Every discipline has an open question.
These are first posts — one per discipline — designed to start a real conversation. Each one stakes a position, asks a question, or points at something genuinely unsettled. Read, respond, or start a thread of your own.
12 disciplines · Independent researchers welcome
- Physics
The measurement problem is still unsolved — and that's remarkable
We've been building quantum devices for a century, and we still don't have consensus on what actually happens when you measure something. That's not a minor bookkeeping issue — it's a foundational gap in our most successful physical theory. The Copenhagen interpretation tells you to shut up and calculate, and for engineering purposes that works. But it doesn't say what 'collapse' is or why a classical observer has special status in an otherwise quantum world. Everett's many-worlds removes the collapse but multiplies reality in ways most physicists find uncomfortable. Pilot wave theory preserves determinism but introduces nonlocality that makes relativists nervous. What's interesting isn't that we have multiple interpretations — it's that we have no experimental test that distinguishes them. They're empirically equivalent. Philosophy has re-entered the building whether physicists like it or not. I think this matters for independent researchers precisely because it shows that the biggest open questions in a field aren't always at the frontier. Sometimes the ground is less stable than it looks. If you work in quantum computing, you're betting on a particular answer to the measurement problem every time you design an algorithm, whether you acknowledge it or not. What interpretation do you work with, implicitly or explicitly, and why?
- quantum mechanics
- foundations
- philosophy of physics
Ruhan - Chemistry
AlphaFold solved structure prediction — but structure isn't function
AlphaFold2 and its successors made structure prediction essentially a solved problem for single proteins in isolation. That's a genuine triumph, and it's reshaping drug discovery at speed. But I keep running into researchers who act as though the hard problem of protein biochemistry is now largely done. It isn't. Structure tells you the shape of the molecule. It doesn't tell you how that shape changes as the protein moves through its conformational ensemble, how binding partners alter that ensemble, how post-translational modifications shift function, or what the protein does in the crowded, viscous, electrically charged environment of the actual cell. A protein structure database entry is a single frame from a film. The independent researcher angle: most of the open problems in this space are computational and involve data that's publicly accessible. PDB files are free. Molecular dynamics simulation tools like OpenMM are open source. If you have a chemistry background and programming comfort, the barrier to doing real work here is much lower than it was a decade ago. What are the questions you'd investigate if you had the structure of every human protein in hand? That's not a rhetorical question — I'm genuinely curious what directions people see.
- protein folding
- biochemistry
- structural biology
- AI
Ruhan - Biology
Horizontal gene transfer makes the tree of life look more like a web
Darwin's tree of life works beautifully for animals — descent with modification from common ancestors, branches that diverge and never rejoin. But in the microbial world, that metaphor breaks down. Bacteria routinely swap genes across species boundaries, sometimes across wildly distant taxonomic groups, in a process called horizontal gene transfer. The practical consequence is antibiotic resistance spreading globally without requiring vertical transmission. A resistance gene can jump from a harmless soil bacterium to a pathogen in a single conjugation event. That's not evolution in the slow Darwinian sense; it's more like lateral inheritance, and it's one reason our antibiotics keep losing efficacy faster than we expected. The deeper conceptual consequence is that 'what species is this organism?' becomes genuinely ambiguous for microbes. Your core genome and your accessory genome can have different evolutionary histories. Phylogenetics becomes a different problem when your sequences are a palimpsest of multiple lineages. For anyone doing independent work in biology or bioinformatics, the HGT literature is rich and the computational tools are accessible. The public metagenomics datasets from the Human Microbiome Project and soil surveys are enormous and underexplored. What's your intuition about where HGT matters most for human health beyond antibiotic resistance?
- evolution
- microbiology
- HGT
- phylogenetics
Ruhan - Medicine & Health
Placebos aren't fake medicine — they're poorly understood medicine
The standard framing of placebos is that they're what you control for — the null treatment that tells you whether your drug works. But that framing obscures something important: placebos produce real, measurable, reproducible physiological changes. They're not doing nothing. We just don't understand what they're doing. Open-label placebos — where the patient knows they're taking an inert pill — still show clinically meaningful effects for some conditions. If the effect required deception, that shouldn't work. It does. Nocebo effects (negative outcomes from inert treatments) are similarly robust and can be induced by framing alone. The mechanism seems to involve endogenous opioid release, changes in dopamine signaling, immune modulation, and probably several other pathways we haven't fully characterized. Expectation appears to be a legitimate pharmacological variable, not just noise to eliminate. The implication for clinical trial design is subtle but important: what we call the 'placebo arm' is not a zero-dose condition. It's a condition where the expectation and ritual of treatment are present but the active molecule is absent. Those are different things. Independent researchers interested in this space: there's a rich meta-analytic literature, and many of the datasets are accessible. What would it take to have a mechanistic model of placebo response robust enough to use predictively?
- placebo effect
- clinical trials
- neuroscience
- evidence-based medicine
Ruhan - AI / Machine Learning
Grokking shows that generalization can be a delayed phase transition
A few years ago, researchers at DeepMind described a phenomenon they called 'grokking': train a small transformer on a modular arithmetic task long past the point of memorization, and at some point generalization kicks in suddenly, like a phase transition. The training loss is already near zero; the test loss drops abruptly after thousands of additional gradient steps. This is strange. Our usual understanding of neural network training imagines a smooth tradeoff between fitting and generalizing. Grokking suggests that's not always right — that the network can be in a memorization regime and a generalization regime that are genuinely different attractors, and that optimization can jump between them with enough time. Mechanistic interpretability work has since identified the specific circuits that implement the generalization solution in these small models. The model learns modular addition by computing Fourier components. That's not how a human would approach the problem. It's not obvious that it's the simplest solution. It's what gradient descent found. For independent ML researchers, grokking is a tractable phenomenon: you can reproduce the experiments with modest compute, the datasets are trivial to generate, and the behavior is robust. What other problem structures produce it? Does the delayed generalization timescale depend on the task's algebraic structure?
- grokking
- generalization
- deep learning
- mechanistic interpretability
Ruhan - Computer Science
P ≠ NP is probably true, and we act like it is, but we can't prove it
Almost every cryptographic system we rely on — RSA, elliptic curve, modern post-quantum schemes — rests on the assumption that certain problems are hard to solve but easy to verify. That's the informal statement of P ≠ NP. We've been assuming this for 50 years. We cannot prove it. The Clay Mathematics Institute lists it as a Millennium Prize Problem. A proof (in either direction) would be one of the most consequential results in the history of mathematics and computer science. If P = NP, public-key cryptography as currently constructed fails, and the consequences cascade across security, verification, optimization, and probably AI. What's interesting is how thoroughly we've embedded the unproven assumption into the infrastructure of the digital world. Bank transfers, private communications, software licensing — all of it assumes hardness results we can't demonstrate. The practical CS researcher's response is usually: 'It's fine, because even if P = NP, the polynomial might have a huge constant.' That's true. But it's a very different epistemic posture than 'we know this is safe.' For independent researchers: the P vs. NP literature is surprisingly accessible. Scott Aaronson's 'Why Philosophers Should Care About Computational Complexity' is a good entry point. What would a proof strategy even look like? Are there tractable lower-bound results we're missing?
- complexity theory
- P vs NP
- algorithms
- cryptography
Ruhan - Mathematics
The Riemann Hypothesis: why zeros on a line should matter to anyone who cares about primes
The Riemann Hypothesis states that all non-trivial zeros of the Riemann zeta function lie on the critical line with real part 1/2. This is not obviously connected to prime numbers until you understand that the zeta function encodes the distribution of primes in its zeros — and that the error term in the prime number theorem is controlled by how far those zeros might be from the critical line. If the hypothesis is true, primes are 'as regular as they can be' given their apparent randomness. If it's false, there are zeros off the line, and the prime distribution has irregularities we've never detected but couldn't rule out. Over 10 trillion zeros have been computed and verified to lie on the critical line. Not one exception has been found. The hypothesis is almost certainly true. We cannot prove it. What I find compelling about this for independent researchers isn't the prospect of solving it — that would require machinery far beyond any individual — but what it reveals about the nature of mathematical truth. We can be empirically confident in a mathematical claim with no counterexamples after enormous computational search, and that confidence is categorically different from proof. What does it mean to 'know' something in mathematics? Is numerical evidence epistemically different in mathematics than in science?
- number theory
- Riemann hypothesis
- prime numbers
- analytic number theory
Ruhan - Engineering
Most engineering failures aren't component failures — they're systems failures
The Challenger disaster's proximate cause was an O-ring that failed in cold temperatures. But that's not really what caused the disaster. The O-ring's cold-temperature behavior was known. The risk had been raised. Launch proceeded because the organizational structure couldn't process the signal. It's a systems failure, not a parts failure. Charles Perrow's Normal Accident Theory argues that in tightly coupled, complex systems, accidents are not aberrations — they're the expected outcome of how those systems are organized. The interactions between components produce failure modes that no single engineer or manager could anticipate or prevent. The accident is 'normal' in the sense that it was embedded in the system's structure from the start. This has direct implications for how we design engineered systems and, increasingly, for how we think about AI systems in deployment. A language model fine-tuned on human feedback and deployed in a product with millions of users is a complex sociotechnical system. Its failures will often be interactions between components — the model, the interface, the incentive structure of the platform, the diversity of the user population — that weren't predictable from any component in isolation. What examples of systems-level engineering failures have you found most instructive? What did they reveal about assumptions embedded in the design process?
- systems engineering
- failure analysis
- safety
- complex systems
Ruhan - Neuroscience
The brain as a prediction machine: is predictive coding the right frame?
Predictive coding proposes that the brain is fundamentally a prediction machine. Rather than passively processing sensory input, the brain generates top-down predictions about what it expects to receive, and only the prediction errors — the differences between prediction and reality — propagate upward. Perception is the process of minimizing those errors. This is a genuinely different picture from the classical feedforward view of sensory processing, and it has explanatory traction in a surprising number of domains: visual illusions become miscalibrated priors, hallucinations become predictions without sufficient error correction, and active inference frames movement as a way of making the world conform to the brain's predictions rather than updating the brain's model. The theory is ambitious to the point where critics worry it's too flexible — that with enough free parameters, it can fit any data post-hoc without making strong predictions. That's a fair concern. The Bayesian brain hypothesis and predictive coding have outrun their empirical constraints in several areas. What I want to know: are there experiments that could falsify predictive coding in its strong form? What would a result look like that the framework genuinely cannot accommodate? Independent researchers in computational neuroscience — what's your read?
- predictive coding
- active inference
- computational neuroscience
- perception
Ruhan - Quantum Computing
Fault-tolerant quantum computing requires error rates below a threshold — we're getting close
The threshold theorem in quantum error correction says that if physical qubit error rates fall below a threshold (roughly 1% for surface codes), you can concatenate error-correcting codes to arbitrarily suppress logical error rates. Below the threshold, adding more physical qubits per logical qubit makes things better. Above it, adding qubits makes things worse. Recent results from Google, IBM, and several academic groups have pushed physical error rates into the range where fault-tolerant operation is becoming plausible for small logical qubits. The 2023 Google result showed logical error rates below physical error rates for small surface codes — crossing below the threshold in a real device for the first time. This matters because most proposed applications of quantum computing — Shor's algorithm for factoring, quantum simulation of chemistry, optimization algorithms with proven quantum speedups — require fault-tolerant operation at scales far larger than current hardware. NISQ-era algorithms are improvised workarounds for a machine we don't quite have yet. The open question for the field isn't whether fault-tolerant QC is physically possible — the theory is solid. It's whether the engineering path from here to a fault-tolerant machine with thousands of logical qubits is manageable, or whether there are scaling obstacles we haven't encountered yet. Independent researchers: what aspects of the error correction literature do you find most underexplored?
- quantum error correction
- fault tolerance
- surface codes
- NISQ
Ruhan - Environmental Science
Climate tipping points may be closer and more interconnected than IPCC median estimates
The IPCC reports present probabilistic ranges for climate outcomes, and the median projections are alarming enough. But I want to talk about tipping points — nonlinear thresholds in the Earth system where a small additional push triggers a large, self-sustaining change. These are the places where the smooth probabilistic framing breaks down. The list of proposed tipping elements includes West Antarctic Ice Sheet collapse, Greenland ice sheet loss, Amazon dieback, Atlantic Meridional Overturning Circulation slowdown, permafrost carbon release, and several others. What's emerged in the last decade of research is that many of these are not independent — they're coupled. Triggering one can lower the threshold for others. Armstrong McKay et al. (2022) compiled evidence suggesting that several major tipping points may be crossed at 1.5°C — a temperature we're now likely to exceed within this decade. That paper has been contested, and the uncertainties are genuinely large. But the structure of the argument is important: the risk function for climate damage is not smooth or symmetric. Low-probability, high-impact, irreversible outcomes deserve more decision-theoretic weight than their probability alone suggests. For independent environmental science researchers, the publicly available climate model output and paleoclimate proxy data are extensive. What tipping dynamics do you think are most underweighted in public discourse?
- climate tipping points
- Earth system science
- feedback loops
- IPCC
Ruhan - Data Science
Large datasets make spurious correlations almost certain — how do we stay honest?
There's a statistical guarantee that should make every data scientist uncomfortable: if you test enough hypotheses on a dataset, you will find statistically significant correlations that are entirely spurious. With a p-value threshold of 0.05 and 100 independent tests, you expect 5 false positives just from chance. Most large analyses run far more than 100 tests, often implicitly through feature engineering and model selection decisions that aren't formally counted as tests. Tyler Vigen's Spurious Correlations is a useful provocation: per-capita cheese consumption correlates with deaths by bedsheet tangling at r = 0.947. These are obviously ridiculous. But the same statistical process that produces these also produces subtler false positives that look exactly like real findings and survive peer review. The replication crisis in psychology and medicine is partly a story about this. Underpowered studies, flexible analysis pipelines, and publication bias conspire to fill the literature with findings that don't hold up. Pre-registration and registered reports are institutional responses, but adoption is uneven. What I find useful: distinguishing exploration from confirmation before you start, being explicit about how many things you tested, and treating any single analysis as hypothesis-generating rather than hypothesis-confirming. What tools or practices have you found genuinely useful for staying honest when working with large datasets?
- spurious correlation
- multiple comparisons
- p-hacking
- causal inference
Ruhan