There are phenomena we know _occur_ yet lack understanding of _how_ they happen. While categorizing these knowledge edges by field seems limiting, it's useful for organization purposes. In [[AI]] - [[black box problem]] — Refers to the difficulty in understanding how complex deep learning models make decisions. We can see what goes on in the training data, the outputs, and the mathematical structure of the network — but we still struggle to understand the decision making process inside the "black box" - Neural network generalization — We know that neural networks can generalize beyond their training data, but we're not sure why they do this so well as opposed to just memorizing training examples - The "[[lottery ticket hypothesis]]" — Within large neural networks, there exist smaller subnetworks that can achieve similar performance when trained in isolation, but we don't fully understand why these "winning tickets" emerge - The emergence of capabilities in LLMs at scale - P vs NP problem — We can identify NP problems ("Non-deterministic Polynomial time" problems where solutions can be verified quickly in polynomial time but we don't know efficient algorithms to find those solution), but we don't know if they can be solved efficiently. Basically — the P vs NP problem asks whether every problem with quickly verifiable solutions also has quickly findable solutions. - [[Traveling salesman problem]] and [[Protein folding]] are examples - [[Algorithmic generalization]] — Why some models avoid overfitting while having capacity to memorize data. Basically, neural networks *could have* billions of parameters — they could memorize the whole training set if they wanted — yet mysteriously, they often learn to generalize well to unseen data. - Researchers debate why this happens — theories include: Implicit regularization from gradient descent, the geometry of high-dimensional space, or properties of natural data distributions. - Benign overfitting — Some neural networks achieve zero training error (perfectly fitting even random noise in the training data) while still performing well on new data. Traditional statistics would predict that fitting noise should harm generalization, but it doesn't. - Foundation model reasoning thresholds — LLMs show nearly no ability to solve multi-step reasoning problems below certain parameter counts, then suddenly demonstrate these abilities at scale. In Physics - [[Quantum mechanics]] — While we can use quantum mechanics with extreme precision (the magnetic movement of an electron has been predicted by quantum electrodynamics and measured with agreement to 14 decimal places) the interpretation of what's actually happening ([[Copenhagen theory]], [[Many-Worlds theory]], etc.) remains the debated - [[quantum measurement problem]] — The unexplained collapse of quantum wave functions upon measurement, e.g., Schrödinger's cat - [[Matter-anitmatter asymmetry]] — Why our universe contains mostly matter despite equal creation - [[Proton radius puzzle]] — Different measurements yield contradicting proton size results - [[Magnetic monopoles]] — Theory predicts they should exist but we've never found one - [[dark matter]] and [[dark energy]] — We can observe the existence of dark matter and energy through its observed gravitational effects but we don't know what it is, despite it making up ~90% of the universe's mass. - [[High-temperature superconductivity]] — We can create materials that superconduct ([[superconductivity]] — materials lose all electrical resistance at certain temperatures, allowing perfect conductivity and strong magnetic fields — enables MRI to maglev trains — would mean lossless power transmission if we could do [[room temperature superconductivity]]) at relatively high temperatures, but the exact mechanism isn't fully understood In Neuroscience - [[Neural geometry]] — We're beginning to see higher-dimensional structures and patterns in neural connectivity — we have a [[Frontier Films]] coming out about this soon. - [[consciousness]] — Biology can't explain how our nearly 100B neuron brain generates consciousness — many would argue the relationship between the brain and consciousness is still up for debate. - [[Memory storage]] — How precisely memories are encoded and retrieved - [[Placebo effect]] — Why belief alone can trigger measurable physiological changes, [Benedtti et al. (2011) "Neurobiological Mechanisms of the Placebo Effect"](https://pmc.ncbi.nlm.nih.gov/articles/PMC6725834/) - [[Dreams]] — Why we dream and what purpose they serve? In Chemistry - [[Sabatier reaction]] — catalytic reduction of CO2 with H2 on a metal catalyst (like nickel) that creates methane and water — we know it works, we're not sure exactly why (example of [[chemical intuition]] — many reactions work in practice before we can explain why) - Water's weird properties — Water expands when it freezes (most substances contract), reaches maximum density at 4C (not at freezing), has unusually high surface tension, and extraordinary heat capacity. - [[Metal-organic frameworks]] — MOF-177 can absorb 140% of its weight in carbon dioxide despite being highly porous (seemingly contradictory). The MOF-5 family shows selective gas absorption that's difficult to predict from structure alone. These frameworks combine metal ions with organic linkers creating structures with tunable pore sizes. In Biology - [[Abiogenesis]] — how non-living chemistry transitioned to living systems - [[Aging mechanisms]] — What exactly drives cellular senescence and organism aging - [[Epigenetic inheritance]] — How some environmental adaptations pass to offspring and others don't - [[Protein folding]] — while we've made huge progress with things like AlphaFold, we still don't fully understand the principles guiding how proteins fold so quickly and reliably. - general [[anesthesia]] — we can reliably render people unconscious using anesthetics like propofol or isoflurane but the precise mechanism is poorly understood - [[sleep function]] — why is sleep a critical function to many living animals?