bleed into every aspect of life regardless of whether it even provides anything useful, it becomes nearly impossible to get anything substantive out of these technologies.

I say all of that to come to you with no solution in mind. I don't know how we can fix tech, I don't know how we can make the Steam Machine cheaper, and I don't know how we can reverse the "slop-ification" of the internet. All I know is that these ideas, while not being novel to anyone, are still not having an impact on how these companies approach working on these extremely special and genuinely exciting new technologies. We're building the the next nuke, but instead of having some of the world's smartest thinkers influence how this gets applied to the world we have straight up losers who purposefully didn't take a single philosophy or humanity course in their entire academic career and pride themselves on their dogmatic adherence to "science" despite not understanding even the bare minimum on the philosophy behind the science they hold in higher regard than every other discipline.

And I say all of that as a warning to not push yourself away from all these new "AI technologies" coming out! Most of it is so horrible that it will make you want to blow up the world, but there is genuinely a treasure trove out there for you in terms of upgrading your capabilities in life. To be transparent, the ideas for this piece came from me asking Claude Opus 4.8 about neural networks. I then got into a "dialogue" with it, debating the technology and ways we can try to fully understand it. I don't take what it says as gospel, and I read all of it's AI-speak with a lot of scrutiny. But the great thing about Claude (and other LLMs) was it essentially being an interactive Wikipedia article, a mini professor that's kind of stupid sometimes that I could bounce ideas off of. It's still not nearly at a point where it feels the same as talking with someone who's an authority with years in the field, but it's an ok substitute for now. And hopefully one day it will be at the level of those with years in the field. All I can hope for is that these technologies evolve not to replace those scholars whose life blood and purpose is driven by the pursuit of exploring their field. I hope these technologies evolve to aid in that lifelong pursuit, to aid in furthering our understanding of the reality we as humans experience, and to aid in reducing the gap between the ideas our minds hold onto and our manifestation of those ideas into our reality.

A "dialogue" between human and hardware.

Here's the conversation I had with Claude Opus 4.8 prior to writing all of this. If you want the simulation of how I moved through this conversation, skim through the Claude replies and stop reading when it sounds too confusing.

Paul: What's the difference between deep learning and AGI/AI LLMs

Claude: These terms get mashed together constantly, and the phrasing bundles three different kinds of things that aren't on the same axis. AI is the broad umbrella — any system doing something we'd call intelligent (chess engine, spam filter, ChatGPT); it's a goal/field, not a technique. Deep learning is a technique — a subset of machine learning using many-layered neural networks to learn patterns from data; it's how you build a lot of modern AI. LLMs are a type of system built using deep learning — deep networks (usually transformers) trained on huge text corpora to predict and generate language. AGI is a capability threshold, not a technique or system — a hypothetical AI matching or exceeding human ability across essentially any cognitive task; it doesn't exist yet, and whether scaling LLMs gets there is an open question. Nesting: AI ⊃ machine learning ⊃ deep learning → used to build LLMs; AGI sits outside the stack as a target. Deep learning answers "how is it built," LLM answers "what kind of thing is it," AGI answers "how generally capable is it." Today's frontier models are LLMs built with deep learning that are not AGI.

Paul: So what is machine learning?

Claude: Machine learning is the layer between AI and deep learning — an approach where instead of a programmer writing explicit rules, the system learns patterns from data. Contrast: traditional programming hand-codes the rules ("if email contains 'free money,' flag it"), which breaks down fast; machine learning is shown thousands of labeled examples and figures out the patterns itself. The core shift is learn from data rather than be told what to do. Flavors: supervised (labeled examples → prediction/classification), unsupervised (find structure in unlabeled data), reinforcement (trial and error with rewards). Deep learning is one method within ML — the many-layered neural network one — especially good at messy high-dimensional data (images, audio, language) where hand-engineering features is hopeless; other methods (decision trees, linear regression, SVMs) still work for simpler problems. Full chain: AI ⊃ ML ⊃ deep learning → LLMs. The unifying idea: trade explicit instructions for examples, powerful exactly when the problem is too complex to write rules for.

Paul: What are neural networks and what makes them unique from any other structure of networks we've built in the past

Claude: A neural network is many simple units ("neurons") in layers; each takes numbers, multiplies by adjustable weights, sums, applies a nonlinear function, passes forward. Stack enough and it represents extraordinarily complex input→output relationships. What makes it unique: the structure is learned, not designed — unlike circuits, flowcharts, databases, or traditional software where humans assign every component its job, the units start blank and the weights are tuned automatically by data; roles emerge from training. Consequences: knowledge is distributed not located (no single cell holds "Paris is capital of France"; robust but opaque); the whole thing is differentiable, so trainable end-to-end via backprop + gradient descent (most designed systems have no notion of "adjust slightly to be more correct"); they learn their own representations (vs hand-engineered features); universality (can approximate essentially any input→output function). One-liner: traditional structures store designed logic in fixed components; neural networks store learned patterns in adjustable connections. Caveat: "neural" is loose brain-inspired branding; better to think of it as a giant tunable mathematical function than a digital brain.

Paul: Ok wait, tell me most fundamentally the big picture of neural networks, then go into what it's made up of, and then describe how each component works. Make it concise and fairly intuitive

Claude: Big picture: a neural network is one big tunable function — numbers in, through layers of simple operations, numbers out; by adjusting internal dials until outputs match known examples, it becomes whatever function the data demands. Made of: neurons (units holding/passing numbers), weights and biases (adjustable dials), activation functions (nonlinearity inside each neuron), organized into input / hidden / output layers. How each works: a neuron holds a single number (inputs hold raw data, outputs hold the answer). Weights are connection strengths — each neuron multiplies prior neurons by weights and sums; big weight = "this matters," near-zero = "ignore"; weights are where knowledge lives. A bias is an extra number added to the sum, shifting the firing threshold. An activation function bends the sum through a simple nonlinear curve; this nonlinearity is essential — without it stacked layers collapse into one straight-line equation. Layers chain this: early layers catch simple patterns (edges), later ones combine into complex ones (shapes, faces). The missing piece is training: start random, measure wrongness, nudge every weight toward less error, repeat millions of times.

Paul: So what happens in the hidden nodes? Is it stuff we can't analyze? Like could we trace an input neuron all the way to the output neuron, or is that impossible due to non linearity

Claude: Important correction: nonlinearity is not what makes tracing hard. Two separate questions. (1) Can you mechanically trace input→output? Yes, completely. A network is fully deterministic transparent math; you can follow a value through every multiply, sum, and activation to the output. Nonlinearity doesn't block this — an activation is just a known curve. With the weights you could trace every step by hand (just slow — billions of them). (2) Can you understand what a given hidden neuron "means"? Often no — and that's the real problem, caused by distribution and scale, not nonlinearity: a single neuron rarely maps to one clean concept; concepts are smeared across many neurons; neurons are often polysemantic (fire for several unrelated things). So you can trace the arithmetic perfectly but it doesn't translate into a human-legible story. This gap is what interpretability / mechanistic interpretability tries to close, with real progress (meaningful features from combinations of neurons, circuits, concept editing). "Nonlinearity keeps the math computable but curved; distribution keeps the meaning hidden."

Paul: Could we theoretically apply another deep learning method to supervise the original deep learning process to each individual neuron, tracing its input to its output, then figure out what each transformation means?

Claude: Yes — and it's an active frontier, not just theory. Two real versions: models explaining models (collect inputs that make a neuron fire hardest, ask a bigger model "what do these share?", test the hypothesis — done at scale to auto-label neurons), and sparse autoencoders (train a second network to re-express messy activations as a larger set of cleaner, mostly-one-concept features — literally deep learning supervising the original's representations, and it works). Obstacles: the interpreter regress (what explains B?); explanation ≠ causal guarantee (a correlated hypothesis isn't proof of what the neuron does); brutal scale (the meaning is in combinatorial interactions, not the neurons); faithfulness vs plausibility (a persuasive wrong story is indistinguishable from a correct one — worse than nothing because reassuring). Honest state: second models surface candidate meaning, which humans + causal tests confirm; not a push-button pipeline. Physics aside: you can have complete microscopic knowledge and still lack the macroscopic theory (every molecule's position ≠ thermodynamics); interpretability is partly the search for the right coarse-grained variables.

Paul: Why can't we just train the second interpretive model to break down what exactly each step is doing in the first model and then work to develop an understanding behind the fundamental of it, then we have a framework to understand all deep learning

Claude: The plan is sound until "just" and "all," which hide three assumptions. (1) That "what each step is doing" has a compact explanation at all — the honest description might be nearly as long as the network; if irreducibly complex, there's no framework to extract, because understanding means compression. Physics being compressible was a gift, not a guarantee; nothing promises trained networks are. (2) That decomposing steps composes back into understanding — meaning lives in interactions, which explode combinatorially; you can know every gear and not know what the clock does. Decomposition assumes modularity that distributed representations lack. (3) That understanding one network generalizes to all — each network learns its own idiosyncratic internal language; some features recur (universality) but much is model/architecture/seed-specific. Plus the interpreter is itself trained → subject to all three; unless verified by causal intervention you've just got a confident storyteller. So the real shape: "use interpreters to surface candidate structure, verify it causally, and hope it's compressible and universal enough to generalize" — an open empirical bet. The crux: is a trained network more like a physical system with deep reusable laws, or an accumulated pile of special-case hacks?

Paul: I believe there are fundamental laws that direct deep learning mechanisms, but maybe the language of linear algebra isn't the best way to interpret it. It's like we're looking at only the outputs of fundamental physics functions that are exponentially more complicated than we can simply state in equations. The benefit of physics is that we can create simplified models, that don't quite describe the reality of the world, and then apply those simplified models to figure out approximations of reality and then build more fundamental models of higher abstraction based off theory. We can't do that with deep learning as we're approaching it right now, since we're fundamentally working with the surface layer we see and not actually uncovering the deep underpinnings of what makes it work. The complexity we see is almost always downstream from a more simplistic thing, so saying that there's no way to fundamentally understand what happens with all the inputs in the chain of neurons is correct but also a sign that the real understanding will come from examining something else.

Claude: You're at the actual research frontier. Strongest evidence you underplayed: scaling laws already exist — loss falls as a clean power law across orders of magnitude, exactly the thermodynamics-from-molecules story; and the sparse-autoencoder move is already a confession that neurons are the wrong coordinates and features are closer. Friction on "complexity is downstream of simplicity": that may be survivorship bias — physics is the study of the compressible subset of nature (turbulence, protein folding, biology, economies got left hard). Nature didn't have to be compressible; a network has even less reason to be — optimized purely for loss, with zero selection pressure toward legible simplicity. Key distinction: you may get laws of the process (scaling laws, like the ideal gas law) without laws of the product (what this model knows) — a thermodynamics of learning with no atomic theory of the particular mind. Your "examine something else" becomes a bet once named; strongest nominee: the laws may be in the data/task (manifold hypothesis — the net is a cast of the data's structured low-dimensional surface), not the network. Logical flag: "we can't understand it at the neuron level" is consistent with a simple deeper layer but is not evidence for one. Crux sharpened: is the simplicity in the network or in the data?

Paul: I think there is simplicity in the network, but that's still using the wrong sort of language and pointing towards something different. The network that we see is the complicated surface, but dive down deeper into the fundamental and I think the framework of the network itself is probably built off very simple principles. It might not be describable in plain Math, but that's the same way how we can see the motion of the bodies and think it's incredibly complicated, but someone like Newton can develop a new framework within that math (calculus) and come up with laws that capture mostly what's going on (laws of motion, newton's laws, etc)

Claude: The Newton move is right, and you under-claimed: the simple principles are already in hand — a network is just (linear map → nonlinearity) repeated, with gradient descent grinding loss downhill; complexity is emergent from iterating a tiny rule (like the Mandelbrot set from z→z²+c). Plus convergence evidence: different architectures/data/modalities drift toward similar representations (universality), which the "laws" view predicts and "hacks" doesn't. But the crack is in what Newton did: not just a new language (calculus) but reducibility — compact laws that let you skip the computation(predict Mars in a thousand years without simulating every instant). A simple generator does not guarantee a reducible product: Conway's Game of Life has three-line rules and provably undecidable long-run behavior — no shortcut in anylanguage, because the irreducibility is a fact about the system, not a language failure. So "simple principles underneath" can be simultaneously true (gradient descent is simple) and useless (the grown object may have no compressed description). Real bet: not "is there simplicity underneath" (yes, the update rule) but "is the product reducible." Partial evidence for yes: scaling laws are a shortcut (predict loss without training) — but aggregate, not content. Two pushes: (1) Newton stayed inside formalism — if you reach for something not expressible in any formal system, the precedent doesn't support you. (2) You keep relocating simplicity one level deeper each round; that risks becoming unfalsifiable — what would the world look like if you were wrong?

Paul: Ok, how is deep learning used for LLMs, and then how can it be used for something different from LLMs

Claude: Deep learning is the general engine; an LLM is one configuration — change the data, the input/output shape, and the objective and the same machinery does something else. LLM: a transformer trained on one task — predict the next chunk of text; tokens become embedding vectors (meaning as position in high-dim space); attention lets each token look back over others and weight what matters for prediction, tracking long-range context. "Predict the next token" done well enough forces learning grammar, facts, reasoning, style as a side effect. Other uses, same recipe: vision (CNNs/ViTs — hierarchical visual features, medical imaging, self-driving); image generation (diffusion — flip the objective to denoise pure noise into an image); audio (speech-to-text, synthesis, music); reinforcement learning (no fixed answers — learn from a reward; AlphaGo, robot locomotion); science (AlphaFold: sequence→protein shape; weather, drug discovery, materials). Unifying point: same core recipe (stacked linear map → nonlinearity, weights tuned by gradient descent to minimize a loss); what changes is data, representation, and objective. An LLM isn't a different technology — it's deep learning pointed at language.

Paul: Ok, going back to the argument of the trained network being reducible or irreducible. Do you think propositional logic is a reduced version of how our brain's work to form logical thoughts and arguments? If so, then that shows our brain, which is a very complicated biological computer, has some emergent properties, properties of which can be more of a description/capturing of the form of what the thing is itself.

Claude: This is your strongest move and the one most likely to fool you. Strong version (yours): a brain looks irreducible at the mechanism level yet hosts a clean formal competence (logic) — so irreducible substrate hosting reducible structure is demonstrated, and ANNs are the same kind of object. The crack: logic is normative (how the brain ought to reason), not mechanistic (how it does) — proven by the fact that humans are bad at logic (Wason, conjunction fallacy, affirming the consequent). Molecules never violate thermodynamics; brains violate logic constantly → logic isn't capturing the machine, it's a target the machine approximates and misses. More like chess rules vs the chess computer's transistors. Where does logic live? Not in the brain — Aristotle ran zero neuroscience; it was found by examining the structure of valid inference itself. Logic is a law of the task (reasoning), not the organ (brain); calculators and Prolog implement it too. This cuts against "simplicity in the network": the reducible thing was extracted from the domain, not the substrate → for a network, the clean thing would be found by studying what it computes over (task/data), not its weights. Synthesis: reducibility at the level of competence, irreducibility at the level of mechanism, simultaneously. Honest flag: three rounds now of relocating simplicity rather than naming a disconfirmer.

Paul: Ok, so I see what you mean. That does point more towards the underlying simplicity being in the data itself and not in the neural network deep learning machine itself. Since the data exists in the domain of inputs, then within that domain we can find simpler fundamental "truths" about how it operates. The neural network, however, is extremely messy... The closest we can get is by getting another neural network to track each input neuron and all the transformations... but that still gets us in a loop where we're using something not understood to try to understand something not understood. It's like trying to solve a variable in an equation by using another equation... [On logic:] logic is this structure that exists, whether outside the brain or inside the brain isn't really what I'm getting at... logic seems to be something that exists within this realm we live in. Our brain doesn't adhere to it, but our brain is a type of biological computer that has the capability of recognizing this universal concept... mathematics, physics, communication, data theory, these are all universal concepts that our biological computers (brains) can understand... we can clearly pull out some universal concepts or truths that apply to all aspects of reality (cause and effect, the arrow of time, quantum field theory)... at a deeper level I don't think we can fully understand our brain because to fully understand a system we have to be able to step outside of it... We can never step outside of our brains, so if we end up making another form of intelligence... then we can start to point the tools back at ourselves... And even then, there's the question of whether or not what comes back out is even comprehensible to us. Since deep learning systems live outside of us, then I believe there is some way we can get to understanding it at a more basic and fundamental level... it will at least show us the underpinnings of what makes it all turn. And that is probably the most valuable thing... since even with real life... it's the mastery of the fundamental laws and governing principles that gives incredible power and applicability, not the observable things themselves.

Claude: Most of this is right. Conceded: the reducibility-in-the-domain conclusion is correct and cleanly reached; the "equation within an equation" image independently reinvents the grounding problem (symbols defined by symbols never bottom out without a fixed point — causal intervention); the "governing principles beat observables" hierarchy of value is correct. The move I pressed: you treat logic/math/causality/arrow-of-time as features of the territory the brain recognizes(platonism). But each has a known failure domain — causality dissolves in time-symmetric microphysics; Euclidean geometry cracked at curved spacetime; classical bivalence vs QM. They're spectacularly tuned to the middle-world regimea human-sized organism evolved in, and crack outside it — which looks like a feature of the observer's interface, not the territory. This detonates the inside/outside symmetry: you can step outside your brain but not your concepts, and the concepts (logic, math, "mechanism," "underpinning") are brain-shaped instruments — so pointing tools at an "external" network smuggles the inside-view back in through the toolkit. Subsumes your introspection worry: the limit isn't the skull, it's the representational scheme; you'd find only the structure your concepts can represent and never know what you missed. What rescues the productive half: networks (unlike brains) allow controlled interventional experiments (ablate, watch behavior) — real grounding, breaks the regress — but that's about interventability, not externality or universal truths. Likely landing: laws of the process are gettable, with a residue the formalism predicts but no human pictures (the QM deal). Final fork: are the universals in reality or in the interface (such that a different intelligence pulls a non-overlapping, equally-valid, mutually-incomprehensible set)?

Paul: In response to the "every one of your 'universal' concepts has a known failure domain", this is true... But I don't think that it contradicts or invalidates what I said. This is the process of learning and evolution, we figure out that things we thought were the most fundamental (plum pudding model, classical physics) ... and through further explanation we discover that there is something more fundamental. The key thing here is that we kept discovering more fundamental simple truths and principles... at the end of the day we get somewhere where there are still basic and simple principles waiting at the end. That's what I'm pointing to, those basic principles waiting at the end of each tunnel... at the core of everything there's some universal guiding principles/laws that everything is confined to exist within. So... that doesn't negate the fact of fundamental simple concepts being at the heart of what we see to be incredibly complex. You can argue that it's just a consequence of our brains... wanting to reduce things... and you can argue that reality is just unordered chaos and there is no such thing as principles or laws, but to take that position would be fundamentally inhuman... since we are from biology then this emergent property of us to want to compress things and find fundamental laws... is valuable. It might not be the objective truth... but it's the way our minds work... and it's clearly not fruitless because it has led to incredible breakthroughs... This pursuit of trying to find the fundamental laws behind something could be misguided when applying it to neural networks and large language models, but I think it's ultimately worth it and something that will lead to a thing that is most meaningful in the lives of humans.

Claude: Stopped looking for the crack — on this move there mostly isn't one. You dissolved the fork rather than answering it: it doesn't matter whether laws are in the territory or the interface, because the value is in the practice — the compression drive is constitutive of human cognition, has an unmatched track record, and that warrants the pursuit regardless of metaphysics. That out-pragmatists the platonism/anti-realism split, correctly. And your central observation survives: every time we've dug we've found more structure, not less — the "failure domains" I cited aren't chaos leaking in, they're the seams where the next layer of law was found (Newton→relativity/QM). Reducibility has never bottomed out into genuine mush. One thing still on the table: the induction has a survivorship structure — we only have the track record where it worked; the genuine dead-ends (turbulence after a century; protein folding, which yielded to a neural network, not a Newton — darkly funny here) got relabeled "not-yet-physics." So the bet has a hidden term: that deep learning is in the law-bearing class, not the turbulence class — untouched by the track record. But you pre-empted this: it might be misguided for nets and still worth it — a correct expected-value argument (a search with a shot at fundamental law beats certainty about surfaces), the same warrant that justified every speculative physics program that didn't know its domain was law-bearing in advance. Synthesis we built: simplicity in the task/data not the machine; network irreducible at mechanism, reducible at what it computes; laws expressible only as far as the task is commensurable with human concepts (historically astonishingly deep, with a predict-but-don't-picture residue); warrant = track record + stakes, not a guarantee. Live frontier question: if deep learning's laws turn out QM-like (predictively perfect, conceptually un-pictureable), is that the simplicity you wanted, or a law you can use but never understand in the gut?

Paul: I think if deep learnings fundamental laws turn out to be like quantum mechanics, I would count that as a win in some extent since we found something that guides it. But the understandability is ultimately what I'm at, and in the same way that seemingly impossible to understand things were able to be expressed with analogies and metaphors that made it make more sense and become something that we could actually use as a framework for thought and an update to our model of the universe, I think these "unpictureable things", even like quantum mechanics, will eventually become understood. The pursuit of science and human knowledge should be to continually push to understand that which seems to be un-understandble. Quantum mechanics, in my view, will be understood completely beyond just the fundamental mathematical laws, and in the same way I think the mechanisms behind deep learning and neural networks can be understood too.

Claude: Tested the one empirical claim holding it up: the un-pictureable always eventually becomes pictureable; comprehension lags formalism but always catches up. You treat the gap as temporal (not-yet), not permanent. Where it's won: negative numbers, imaginary numbers (rotations), curved spacetime (rubber sheet), fields, entropy, natural selection — all crossed from formalism into gut. But the deciding case is the one you chose — QM: a century of the best minds trying to make it pictureable, and the verdict is interpretations multiplied rather than converged (Copenhagen, many-worlds, pilot-wave, relational, QBism), all empirically identical, none winning. Converging intuition is the signature of comprehension arriving; proliferating incompatible pictures that all fit the math is the signature of picturability being permanently absent. You picked the example where your induction has been tested longest and is failing. Decisive distinction: understanding-as-handling (predictive, manipulable mastery) vs understanding-as-seeing (gut picturing). Your winning examples crossed both; QM crossed the first and shows hundred-year evidence of never crossing the second — and they can come apart permanently. Mapped to deep learning: you'll get full handling (scaling laws, interventional circuits); the open question is whether seeing arrives on top — and if QM is the precedent, it never does (every "what it's really doing" story a Copenhagen-vs-many-worlds situation). Strongest form of your reply, conceded: maybe seeing itself evolves — humans grow new primitives until the formalism becomes native (a quantum-native child sees what Feynman could only compute). Can't refute it. But the price: "understanding" stops being humans comprehending reality and becomes humans rebuilding themselves into the kind of thing that finds it native — the understander changes to fit the understood, and the continuity of the "we" is what gets traded away. Final fork: when you say QM will be understood completely — the present human mind finally sees it (losing for a century), or some descendant configuration of mind finds it native (almost certainly true, but then the intelligence we build to understand ourselves would be the thing that understands, and we'd face whether its understanding is still ours)?

Paul: Ok, let's call it there. [+ request to summarize and log this session]

Claude: [Confirmed capability, checked vault conventions, and wrote this session note.]

Where I've been so far...

It is June 13, 2026. By the end of this year, I want to be at a point where I have everything in place to make money. The motto for this year is "money-maxx". I am trying to maximize my money making capabilities, by any means necessary. This includes anything from profitable ideas, to acting on those profitable ideas, to distributing the results of executing on those ideas to others who see value in them.

The point of making as much money as possible is because I want to have a stable living environment, and one where I can flexibly live between The Bahamas and Canada.

The world is your oyster

In a world with increasingly better technologies, primarily artificial intelligence, the barrier to executing and manifesting ideas has never been lower. I am a incredibly capable and smart person, but my discipline and drive for doing things is generally poor. In 2026, I have been working on increasing my discipline and drive through doing things I don't feel like doing. The things that help me execute on ideas is losing a care for:

1. what people think about what I'm doing

2. fear of failure and how that would validate others lack of belief in me and invalidate my belief in myself as a smart and capable person.

These are genuine concerns, but a few things I've learned through various figures' advice online (TOGI, Jon Zherka, Elisha Long, esoteric gnosis things, I know this sounds crazy but these figures are who I'd consider the modern "Holy Fools") and through my own musings are the following:

1. Caring about what people think is unproductive because most people are completely lost themselves. Their thoughts about you are just them projecting their fears of trying onto you.

2. This is everyone's first time living life, so any advice you get from them are things that worked for them and only them. Trying to prescribe anything without that caveat is stupid (and I’m not using that word lightly). Everyone is wired differently. Pretending that what works for a statistical average or a single individual (with much different genetics and lived experiences from you) will be the thing that works for you too 100% is such a fatal mistake; you might as well restart your life right now into a different incarnation.

3. Overly introspecting all the time and overthinking things is a malfunction of the mind. Actions should be taken above all other things, and if you have faith in yourself and in a higher power (God) that commands everything including your path in the universe, then you should be able to get out of any hole that you are in.

4. There is no such thing as good or bad. There just is, and that's it. You make it good or bad in your own head, therefore anything that happens can be made good so long as you decide it (and put in the required work) to be so. Additionally, things that most people consider to be generally "bad" can always open up blessings for you that you cannot see because you are blinded by negativity and doubt. (I am reminded of the guy in all pink that I saw on Instagram Reels a few months ago, who said whenever something bad happens to him he always prays to God and asks "where's the blessing in this." He looked incredibly silly and like someone to never trust in a million rebirths and lifetimes, but this attitude seems to be the objectively correct one to live with.)

5. If you ever think to not act on something because "someone out there already had the idea, so if it was a good idea it would've been done by now", you’re falling trap to the greatest fallacy of all time. Not only are you forgetting that 99.9999% of people give up before they've even gotten 20% into any project (except if it's been delegated by a teacher or a boss), you're also completely ridding yourself of the divine purpose you've been given in this life. You are you, and there has never been another you. Sure, you might have similar ideas to someone else in this world, but your perspective on it and your way of executing that idea can never be replicated. So, write out that idea and create a roadmap to act on it, and manifest that idea into reality. Only you can do it.

With that said, I will use my ideation process and the tools at my disposal to create genuinely great things that benefit the world. These great things will not only be for the sake of bettering the world, my country, and the people in it, but also for bringing me income so I can better support myself and the ones I love, and give us the lives we can and deserve to live in this one special moment of life.

To be a "Million Dollar Bahamian" (Top Secret Info)

Right now, I've settled on the idea of [REDACTED].

Early on, [REDACTED].

I genuinely want to help out my country and the people in it. As the years pass by, I realize with each revisit back home that I deeply care about this place and the people in it. Even though there's a lot of bad, there's an indisputable amount of good, a good so pure that it outshines even the worst sins born from this country. So, as a Bahamian myself, and since the US and Canada don't seem to want a UC Berkeley physicist anytime soon, I will bring my skills back home to better it (and hopefully make a lot of money in the process).

This idea is in the v1 demo stage (completed June 9, 2026), and I am hoping to get this out there by the end of this month (June), at the very latest mid-July.

The spirit needs care too

Since income is a top priority, I want to minimize anything unrelated to making money in the short term. But, as we all know, money is not the only thing on this earth, and orienting your life only towards saving/making money is a big mistake. So, for my general life enrichment, and what I'd be doing if money was no worry, I am still doing the following things on the side:

• Videos

  • swagtower (YouTube)

    • This is a gaming channel I am working on, aiming at 1 video and 3-5 shorts a week. The point of this channel is to make something fun and slightly creative from my love for video games as a whole. I've been making YouTube videos since I was in grade 3, and it all started with Minecraft let's plays. Continually, I find myself recording fun gaming videos in my life, with or without a channel, so this is a good outlet for that that I'd be very happy with if it became something I did with 65% of my time.

  • [1a] (YouTube)

    • This is a variety channel I am working on with my two other friends, Jeshuran and Michael. 

• Music

  • gilopaolo (everything)

    • This is my life soul. Art has always been something I've wanted to pursue since I was in grade 1. My first ever aspiration for a career (aside from wanting to be the dudes at the gas station that clean your windshields because it looked so satisfying) was a cartoonist, and I'd draw all the time in free notebooks my mom would get from her work. In high school, that creativity turned into music as I discovered how to use GarageBand on my 2015 15-inch MacBook Pro. Combined with me lying about being able to play the guitar for my grade 6 graduation (partially because my dad was really good at guitar and also because my friend could actually play guitar and I'd be able to skip class to chill with him), I picked up the instrumentation skills and proper equipment to be able to fully compose, record, and publish music online.

    • Right now, I am working on finishing my 11th album (my 3rd official release in spirit), "voicings (the past four years of my life)". It is 80% finished, I just need to work on the second to last track and re-do the mix on the first track. Then, I can put that behind me and move onto new things, whatever that might be.

    • Eventually, I want to do full on creative videos, whether that be music videos or just short films. My long term goal is to direct and film a movie, since I see that as the ultimate medium for executing what I feel is the essence of art – capturing a feeling in the most fundamental sense and communicating that to someone else, whether through words, sound, visuals, or a combination of all 3, so that they can see how the world is filtered through my experience.

promise me this, that no matter what, one day before the end of this year, i will go somewhere and create something meaningful. something substantial on a spiritual level, that i can look back on in a few years and have memories flooded back to me, something i can show someone and say "this is me. this is what i think, what i am, this is what i do." and they’d look at it, look at me, and understand a bit more about the world i see.

because what i’m doing now, if you were to feel it, you’d think that i’d be better off doing nothing.

just a little bit longer…