Notes on AI, machine learning, data systems, mathematics, statistics, and the psychology of building reliable things.
302 notes
- The Cut Survives The ContractionsKarger's min-cut algorithm is a small Monte Carlo argument: a minimum cut is found if none of its edges are contracted. A brute-force oracle and si...
- The Password Hash Rents MemoryA password hash is not just a slow hash. Memory-hard schemes make offline guessing rent RAM, burn bandwidth, carry per-user salts, and fit inside a...
- The Hash Join Gets a SpillwayA hash join larger than memory survives by partitioning both inputs into matching batches. Spilling is not a side effect; it is the algorithm's esc...
- The Tree Rotates Toward ReuseA splay tree keeps no balance bits. It pays rotations after each access so the next access can be cheaper if the trace has locality.
- The Difference Peels Out of the TableAn invertible Bloom lookup table lets two large, similar sets subtract their summaries and recover only the keys that differ by peeling pure cells.
- The Posting List Splits Its BitsElias-Fano encoding stores a monotone integer list by writing low bits literally and turning high bits into a selectable unary bitmap.
- The Counter Only Remembers the ExponentMorris approximate counting trades exact increments for a tiny logarithmic register, with an unbiased estimator and a variance bill you can compute.
- The Smallest Counter Takes the BlameSpace-Saving finds stream heavy hitters by replacing the smallest counter and carrying its value forward as an explicit error receipt.
- The Alphabet Splits Into BitsA wavelet tree makes rank, select, access, and range quantile queries by recursively turning a large alphabet into stable bitvector decisions.
- The Best Candidate Waits For A RecordThe secretary problem is not really about hiring. It is a finite optimal-stopping contract: spend a sample, then accept the first record, and verif...
- The Quantile Keeps Five MarkersThe P2 algorithm estimates one streaming quantile with five moving markers. Its strength is fixed memory; its price is that the error is empirical,...
- The Line Carries an Error LedgerBresenham line drawing is not just old graphics lore. It is an incremental proof that one integer error term can choose adjacent pixels without mul...
- The Range Minimum Hides a TreeA Cartesian tree turns array order into inorder traversal and range minima into lowest common ancestors. The monotone stack is the linear-time rece...
- The Ready Queue Is a ProofTopological sorting is not just a way to print a build order. The zero-indegree queue is a local certificate of legality, and the critical path say...
- The Row Group Keeps a Bounding BoxZone maps and data-skipping indexes do not find rows. They prove that whole row groups cannot match a predicate, and their power comes from physica...
- The Heartbeat Is a Suspicion ScorePhi-accrual failure detectors do not turn silence into an immediate yes/no verdict. They convert heartbeat delay into a tail probability, then let ...
- The Rumor Needs a FanoutEpidemic dissemination turns broadcast into repeated random contacts. Push spreads quickly at the beginning, pull cleans up the tail, and push-pull...
- The Snapshot Draws a CutChandy-Lamport snapshots record a meaningful global state without stopping a distributed system. The trick is to record both local process state an...
- The Counterexample Needs a LoopSafety bugs can stop at a bad prefix. Liveness bugs need an infinite witness, so finite-state model checkers hand you a prefix plus a loop.
- The Page Split Climbs Toward the RootA B+ tree index stays shallow because each node is a page-sized roster of keys. Insertions are local until a full page splits, promotes a separator...
- The Residual Graph Is the Audit TrailMax-flow algorithms do not just push capacity forward. The residual graph records the unused lanes and the undo lanes, and the final unreachable se...
- The Vacuum Has a HorizonMVCC makes reads and writes pass each other by keeping old row versions. Vacuum can clean only behind the oldest snapshot that might still see them.
- The Queue Has an AreaLittle's Law is not a magic queueing formula. It is an accounting identity over the right observation window, and cropped logs can quietly lose the...
- The Estimate Has Two ReceiptsA doubly robust estimator checks an observational causal estimate against both the treatment assignment model and the outcome model. One correct re...
- The Discovery Budget Has a ClockOnline false discovery control is not batch FDR with a refresh button. A stream of tests needs thresholds that are chosen before the next p-value a...
- The Isolation Boundary Needs ConsentCross-origin isolation is not one magic header. COOP isolates the opener relationship, COEP sets an embedder contract, and CORP or CORS lets resour...
- The Request Carries Its ContextFetch Metadata headers let a server classify the browser shape of a request before application code spends work or trusts ambient cookies.
- The Cookie Has a Site BoundarySameSite is not a domain flag or a CORS setting. It is a browser retrieval rule for deciding whether an ambient cookie rides along with a request.
- The Script Has a Guest ListContent Security Policy for scripts is not one rule. Host allowlists, nonces, hashes, and strict-dynamic answer different questions about which Jav...
- The Code Needs a ProofPKCE turns an OAuth authorization code from a bearer artifact into a two-step receipt: the front channel carries a challenge, and the token endpoin...
- CORS Is Not a FirewallCross-Origin Resource Sharing is a browser response-sharing protocol. Preflight requests ask whether a script may send a shaped request; the final ...
- The Cache Entry Needs a WitnessHTTP caching is not only a TTL. Freshness says when a stored response may be reused without asking; validators say how to ask the origin whether th...
- The Bucket Has a ClockA token bucket is not a vague request quota. It is a small arrival-curve contract: spend saved credit on bursts, refill at a steady rate, and prove...
- The Stack Knows Which Cycle ClosedTarjan's SCC algorithm turns one depth-first search into a component partition by keeping lowlink values and popping only when an open cycle can no...
- The Matching Arrives in WavesHopcroft-Karp speeds up bipartite maximum matching by batching a maximal set of shortest augmenting paths instead of repairing the matching one pat...
- The Diff Keeps a Furthest FrontierMyers' diff algorithm turns longest common subsequence into a shortest path problem, then searches by remembering only the furthest point reached o...
- The Puzzle Chooses Its Tightest ColumnAlgorithm X is backtracking with an accounting discipline: express the puzzle as exact cover, branch on the most constrained uncovered column, and ...
- The Diameter Waits For An AntipodeThe farthest pair in a planar point set is not hiding in the interior. Build the convex hull, rotate a support line around it, and the diameter app...
- The Operator Carries Its Binding PowerPratt parsing turns expression grammar into a table of token behaviors: each token knows how to start an expression, how to extend one, and how tig...
- The Tree Fits in the ParenthesesSuccinct tree representations separate topology from navigation: a rooted ordered tree can live in a balanced-parentheses bitstring, while rank, se...
- The Mean Needs a FirebreakMedian-of-means estimates a mean by splitting samples into blocks, averaging locally, and letting the median discard a few damaged blocks.
- The Mid Has a Queue ShadowThe midprice ignores how much size is waiting at the best bid and ask. A one-step microprice treats the next quote move as a queue-depletion race.
- The Minima March RightSMAWK searches a totally monotone matrix without materializing it. The row minima are not independent; total monotonicity lets one comparison disca...
- The Piles Remember the FrontierPatience sorting computes a longest increasing subsequence by keeping one small tail value for every possible length. The piles are not the answer;...
- The Integrator Kept ChargingA PI controller with a saturated actuator can keep accumulating error after more command can no longer help. Anti-windup is bookkeeping for that hi...
- The Search Path Has an AddressA binary search tree can do the same comparisons and still touch very different memory blocks. A cache-oblivious layout changes the address sequenc...
- The Cell Has Two DiagonalsMarching squares turns a sampled scalar field into contour segments with a 16-case table. Two cases are ambiguous because the corners alone do not ...
- The Transcript Is a Key InputTLS 1.3 does not only derive keys from Diffie-Hellman bytes. It feeds the handshake transcript into HKDF labels, CertificateVerify, and Finished so...
- The Bucket Becomes the QueueHash tables earn constant-time behavior from a distributional assumption. Hash-flooding attacks spend chosen input to break that assumption and tur...
- The Timeout Is Not a StopwatchTCP's retransmission timeout is a guarded suspicion about silence: estimate the path, price the variance, and refuse ACK samples whose origin is am...
- The Nonce Has One JobAES-GCM is fast because each nonce/key pair creates a one-time encryption stream and a one-time authentication mask. Reuse the nonce and both contr...
- A Gap Is an Event, Not a CharacterNeedleman-Wunsch alignment is a dynamic program over possible paths through a grid. Gotoh's affine-gap version adds state so one long insertion or ...
- The Run Stack Remembers the OrderTimsort is often summarized as mergesort plus insertion sort. The more interesting idea is that it treats existing order as data: find natural runs...
- The Parent Pointer Learns a ShortcutUnion-find keeps a changing partition by storing parent pointers. Union by rank keeps the trees shallow; path compression makes every search repair...
- The Error Bar Is CurvatureFisher information turns local likelihood curvature into a lower bound on estimator variance. The Cramer-Rao bound is a geometry statement before i...
- The Old Code Still Had a SwitchKnight Capital's 2012 loss was not just a bad deploy. It was stale production code, a reused flag, missing router-level controls, and a live market...
- The Texture Needs a Smaller CopyMipmaps are prefiltered promises about pixel footprint: when one screen pixel covers many texels, sample a smaller copy instead of one lucky texel.
- The Page Must Not Outrun the LogWrite-ahead logging makes durability a recoverable story: before a dirty page reaches disk, the log record that explains it must already be durable.
- The Ruler Is a RatioWeber's law, Fechner's logarithmic scale, and Stevens' power law are not three interchangeable slogans. They are different ways to turn physical in...
- The Image Pays One Pixel at a TimeSeam carving turns resizing into a shortest-path problem: remove one low-energy connected pixel path, recompute the image, and audit what the resiz...
- The Fingerprint Has Two HomesA cuckoo filter keeps only short fingerprints, so deletion works because each fingerprint still carries enough information to find its other bucket.
- The Square Root Was in the ExponentQuake III's famous fast inverse square root is not dark magic. It is a logarithm hiding in IEEE float bits, followed by one Newton repair step.
- The Auction Should TickA continuous limit order book can turn public information into a private speed prize. Frequent batch auctions change the question from who arrived ...
- The Conflict Writes a ClauseCDCL solvers do not just backtrack when a SAT branch fails. They turn the failure into a resolvent, jump past irrelevant decisions, and make the sa...
- The Cursor Needs an AddressCollaborative text editing is not just merging strings. Every inserted character needs a stable identity, and the wrong address scheme can make two...
- The Feature Has a Shadow TwinKnockoff filters control false discoveries by making every feature race a synthetic twin that has the same correlation structure but no direct clai...
- The Alarm Has a Run LengthCUSUM and Page-Hinkley detectors are not magic drift bells. They are stopping rules that spend false-alarm budget to buy detection delay.
- The Distance Table Replaces the VectorProduct quantization makes vector search fast by turning every stored embedding into a handful of small codebook IDs. The query stays real; the dat...
- The Distance Is an XOR ShadowKademlia routes through a distributed hash table by treating closeness as XOR distance. That turns lookup into a prefix game with k-buckets, stale ...
- The Arbitrage Is a Negative CycleCurrency arbitrage becomes a graph problem after one logarithm. Multiplicative exchange rates turn into additive path weights, and a profitable rou...
- The Watermark Is a Promise About the PastEvent-time stream processing is a negotiation between correctness and latency. A watermark is the system's current promise about which old events a...
- The Dictionary Has Failure LinksAho-Corasick turns many string patterns into one automaton. The trick is not only a trie; it is knowing where to fall when a partial match breaks.
- The Interior Path Keeps Its DistanceInterior-point methods solve constrained optimization by making every wall expensive before touching it. A small LP lab follows the central path an...
- The Variable Order Is the AlgorithmReduced ordered binary decision diagrams are canonical only after you choose an order. The wrong order can make a small Boolean function look enorm...
- The Triangle Has a Sign BitThe orientation predicate looks like a tiny determinant. In computational geometry, that sign bit is a topology decision, and ordinary floating-poi...
- The Graph Forgets the TrianglesMessage-passing graph neural networks inherit a precise blind spot from color refinement. The 1-WL test can be strong, but it can also stare at two...
- The Graph Is an Equivalence ClassObservational causal discovery usually cannot identify one directed graph. Markov equivalence says the data sees a skeleton and unshielded collider...
- The Signature Spends a Secret OnceIn DSA-style signatures, the per-message nonce is not harmless randomness. Reuse it, shrink it, or put it on a clock, and the private key turns int...
- The Bin Is a FilterA DFT bin is not a tiny bucket that catches one exact frequency. It is a window-shaped filter, and spectral leakage is what happens when the tone f...
- The Collider Opens When You LookD-separation is the graph rule behind conditional independence in Bayesian networks: chains and forks close when you condition, but colliders open ...
- Every Path Has a GatekeeperDominator trees turn control flow into a question a compiler can answer: which blocks must every path pass through before reaching this point?
- The Map Becomes a StringA spatial index is often a strange act of compression: take a rectangle on a grid, turn it into a few one-dimensional key ranges, and hope locality...
- The Max Has a TemperatureEntropy-regularized control does not merely add random exploration after planning. It changes the Bellman backup itself: max becomes log-sum-exp.
- The Tree Gets More Lottery TicketsGreedy tree splits compare the best cut each feature can find. A feature with many possible cuts gets more chances to look important by accident.
- The Delta Is the QueryIncremental view maintenance works when change is smaller than state. The hard part is choosing which partial answers deserve to be remembered.
- One Shadow Is EnoughA collision detector does not need to understand two convex shapes all at once. If one projection leaves daylight between them, the case is closed.
- Move Two Knobs Or Miss The ModelOne-at-a-time sensitivity analysis walks along coordinate axes. Sobol indices ask how much output variance each input owns, including the variance ...
- The Path Is Not Its MarginalsViterbi decoding and posterior decoding both look through the same HMM trellis, but they optimize different losses. The most likely story is not al...
- The Menu Has EntropyHick-Hyman law is not a bumper sticker for removing choices. It is a small information-theoretic model of choice reaction time, and it gets useful ...
- The Cut Has a DenominatorMinimum cuts like cheap separations. Normalized cuts ask whether the separated set had enough internal graph volume to deserve being called a cluster.
- The Average Arrives LateRobbins-Monro stochastic approximation learns a root from noisy experiments. The step size has to keep moving long enough to find the target, then ...
- The Record Has No FieldsHyperdimensional computing stores symbols as wide random vectors. Binding and bundling let a record live in one vector, but cleanup only works whil...
- The Checksum Is a Polynomial RemainderA CRC is not a small hash pasted onto a packet. It is polynomial division over GF(2), and an error is missed exactly when the error polynomial is d...
- The Memory Chooses Its DecayMamba-style state-space models are recurrent, but the important move is selectivity: the update rule can decide which tokens write, which tokens fa...
- The List Grows TowersSkip lists replace tree rotations with coin flips. The trick works because a random tower of forward pointers gives logarithmic search in expectati...
- The Curve Moves in Three ShadowsYield curves are high-dimensional objects, but many daily moves look like mixtures of level, slope, and curvature. PCA makes that visible; risk sti...
- The Counterfactual Is a PortfolioSynthetic control builds a missing no-treatment path as a weighted portfolio of untreated units. The weights are transparent; the causal claim is n...
- The Motif Is a Nearest NeighborMatrix profiles turn time-series motif discovery into a nearest-neighbor ledger over sliding windows. The smallest entry finds repeated shape; the ...
- The Kernel Keeps a Few WitnessesNystrom kernel approximation replaces a full Gram matrix with a few landmark columns. The hard part is not the formula; it is choosing which witnes...
- The Posterior Pays RentPAC-Bayes bounds let the learned distribution depend on the data, but charge it KL divergence from a data-independent prior.
- The Normal Equations Square the TroubleLeast squares is one objective but not one computation. Forming A^T A can square the condition number; QR keeps the geometry closer to the data.
- The Label Is a Latent VariableMajority vote treats annotators as interchangeable sensors. Dawid-Skene style truth inference treats the true label and each worker's confusion mat...
- The Line Has a Breakdown PointLeast squares is a voting rule with a very loud microphone for far-away points. Robust regression is not one trick; Huber, Theil-Sen, and RANSAC fa...
- The Sum Has a ScheduleFloating-point addition is commutative but not associative. Once a sum is parallelized, shuffled, vectorized, or compensated, the reduction order b...
- The Remainder Is a Quotient EstimateBarrett and Montgomery reduction replace division by a fixed modulus with multiplication, shifts, and a small correction loop. The trick is not avo...
- The Type Checker Finds the Most General ShapeHindley-Milner inference feels like magic because it returns a principal type. Under the hood it is a small equation solver with one crucial habit:...
- The Plan Is a Row-Count ForecastA SQL optimizer does not merely choose joins. It makes a forecast about intermediate rows, then spends real I/O and CPU as if that forecast were true.
- The Event Loop Is a QueueAsync I/O removes waiting, not work. A single slow callback can turn a calm service into a tail-latency machine unless the queue is visible and bou...
- The Model Still Owes a FilterLearned Bloom filters are not classifiers with nicer marketing. The model may save bits, but the one-sided membership guarantee still has to be pai...
- The Cache Forgets the MiddleKV-cache eviction is a lossy attention policy. Sliding windows, attention sinks, heavy hitters, and observation-window voting each make a different...
- The Kernel Decides What a Difference IsMaximum mean discrepancy turns a vague distribution-shift question into a precise measurement. The catch is that the kernel is not decoration: it d...
- The Username Is Not a StringUnicode lets identifiers speak many languages, but raw string equality is the wrong security boundary. Normalization, case folding, script policy, ...
- The Object Has More Than One DoorGFlowNets are easiest to misunderstand if you think they optimize for the single best object. Their stranger job is to sample objects in proportion...
- The Rounder Is the TurnstileLow-precision training is not just a bit-width choice. Once updates are smaller than a grid step, the gate rule decides whether they are blocked or...
- The Address Came Back WrongThe ABA problem is what happens when compare-and-swap sees the same pointer twice and mistakes equality for history.
- The Edge Pixel Is a ContractAlpha compositing works when color, coverage, and color space agree. Premultiplied alpha is the contract that keeps filtered edges from lying.
- The One Percent Chance Has a PriceCumulative prospect theory is a model of how risky choices can be priced relative to a reference point, not by expected value alone.
- The Failing Test Case Shrinks by Asking HalvesDelta debugging turns a huge failing input into a small witness, but its promise is 1-minimality under a test oracle, not omniscience.
- The Front Door Opens Through the MiddleThe front-door criterion can identify a causal effect with hidden treatment-outcome confounding, but only by using a mediator in a very particular ...
- A Network Is a Kernel Until It MovesThe neural tangent kernel turns training into a fixed linear model in the infinite-width limit. A small lab shows what that approximation looks lik...
- The Bitmap Is a Container MapRoaring bitmaps compress integer sets by cutting the universe into 16-bit chunks and choosing a local container for each chunk's shape.
- The Prefix Sum Learns to MoveA Fenwick tree keeps an editable cumulative frequency table by storing suffix blocks whose widths are written in the low bits of their indexes.
- Ask Two Queues, Not the Whole FleetThe power-of-two-choices result says a load balancer can crush the maximum-load tail by comparing two random servers instead of trusting one random...
- The Loaded Die Has a TableThe alias method turns a fixed categorical distribution into equal-width boxes, so a draw needs one random column and one threshold comparison.
- The Gradient Needs a RulerA gradient is not a direction until a geometry says what a unit step means. On the probability simplex, the wrong ruler can turn a distribution fit...
- The Baseline Is a Control VariateIn policy gradients, a baseline is allowed because the score has zero mean. The best baseline is not always the average reward.
- The Likelihood Ratio Is a Running ScoreA sequential test is not permission to keep refreshing a p-value. It is a pre-committed stopping rule whose state is the accumulated log evidence.
- The Random Timeout Is a Vote SpacerRaft's election timeout is not just a wait. It is a randomized spacing device that makes split votes rare without making safety depend on clocks.
- The Prefix Sum Is an Address MachineParallel scan looks like a cumulative sum, but its real power is turning local facts into global addresses with logarithmic synchronization.
- The Weight Is a Receipt for Being MissedHorvitz-Thompson weighting turns unequal sampling chances into a finite-population estimator, but the variance bill arrives in the weights.
- The Instant Has to Fit Inside the CallLinearizability asks whether every concurrent operation can be assigned one legal instant between invocation and response.
- The Write Did Not Happen EverywhereA tiny store-buffering litmus test shows why memory order is a contract, not a timestamp.
- The Singleton Points Beyond the SampleGood-Turing and unseen-species estimators turn one-off observations into evidence about the probability mass you have not seen yet.
- The Pixel Is an ExpectationPath tracing works because a pixel can be written as an integral, and a ray can be treated as a random variable.
- The Hedge Is a Variance MeterA delta-hedged option does not merely bet on direction. It keeps a ledger of realized variance, rebalancing error, jumps, and transaction costs.
- A Dataset Can Be a Pull RequestData poisoning is what happens when training examples become a write path into model behavior.
- The Bounds Check Was Not a BoundarySpectre-style attacks broke a quiet assumption: discarded speculative work can still leave measurable traces in microarchitectural state.
- The Page Is Not a DeveloperPrompt injection is a control-plane bug in LLM applications: untrusted text is allowed to compete with trusted instructions, tools, and secrets.
- The Watermark Is a Tail TestLLM text watermarks are not visible signatures. They are statistical evidence accumulated token by token, and their reliability lives in entropy, l...
- The Mask Is a Second ClockDiscrete diffusion language models make generation a schedule over unknown tokens, not just a left-to-right stream. The hard part is deciding what ...
- The Query Remembers Only New FactsRecursive Datalog queries are fixed points over relations. Semi-naive evaluation makes the fixed point fast by propagating only the facts that were...
- The Error Changes GridsMultigrid is fast because it does not ask one iteration to see every scale. It smooths oscillatory error, moves the remaining error to a coarser gr...
- The Filter Peels Before It AnswersXor filters look like a three-probe membership test. The construction is stranger: build a random hypergraph, peel it, then solve a sparse XOR syst...
- The Search Distribution Learns CoordinatesCMA-ES is not magic mutation. It is a rank-based optimizer that lets a Gaussian search distribution learn the valley it is walking through.
- Two Thirds Is Not a VibeByzantine quorum certificates work because two possible histories are forced to share an honest voter, not because 67% has mystical distributed-sys...
- Discoveries Need a DenominatorFalse discovery rate control is not a ritual after p-values. It is a contract about the list you are willing to publish.
- Pixels Can Keep One LightReservoir resampling turns a stream of many possible light samples into one carried sample plus a weight ledger. The magic is not the sample; it is...
- Waiting for the Wrong FluctuationKramers escape explains why a system can look settled while rare noise quietly prices the exit.
- Grids Have Speed LimitsThe CFL condition is not a magic stability spell. It is a geometric warning: a time step cannot outrun the stencil that is supposed to carry inform...
- Loops Turn Evidence Into RumorBelief propagation is exact on trees because every message summarizes disjoint evidence. On a loopy graph, the same local rule can double-count its...
- Controllers Solve BackwardFinite-horizon LQR turns a future cost into a present feedback law. The Riccati recursion runs backward; the controller runs forward.
- Holes Have to PersistPersistent homology turns a point cloud into a scale experiment. Components and loops are allowed to appear and die; the barcode records which ones...
- Fixed Points Remember MistakesAnderson acceleration turns a slow fixed-point iteration into a small residual-canceling least-squares problem. The magic is real, but only if you ...
- Sketches Shrink Every DirectionFrequent Directions is a deterministic streaming sketch for covariance. It keeps a tiny matrix, repeatedly shrinks the spectrum, and still promises...
- Normalizers Hide in the PathAnnealed importance sampling estimates a partition function by walking through easier distributions and averaging the resulting path weights.
- Epsilon Is Not a PurseDifferential privacy is easier to reason about when epsilon is a likelihood-ratio threshold, delta is tail slack, and the accountant keeps the whol...
- Critical Points RememberAt the Ising critical point, local Monte Carlo updates keep revisiting the same large domains. A cluster move changes the unit of motion.
- TFLOP Numbers Are Not SpeedometersA processor's peak FLOP rate is only one roof. The other roof is memory bandwidth, and the useful question is how much arithmetic each byte gets to...
- Eigenvalues Are Not WeatherA stable matrix can amplify disturbances before it damps them. The missing diagnostic is not another eigenvalue, but the geometry of the eigenvecto...
- First Digits Live on a CircleBenford's law is not a numerology trick about 1s. It is a statement about logarithmic mantissas, scale invariance, and when a dataset has no prefer...
- Urns Remember Without OrderA Polya urn is reinforced and path-dependent, but still exchangeable. The order of the draws can disappear even when early luck never does.
- Cuckoo Hashing Fails as a GraphThe basic cuckoo table does not fail because the insertion loop is unlucky. It fails when the hidden bipartite graph contains too many edges for it...
- Flocking Is an Order ParameterA flock does not need a leader. In the Vicsek model, local alignment, finite speed, and angular noise are enough to create a measurable transition ...
- CG Buys One More DegreeConjugate gradient is not just gradient descent with a better memory. Each iteration buys another degree of polynomial freedom to suppress the eige...
- Stop Before the Vector Turns AroundA two-amplitude Grover lab for quantum search, amplitude amplification, overshoot, phase calibration, and the square-root limit of black-box search.
- Random Dots Need Personal SpaceA graphics and games lab for comparing random points, jittered grids, best-candidate sampling, and Bridson Poisson disks with gaps, voids, and spec...
- Momentum Loans for the PosteriorHamiltonian Monte Carlo proposes by moving like a physical system: conserve energy, preserve volume, then let Metropolis audit the numerical error.
- Ask Random Vectors for the TraceHutchinson's estimator turns random vectors into a matrix-free trace meter. The trick is unbiased; the hard part is controlling variance when the s...
- Mistakes Burn CapitalMultiplicative weights is a loss ledger with teeth: every expert keeps capital, bad rounds shrink it, and the learner gets a certificate against a ...
- Next Reaction Wins the RaceGillespie's stochastic simulation algorithm samples chemical reaction paths exactly by letting propensities compete as exponential clocks.
- Cliffs Live at InfinityPercolation has a clean infinite-lattice threshold, but every experiment sees a rounded finite box. The browser can show the cliff only by watching...
- Last Column Knows Where to SearchThe FM-index turns the Burrows-Wheeler transform into a searchable compressed index: exact matching becomes a right-to-left interval-narrowing game.
- Every Tile Starts as a MaybeWave Function Collapse looks like procedural magic, but the useful version is a constraint solver with a texture memory: local patterns, entropy ch...
- Buttons Make Promises to HandsFitts' law treats pointing as information transmission: distance, width, time pressure, and endpoint scatter decide whether a target is actually us...
- Next FLOP Has to ChooseScaling laws are not just a slogan about bigger models. Under a fixed training budget, parameters and tokens trade off, and data scarcity bends the...
- Leases Expire; Writes Do NotLease expiry is a time-based promise, but stale side effects need an epoch check at the resource. Clock padding and fencing tokens solve different ...
- When the Queue Answers Too LateTCP congestion control is a delayed conversation: the congestion window probes, the bottleneck replies with loss or delay, and a fat buffer can mak...
- Game Trees Tell the Wrong StoryAlpha-beta search is drawn as a tree, but repeated positions make the real work a graph. A transposition table is safe only when it stores identity...
- Stopwatches Are Not CalendarsWall clocks name calendar instants; monotonic clocks measure elapsed time. Mixing them turns NTP corrections, clock steps, and lease expiries into ...
- Smiles Pay in ButterfliesA volatility smile is not just a curve of Black-Scholes inputs. Through call-price curvature, it encodes a risk-neutral density, and the butterfly ...
- Search That Refuses to RereadKMP string search is linear because a mismatch leaves evidence behind. The failure table remembers which prefix is also a suffix.
- Do Not Spend the Budget on the Already GoodA treatment can be bad on average and still worth targeting. CATE and uplift models ask who changes because of treatment, not who would have done w...
- Denominators Can Learn the FutureImmortal time bias is a time-zero bug in causal survival analysis. A treatment can look protective because the treated group had to survive long en...
- Prediction Errors Leave FootprintsTD(lambda) gives the next surprise a trail to follow, so delayed reward can reach the states that made it likely.
- Heatmaps Are Not CircuitsActivation patching can be a real causal intervention, but a restored logit is evidence about a contrast, a metric, and a site, not an explanation ...
- Your Example Brought Its NeighborsBatch normalization is stateful and contextual: in training mode each activation is normalized by the mini-batch around it, then eval mode switches...
- Step Sizes Are DaresA learning rate is not just speed. It is a numerical stability promise about the sharpest direction the optimizer is willing to survive.
- Thresholds Only Speak NearbyRegression discontinuity turns a hard rule into a narrow causal comparison. The useful question is not whether the rule is sharp, but whether anyth...
- Samplers Can Pace in the Wrong RoomA Metropolis chain can accept most proposals and still fail to explore. Convergence is a claim about what the chain has forgotten, not how busy it ...
- Sharpe Ratios Have ClocksAnnualizing a Sharpe ratio by sqrt(252) is not a law of nature. It is an assumption about the sampling clock, the autocovariances, and how returns ...
- Vote Gaps Draw the BallRandomized smoothing turns a classifier into a Gaussian vote. When the top class wins by enough, the prediction earns a certified L2 ball.
- Prediction Is a Witness, Not a LeverGranger causality is incremental forecasting. It becomes causal evidence only after the conditioning set, lag order, and measurement process surviv...
- Give Every Candidate a ClockThe Gumbel-max trick turns a categorical draw into a race: add one random key to every log weight, take the winner, or sort the same keys for sampl...
- Every Token Gets a Clock HandRotary position embeddings do not paste a position vector onto a token. They rotate query and key coordinates so attention reads relative displacem...
- Pager Starts DrummingA Hawkes process notices when an event is also fuel for the next event. The mean rate can look harmless while the clock is carrying aftershocks.
- Missing Shards Are EquationsReed-Solomon erasure coding is not backup magic. It is the promise that any k honest surviving shards contain enough GF(256) equations to rebuild t...
- Crash Risk Lives in the CornerA portfolio can keep the same one-name loss distributions and broad rank association while changing its joint crash probability. The missing object...
- Smoothing Drew the SpotsReaction-diffusion systems bend intuition: local mixing plus nonlinear reactions can turn a nearly uniform field into spots, worms, fronts, and decay.
- Matching, but With Soft EdgesEntropic optimal transport turns a hard matching problem into matrix scaling. Sinkhorn repeats row and column normalization until a soft transport ...
- Prompts Are Ratios, Not CommandsClassifier-free guidance is not magic obedience. In the score idealization, the slider raises a conditional-vs-unconditional density ratio and buys...
- Case File: The Heap Was Not FullHigh RSS is not always a leak. Fragmentation, size classes, phase changes, and thread-local caches can make the allocator's ledger disagree with li...
- Delete Buttons Owe CounterfactualsMachine unlearning is not just deleting a row. It is the demand that a deployed model behave as if that row had never entered training.
- Do Not Reassign the World to Add One BoxConsistent hashing is mostly a remapping contract: when the fleet changes, only the keys that genuinely belong on the new owner should move.
- A* Trusts a Promise, Not a HunchA* is fast when its heuristic makes a mathematical promise about the remaining path cost. Break that promise and the speedup becomes an explicit bet.
- Memtables Borrow From TomorrowAn LSM tree makes writes cheap by turning random updates into sorted runs. Compaction is where the sorting loan comes due.
- One Transcript Can Be FakeA zero-knowledge proof is not a magic blackout curtain. In a Sigma protocol, a fake transcript can look real, while two forked answers reveal the w...
- Where the Half-Bits GoEntropy coding is the quiet accounting beneath compression: the model prices each symbol, and the coder pools those fractional prices into a revers...
- Clocks Can Say ConcurrentLamport clocks preserve causal order, but scalar timestamps still sort concurrent events. Vector clocks keep the partial order visible instead of i...
- CPUs Have to Guess Right NowBranch prediction is online learning at the front of the processor: tiny adaptive models guess control flow before the real label arrives.
- Let the Clause SleepModern SAT solvers are fast partly because most clauses stay dormant. Two watched literals turn unit propagation into a lazy alarm system for memor...
- Small Models Can Be ScoutsSpeculative decoding is exact because the draft model is only a proposal distribution. The verifier keeps the overlap and samples the missing targe...
- Side Deals Are the BugDeferred acceptance is not a sorting algorithm. It is a way to remove every mutually tempting off-book deal while choosing one endpoint of the stab...
- p95 Needs Its CrowdA p95 is a rank statement about a population, not a number you can average across shards. Sketches work because they keep rank evidence alive.
- Tails Come Back Through the Front DoorFFT convolution is fast, but the finite DFT lives on a circle. Without enough zero padding, the end of the answer wraps around and pollutes the beg...
- Gradients Keep ReceiptsReverse-mode automatic differentiation is not numerical differencing and not symbolic algebra. It is a transposed execution of the program, and the...
- Ask the Matrix Where It EchoesRandomized SVD works because a matrix can be interrogated by random probes. The probes do not inspect every entry; they listen for the subspace whe...
- Give QMC the Big Move FirstQuasi-Monte Carlo option pricing depends on the map from uniforms to paths. A Brownian bridge keeps the Brownian law but moves coarse path variance...
- Prices Forecast With a TabPrediction markets and proper scoring rules are the same incentive idea wearing different interfaces. LMSR turns a log score into an always-on mark...
- Trees Spend ExperimentsMonte Carlo Tree Search is not just lookahead. It is an adaptive sampling rule: spend simulations where uncertainty and upside are still worth payi...
- The Tokenizer Comes With a PhrasebookA BPE merge table is a learned phrasebook from a reference corpus. It gives familiar strings short names and leaves unfamiliar strings to spell the...
- Ladders Need Error BarsElo-like ratings are not prizes. They are online estimates of latent skill from noisy pairwise comparisons, and the dangerous omission is uncertainty.
- Games Guess the PastRollback netcode does not make packets faster. It lets online games run now, predict missing input, and repair the timeline when reality arrives.
- Rare Targets Teach You to QuitLow-prevalence visual search is not ordinary perception with fewer positives. Rarity changes item criteria, quitting thresholds, and mistakes.
- Matrices Were Never RequiredFlashAttention is exact attention scheduled differently: keep the running max, rescale the denominator, and avoid writing the full attention matrix...
- Same Bid, Different FuturesIn a limit order book, two orders at the same price are not the same trade. FIFO priority changes fill probability, adverse selection, and the valu...
- Patterns Can Hide Search TreesA regex can look like a declarative shape while the engine underneath is doing a depth-first search through guesses.
- Experiments Hide in the ResidualsDouble machine learning is not a confounding eraser. It builds a residual score so nuisance-model mistakes have less leverage over the causal number.
- Nearest Neighbor Search Needs RoadsHNSW-style vector search works because the index is navigable. Local edges give precision, long edges give reach, and the search beam is the fuel b...
- Put Both Rankers on the Same PageA/B tests compare rankers through different users and queries. Interleaving makes them share one result page, turning ranking evaluation into a pai...
- Folds Remember Label WindowsRandom cross-validation can leak through time even when every feature is legal. The culprit is often the label interval, not an obviously forbidden...
- Review Queues Spend MemorySpaced repetition is not a productivity spell. It is a control problem over memories whose recall probabilities decay at different rates.
- Blurry Interpreters Live in Your CompilerStatic analysis runs the program in a smaller universe: types, signs, intervals, shapes, effects, or any abstraction cheap enough to compute and so...
- Averages Lose Innocence in 3-DStein's paradox says the coordinatewise sample mean can be beaten uniformly once you estimate three or more normal means at the same time.
- Keep the Rewrite Around Until the EndEquality saturation turns compiler optimization from a brittle sequence of rewrites into a search over equivalent programs.
- Good Reward Hints Cancel on the LoopPotential-based reward shaping can make reinforcement learning easier without changing the task. Arbitrary hints can quietly create a new game.
- Stores Can Save Live ObjectsA garbage collector is a graph traversal racing a mutating graph. The write barrier is the small memory of a store that keeps the traversal true.
- Bring the Receipt, Not the Whole LogA Merkle proof does not say trust me. It says this exact item belongs to this exact committed root, and here are the sibling hashes to check.
- Two Clean Snapshots Can Break the WorldDatabase isolation levels are not adjectives about safety. They are claims about which shapes of history are allowed to exist.
- Adam Chooses the Units Before It StepsAdam is not just SGD with momentum. It keeps a per-coordinate memory of gradient scale, then silently decides what a unit step means.
- If You Keep Looking, Use a BankrollE-values and e-processes turn sequential evidence into a nonnegative martingale, so repeated looks become part of the design instead of a loophole.
- AUC Only Sees the PairROC AUC is not accuracy, calibration, alert quality, or deployment utility. It is the probability that one random positive outranks one random nega...
- Hash Tables Are Fast Until the Tail ArrivesOpen-addressed hash tables do not merely get fuller. Their probe distributions change shape.
- SHAP Bars Need a Chosen WorldSHAP values are not pure feature truth. They are Shapley values for a particular missing-feature game, background population, and prediction scale.
- LoRA Bets on the Missing DirectionsLoRA is not magic compression. It freezes the base model and asks whether the task-specific update mostly lives in a small spectral neighborhood.
- Collisions Keep the ReceiptsCount-Min is not a tiny dictionary. It is a one-sided bargain: lose identities, keep upper bounds, and budget the collision mass.
- PageRank Reenters Through the Side DoorPageRank is a damped random walk, and the teleport vector is the quiet place where outside judgment enters the graph.
- Reliable Bits Need Empty SpaceError correction is not a spell that makes bits reliable. It is a geometric bargain: spend redundancy to move valid messages farther apart.
- Formulas Can Lie to the ComputerVariance looks like a formula, but in finite precision arithmetic the algorithm you choose can decide whether the answer is accurate, noisy, or imp...
- Keep the Distances, Audit the DecisionsThe Johnson-Lindenstrauss lemma is not a slogan about making data smaller. It is a promise about finite geometry, with margins the application stil...
- Every Eviction Is a PredictionA cache replacement policy is not housekeeping. It is an online prediction about which object will be useful before memory runs out.
- Averages Mix Before They JudgeSimpson's paradox is not an arithmetic prank. It appears when a mixture is asked to answer a causal question it was not built to answer.
- Ask Who the Instrument MovedInstrumental variables do not recover the effect for everyone. Under strong assumptions, they recover the effect for people whose treatment was mov...
- Bootstrap the Assumption, Not the WishResampling is not magic. A bootstrap interval works by replaying a data-generating contract, and the wrong contract can be confidently wrong.
- Winning Can Be Bad NewsIn a common-value auction, winning is not just success. It is evidence that your estimate was unusually optimistic.
- Clusters Remember by OverlapReplicas do not remember because they agree in spirit. They remember when the next decision is forced to meet the last one.
- Who Leaves the Curve?Kaplan-Meier curves do not magically solve missing follow-up. They keep censored futures in the accounting and expose the assumption doing the work.
- Inverses Turn Noise Into BetsMean-variance optimization is elegant until the covariance matrix is estimated from too little history. The inverse turns tiny noisy eigenvalues in...
- Benchmarking Through the OutageClosed-loop latency tests can stop sending work exactly when the service is slow. Coordinated omission turns a tail-latency incident into a tidy pe...
- Do Not Trust a Naked ECECalibration plots are measuring instruments. Change the ruler, the bins, or the slice, and the confidence story can change with it.
- Sometimes the Posterior Needs a CrowdParticle filters approximate a nonlinear, non-Gaussian filtering distribution with weighted samples. The hard part is not drawing samples; it is ke...
- Button Presses Carry ClocksA drift-diffusion model treats a decision as noisy evidence accumulating until it crosses a threshold. The boundary prices time, mistakes, and unce...
- Yesterday Can Tighten an A/B TestCUPED borrows pre-period signal to reduce variance. Move the timing line past treatment, and precision quietly turns into bias.
- How Far Should a Noisy Sensor Move You?A Kalman filter is not a smoothing trick. It is a recursive argument about how much to trust a model versus a noisy measurement.
- Similar Pages Should Share a Lottery TicketMinHash turns document overlap into a collision event whose probability is the Jaccard similarity.
- Bit Arrays That Only Say MaybeA Bloom filter is a memory bargain: it can say definitely not, or maybe, but it never proves membership.
- Second Tries Are Still TrafficA retry is not free resilience. Under overload, it can become a traffic generator aimed at the component already failing.
- Fuzzy Joins Write FactsWhen there is no shared key, a data join becomes a statistical decision about identity under measurement error.
- Privacy Starts With a Volume KnobDP-SGD is often described as SGD plus noise. The better description is clipped influence plus accounted randomness.
- Midpoint as TruceA bid-ask spread is not one cost. It prices immediacy, adverse selection, inventory risk, and the mechanics of keeping a market open.
- Control Groups Hold Ghost SlopesDifference-in-differences is not before-minus-after. It is a wager that the control group carries the untreated trend the treated units never got t...
- Leaderboards Have Seating ChartsLLM-as-judge evaluation is not just a model call. It is a measurement protocol with position effects, verbosity effects, calibration choices, and s...
- Loss Functions Never See the Missing PairPreference optimization is not a direct tap into human values. It is a policy update seen through pairwise labels, a reference model, a KL temperat...
- Counting Without NamesHyperLogLog is the strange bargain behind large-scale distinct counts: forget identities, keep rare hash patterns, and still estimate how many uniq...
- Witnesses Are Not CamerasA memory report is not raw footage. It is reconstructed from noisy traces, gist, priors, cues, and a decision threshold.
- Merge Functions Become ProductsCRDTs are not a trick for avoiding distributed systems. They are a way of making merge algebra explicit: what grows, what can be forgotten, and wha...
- Alpha Was Hiding in the TimestampAt high frequency, a lead-lag signal can be market structure, a data-engineering artifact, or both. The Epps effect is the warning label.
- Context Windows Have Rent DueLong-context inference is not only attention math. Every live token leaves key and value vectors in GPU memory, and the serving stack pays that ren...
- Every Trial Spends the BudgetBayesian optimization is not optimizer dust. It is a policy for choosing the next expensive experiment under uncertainty, and its failures belong i...
- Models Forget the Weird Stuff FirstRecursive training can look healthy in the average case while sampling, filtering, and reuse quietly sand away rare modes.
- Dashboards Can Change the ExperimentA fixed-horizon p-value is a different object once the team checks it every morning and ships the first good-looking day.
- Risk Dials Have DelaysScaling exposure by estimated volatility is not a free Sharpe-ratio machine. It is a lagged control loop with estimator error, leverage caps, turno...
- Last Layer, Solver IncludedDiffusion and flow-matching models do not end at the neural network. The path, vector field, and numerical solver decide what distribution appears.
- Models Overbook Feature SpaceNeural networks may store more features than dimensions. Sparsity makes that bargain possible, and interference sends the bill.
- Every Expert Has a Line OutsideSparse mixture-of-experts models promise cheap active compute, but the router is really allocating scarce expert capacity under bursty demand.
- Straight Lines That Lie PolitelyHeavy tails are easy to see and hard to prove. A straight-ish log-log plot is a lead, not a verdict.
- One Outlier Can Spend the Whole ScaleQuantizing a language model is not just rounding weights. A few large activation channels can spend the whole numerical budget for everyone else.
- Draft Tokens Have to Earn Their KeepSpeculative decoding is often sold as free speed. The real question is whether accepted draft tokens repay the draft model, verifier overhead, and ...
- Vectors That Show Up EverywhereNearest-neighbor retrieval can fail because one vector becomes everybody's neighbor. Hubness is not a vague curse of dimensionality; it is a measur...
- Build a Copy Machine You Can SeeInstead of treating in-context learning as magic, build a browser-sized induction circuit that finds a repeated pattern and copies what followed it.
- After the Decision, the Trade BeginsA portfolio model has a decision price. The desk has a path through liquidity, volatility, impact, patience, and information leakage.
- Which Zero-Error Model Did You Get?Double descent is not a moral boundary between underfitting and overfitting. It is a clue that geometry, noise, and implicit regularization selecte...
- Feeds Teach Users, Then Take NotesA recommender does not merely observe preference. It allocates exposure, changes what users can learn about, and trains on behavior it helped create.
- Leaderboards Make Blunt InstrumentsAI benchmarks do not just score models. They build measuring instruments with ranges, reliability limits, blind spots, and exposure risks.
- Beautiful Backtests Come From a CrowdA beautiful backtest is weak evidence until the ideas, parameters, filters, assets, and windows that lost are counted too.
- Saying No Needs a Price ListA model should not answer because confidence cleared a universal threshold. The stop rule belongs to error costs, abstention costs, and the value o...
- Coverage Comes With Fine PrintConformal intervals are powerful because their guarantee is explicit: coverage under an exchangeability story, not magic under every deployment shift.
- Metrics Become Control SurfacesMetrics fail under pressure because selection, gaming, and feedback change the data-generating process that made the metric useful.
- Thirty-Two Tries Can Be One MistakeRepeated reasoning samples buy intelligence only when attempts are diverse, the selector is reliable, and the system knows when not to answer.
- RAG Is the Question of What Gets ReadLong-context models changed retrieval, but they did not remove the dispatch problem: deciding what evidence deserves attention, placement, and an a...
- Trust First, Then Size the BetKelly sizing is a clean formula with an unforgiving dependency: calibrated probabilities, executable odds, and a drawdown budget that survives cont...
- At 3 AM, Ask the CacheA production pattern for sparse, long-tailed estimates: robust rollups, partial pooling, visible fallback levels, and the discipline to ask for ML ...
- Notes With the Wires Left OutA small charter for these notes: technical ideas should show their evidence, their scars, and the places where a reader can push back.
Machine learning47
- The Graph Forgets the TrianglesLab
- The Max Has a TemperatureLab
- The Tree Gets More Lottery TicketsLab
- The Path Is Not Its MarginalsLab
- The Cut Has a DenominatorLab
- The Record Has No FieldsLab
- The Memory Chooses Its DecayLab
- The Kernel Keeps a Few WitnessesLab
- The Posterior Pays RentLab
- The Label Is a Latent VariableLab
- The Cache Forgets the MiddleLab
- The Object Has More Than One DoorLab
- The Rounder Is the TurnstileLab
- A Network Is a Kernel Until It MovesLab
- The Gradient Needs a RulerLab
- The Baseline Is a Control VariateLab
- The Search Distribution Learns CoordinatesLab
- Loops Turn Evidence Into RumorLab
- Mistakes Burn CapitalLab
- Do Not Spend the Budget on the Already GoodLab
- Prediction Errors Leave FootprintsLab
- Heatmaps Are Not CircuitsLab
- Your Example Brought Its NeighborsLab
- Step Sizes Are DaresLab
- Vote Gaps Draw the BallLab
- Give Every Candidate a ClockLab
- Matching, but With Soft EdgesLab
- Prompts Are Ratios, Not CommandsLab
- Delete Buttons Owe CounterfactualsLab
- Gradients Keep ReceiptsLab
- Ask the Matrix Where It EchoesLab
- Put Both Rankers on the Same PageLab
- Good Reward Hints Cancel on the LoopLab
- Adam Chooses the Units Before It StepsLab
- SHAP Bars Need a Chosen WorldLab
- LoRA Bets on the Missing DirectionsLab
- Do Not Trust a Naked ECELab
- Privacy Starts With a Volume KnobLab
- Leaderboards Have Seating ChartsLab
- Loss Functions Never See the Missing PairLab
- Context Windows Have Rent DueLab
- Every Trial Spends the BudgetLab
- Models Forget the Weird Stuff FirstLab
- Last Layer, Solver IncludedLab
- Models Overbook Feature SpaceLab
- Build a Copy Machine You Can SeeLab
- Which Zero-Error Model Did You Get?Lab
AI systems22
- The Distance Table Replaces the VectorLab
- A Dataset Can Be a Pull RequestLab
- The Page Is Not a DeveloperLab
- The Watermark Is a Tail TestLab
- The Mask Is a Second ClockLab
- Next FLOP Has to ChooseLab
- Every Token Gets a Clock HandLab
- Small Models Can Be ScoutsLab
- Trees Spend ExperimentsLab
- The Tokenizer Comes With a PhrasebookLab
- Matrices Were Never RequiredLab
- Nearest Neighbor Search Needs RoadsLab
- Every Expert Has a Line OutsideLab
- One Outlier Can Spend the Whole ScaleLab
- Draft Tokens Have to Earn Their KeepLab
- Vectors That Show Up EverywhereLab
- Feeds Teach Users, Then Take NotesLab
- Leaderboards Make Blunt InstrumentsLab
- Saying No Needs a Price ListLab
- Thirty-Two Tries Can Be One MistakeLab
- RAG Is the Question of What Gets ReadLab
- At 3 AM, Ask the CacheLab
Computer science83
- The Cut Survives The ContractionsLab
- The Password Hash Rents MemoryLab
- The Tree Rotates Toward ReuseLab
- The Posting List Splits Its BitsLab
- The Counter Only Remembers the ExponentLab
- The Smallest Counter Takes the BlameLab
- The Alphabet Splits Into BitsLab
- The Line Carries an Error LedgerLab
- The Range Minimum Hides a TreeLab
- The Ready Queue Is a ProofLab
- The Snapshot Draws a CutLab
- The Counterexample Needs a LoopLab
- The Residual Graph Is the Audit TrailLab
- The Queue Has an AreaLab
- The Stack Knows Which Cycle ClosedLab
- The Matching Arrives in WavesLab
- The Diff Keeps a Furthest FrontierLab
- The Puzzle Chooses Its Tightest ColumnLab
- The Diameter Waits For An AntipodeLab
- The Operator Carries Its Binding PowerLab
- The Tree Fits in the ParenthesesLab
- The Minima March RightLab
- The Piles Remember the FrontierLab
- The Search Path Has an AddressLab
- The Cell Has Two DiagonalsLab
- The Transcript Is a Key InputLab
- The Bucket Becomes the QueueLab
- The Timeout Is Not a StopwatchLab
- The Nonce Has One JobLab
- A Gap Is an Event, Not a CharacterLab
- The Run Stack Remembers the OrderLab
- The Parent Pointer Learns a ShortcutLab
- The Fingerprint Has Two HomesLab
- The Conflict Writes a ClauseLab
- The Distance Is an XOR ShadowLab
- The Dictionary Has Failure LinksLab
- The Variable Order Is the AlgorithmLab
- The Signature Spends a Secret OnceLab
- Every Path Has a GatekeeperLab
- The Checksum Is a Polynomial RemainderLab
- The List Grows TowersLab
- The Normal Equations Square the TroubleLab
- The Type Checker Finds the Most General ShapeLab
- The Model Still Owes a FilterLab
- The Prefix Sum Learns to MoveLab
- The Loaded Die Has a TableLab
- The Prefix Sum Is an Address MachineLab
- The Pixel Is an ExpectationLab
- The Bounds Check Was Not a BoundaryLab
- The Filter Peels Before It AnswersLab
- Sketches Shrink Every DirectionLab
- Epsilon Is Not a PurseLab
- Cuckoo Hashing Fails as a GraphLab
- Stop Before the Vector Turns AroundLab
- Last Column Knows Where to SearchLab
- When the Queue Answers Too LateLab
- Search That Refuses to RereadLab
- Smoothing Drew the SpotsLab
- A* Trusts a Promise, Not a HunchLab
- One Transcript Can Be FakeLab
- Where the Half-Bits GoLab
- Clocks Can Say ConcurrentLab
- CPUs Have to Guess Right NowLab
- Let the Clause SleepLab
- Side Deals Are the BugLab
- Tails Come Back Through the Front DoorLab
- Keep the Rewrite Around Until the EndLab
- Bring the Receipt, Not the Whole LogLab
- Hash Tables Are Fast Until the Tail ArrivesLab
- Collisions Keep the ReceiptsLab
- PageRank Reenters Through the Side DoorLab
- Reliable Bits Need Empty SpaceLab
- Formulas Can Lie to the ComputerLab
- Keep the Distances, Audit the DecisionsLab
- Every Eviction Is a PredictionLab
- Clusters Remember by OverlapLab
- Benchmarking Through the OutageLab
- Similar Pages Should Share a Lottery TicketLab
- Bit Arrays That Only Say MaybeLab
- Second Tries Are Still TrafficLab
- Fuzzy Joins Write FactsLab
- Counting Without NamesLab
- Merge Functions Become ProductsLab
Gaming5
Physics6
Applied math8
Computer graphics5
Data systems9
- The Difference Peels Out of the TableLab
- The Row Group Keeps a Bounding BoxLab
- The Page Split Climbs Toward the RootLab
- The Page Must Not Outrun the LogLab
- The Bitmap Is a Container MapLab
- Two Thirds Is Not a VibeLab
- Missing Shards Are EquationsLab
- Do Not Reassign the World to Add One BoxLab
- p95 Needs Its CrowdLab
Software engineering39
- The Hash Join Gets a SpillwayLab
- The Heartbeat Is a Suspicion ScoreLab
- The Rumor Needs a FanoutLab
- The Vacuum Has a HorizonLab
- The Isolation Boundary Needs ConsentLab
- The Request Carries Its ContextLab
- The Cookie Has a Site BoundaryLab
- The Script Has a Guest ListLab
- The Code Needs a ProofLab
- CORS Is Not a FirewallLab
- The Cache Entry Needs a WitnessLab
- The Bucket Has a ClockLab
- The Old Code Still Had a SwitchLab
- The Watermark Is a Promise About the PastLab
- The Triangle Has a Sign BitLab
- The Map Becomes a StringLab
- The Delta Is the QueryLab
- One Shadow Is EnoughLab
- The Sum Has a ScheduleLab
- The Remainder Is a Quotient EstimateLab
- The Plan Is a Row-Count ForecastLab
- The Event Loop Is a QueueLab
- The Username Is Not a StringLab
- The Address Came Back WrongLab
- The Failing Test Case Shrinks by Asking HalvesLab
- Ask Two Queues, Not the Whole FleetLab
- The Random Timeout Is a Vote SpacerLab
- The Instant Has to Fit Inside the CallLab
- The Write Did Not Happen EverywhereLab
- The Query Remembers Only New FactsLab
- TFLOP Numbers Are Not SpeedometersLab
- Leases Expire; Writes Do NotLab
- Stopwatches Are Not CalendarsLab
- Case File: The Heap Was Not FullLab
- Memtables Borrow From TomorrowLab
- Patterns Can Hide Search TreesLab
- Blurry Interpreters Live in Your CompilerLab
- Stores Can Save Live ObjectsLab
- Two Clean Snapshots Can Break the WorldLab
Statistics44
- The Best Candidate Waits For A RecordLab
- The Quantile Keeps Five MarkersLab
- The Estimate Has Two ReceiptsLab
- The Discovery Budget Has a ClockLab
- The Mean Needs a FirebreakLab
- The Error Bar Is CurvatureLab
- The Feature Has a Shadow TwinLab
- The Alarm Has a Run LengthLab
- The Graph Is an Equivalence ClassLab
- The Collider Opens When You LookLab
- Move Two Knobs Or Miss The ModelLab
- The Average Arrives LateLab
- The Counterfactual Is a PortfolioLab
- The Motif Is a Nearest NeighborLab
- The Line Has a Breakdown PointLab
- The Kernel Decides What a Difference IsLab
- The Front Door Opens Through the MiddleLab
- The Likelihood Ratio Is a Running ScoreLab
- The Weight Is a Receipt for Being MissedLab
- The Singleton Points Beyond the SampleLab
- Discoveries Need a DenominatorLab
- First Digits Live on a CircleLab
- Urns Remember Without OrderLab
- Momentum Loans for the PosteriorLab
- Denominators Can Learn the FutureLab
- Thresholds Only Speak NearbyLab
- Samplers Can Pace in the Wrong RoomLab
- Prediction Is a Witness, Not a LeverLab
- Experiments Hide in the ResidualsLab
- Averages Lose Innocence in 3-DLab
- If You Keep Looking, Use a BankrollLab
- AUC Only Sees the PairLab
- Averages Mix Before They JudgeLab
- Ask Who the Instrument MovedLab
- Bootstrap the Assumption, Not the WishLab
- Who Leaves the Curve?Lab
- Sometimes the Posterior Needs a CrowdLab
- Yesterday Can Tighten an A/B TestLab
- How Far Should a Noisy Sensor Move You?Lab
- Control Groups Hold Ghost SlopesLab
- Dashboards Can Change the ExperimentLab
- Straight Lines That Lie PolitelyLab
- Coverage Comes With Fine PrintLab
- Metrics Become Control SurfacesLab
Quant finance21
- The Mid Has a Queue ShadowLab
- The Auction Should TickLab
- The Arbitrage Is a Negative CycleLab
- The Curve Moves in Three ShadowsLab
- The Hedge Is a Variance MeterLab
- Smiles Pay in ButterfliesLab
- Sharpe Ratios Have ClocksLab
- Pager Starts DrummingLab
- Crash Risk Lives in the CornerLab
- Give QMC the Big Move FirstLab
- Prices Forecast With a TabLab
- Same Bid, Different FuturesLab
- Folds Remember Label WindowsLab
- Winning Can Be Bad NewsLab
- Inverses Turn Noise Into BetsLab
- Midpoint as TruceLab
- Alpha Was Hiding in the TimestampLab
- Risk Dials Have DelaysLab
- After the Decision, the Trade BeginsLab
- Beautiful Backtests Come From a CrowdLab
- Trust First, Then Size the BetLab
Psychology8
Meta1
Probabilistic data structures10
Sketches, filters, and hash tables that trade a little accuracy for a lot of space.
- Counting Without NamesLab
- Bit Arrays That Only Say MaybeLab
- Similar Pages Should Share a Lottery TicketLab
- Collisions Keep the ReceiptsLab
- Hash Tables Are Fast Until the Tail ArrivesLab
- Cuckoo Hashing Fails as a GraphLab
- The Filter Peels Before It AnswersLab
- The Model Still Owes a FilterLab
- The Fingerprint Has Two HomesLab
- The Difference Peels Out of the TableLab
Trustworthy experiments9
How A/B tests and metrics quietly mislead, and what keeps them honest.
- Metrics Become Control SurfacesLab
- Dashboards Can Change the ExperimentLab
- Yesterday Can Tighten an A/B TestLab
- If You Keep Looking, Use a BankrollLab
- Discoveries Need a DenominatorLab
- The Feature Has a Shadow TwinLab
- The Cursor Needs an AddressLab
- The Discovery Budget Has a ClockLab
- The Estimate Has Two ReceiptsLab
Faster LLM inference6
The systems tricks that make large models cheap enough to serve.
Quantitative trading8
Sizing, executing, and stress-testing a strategy without fooling yourself.
Calibrated confidence4
Making a model's stated uncertainty mean what it actually says.
Distributed systems9
Time, consistency, and failure when one machine is no longer enough.
- Two Clean Snapshots Can Break the WorldLab
- Clocks Can Say ConcurrentLab
- Do Not Reassign the World to Add One BoxLab
- Missing Shards Are EquationsLab
- Leases Expire; Writes Do NotLab
- Two Thirds Is Not a VibeLab
- The Snapshot Draws a CutLab
- The Rumor Needs a FanoutLab
- The Heartbeat Is a Suspicion ScoreLab
- The Cut Survives The ContractionsKarger's min-cut algorithm is a small Monte Carlo argument: a minimum cut is found if none of its edges are contracted. A brute-force oracle and si...
- The Password Hash Rents MemoryA password hash is not just a slow hash. Memory-hard schemes make offline guessing rent RAM, burn bandwidth, carry per-user salts, and fit inside a...
- The Hash Join Gets a SpillwayA hash join larger than memory survives by partitioning both inputs into matching batches. Spilling is not a side effect; it is the algorithm's esc...
- The Tree Rotates Toward ReuseA splay tree keeps no balance bits. It pays rotations after each access so the next access can be cheaper if the trace has locality.
- The Difference Peels Out of the TableAn invertible Bloom lookup table lets two large, similar sets subtract their summaries and recover only the keys that differ by peeling pure cells.
- The Posting List Splits Its BitsElias-Fano encoding stores a monotone integer list by writing low bits literally and turning high bits into a selectable unary bitmap.
- The Counter Only Remembers the ExponentMorris approximate counting trades exact increments for a tiny logarithmic register, with an unbiased estimator and a variance bill you can compute.
- The Smallest Counter Takes the BlameSpace-Saving finds stream heavy hitters by replacing the smallest counter and carrying its value forward as an explicit error receipt.
- The Alphabet Splits Into BitsA wavelet tree makes rank, select, access, and range quantile queries by recursively turning a large alphabet into stable bitvector decisions.
- The Best Candidate Waits For A RecordThe secretary problem is not really about hiring. It is a finite optimal-stopping contract: spend a sample, then accept the first record, and verif...
- The Quantile Keeps Five MarkersThe P2 algorithm estimates one streaming quantile with five moving markers. Its strength is fixed memory; its price is that the error is empirical,...
- The Line Carries an Error LedgerBresenham line drawing is not just old graphics lore. It is an incremental proof that one integer error term can choose adjacent pixels without mul...
- The Range Minimum Hides a TreeA Cartesian tree turns array order into inorder traversal and range minima into lowest common ancestors. The monotone stack is the linear-time rece...
- The Ready Queue Is a ProofTopological sorting is not just a way to print a build order. The zero-indegree queue is a local certificate of legality, and the critical path say...
- The Row Group Keeps a Bounding BoxZone maps and data-skipping indexes do not find rows. They prove that whole row groups cannot match a predicate, and their power comes from physica...
- The Heartbeat Is a Suspicion ScorePhi-accrual failure detectors do not turn silence into an immediate yes/no verdict. They convert heartbeat delay into a tail probability, then let ...
- The Rumor Needs a FanoutEpidemic dissemination turns broadcast into repeated random contacts. Push spreads quickly at the beginning, pull cleans up the tail, and push-pull...
- The Snapshot Draws a CutChandy-Lamport snapshots record a meaningful global state without stopping a distributed system. The trick is to record both local process state an...
- The Counterexample Needs a LoopSafety bugs can stop at a bad prefix. Liveness bugs need an infinite witness, so finite-state model checkers hand you a prefix plus a loop.
- The Page Split Climbs Toward the RootA B+ tree index stays shallow because each node is a page-sized roster of keys. Insertions are local until a full page splits, promotes a separator...
- The Residual Graph Is the Audit TrailMax-flow algorithms do not just push capacity forward. The residual graph records the unused lanes and the undo lanes, and the final unreachable se...
- The Vacuum Has a HorizonMVCC makes reads and writes pass each other by keeping old row versions. Vacuum can clean only behind the oldest snapshot that might still see them.
- The Queue Has an AreaLittle's Law is not a magic queueing formula. It is an accounting identity over the right observation window, and cropped logs can quietly lose the...
- The Estimate Has Two ReceiptsA doubly robust estimator checks an observational causal estimate against both the treatment assignment model and the outcome model. One correct re...
- The Discovery Budget Has a ClockOnline false discovery control is not batch FDR with a refresh button. A stream of tests needs thresholds that are chosen before the next p-value a...
- The Isolation Boundary Needs ConsentCross-origin isolation is not one magic header. COOP isolates the opener relationship, COEP sets an embedder contract, and CORP or CORS lets resour...
- The Request Carries Its ContextFetch Metadata headers let a server classify the browser shape of a request before application code spends work or trusts ambient cookies.
- The Cookie Has a Site BoundarySameSite is not a domain flag or a CORS setting. It is a browser retrieval rule for deciding whether an ambient cookie rides along with a request.
- The Script Has a Guest ListContent Security Policy for scripts is not one rule. Host allowlists, nonces, hashes, and strict-dynamic answer different questions about which Jav...
- The Code Needs a ProofPKCE turns an OAuth authorization code from a bearer artifact into a two-step receipt: the front channel carries a challenge, and the token endpoin...
- CORS Is Not a FirewallCross-Origin Resource Sharing is a browser response-sharing protocol. Preflight requests ask whether a script may send a shaped request; the final ...
- The Cache Entry Needs a WitnessHTTP caching is not only a TTL. Freshness says when a stored response may be reused without asking; validators say how to ask the origin whether th...
- The Bucket Has a ClockA token bucket is not a vague request quota. It is a small arrival-curve contract: spend saved credit on bursts, refill at a steady rate, and prove...
- The Stack Knows Which Cycle ClosedTarjan's SCC algorithm turns one depth-first search into a component partition by keeping lowlink values and popping only when an open cycle can no...
- The Matching Arrives in WavesHopcroft-Karp speeds up bipartite maximum matching by batching a maximal set of shortest augmenting paths instead of repairing the matching one pat...
- The Diff Keeps a Furthest FrontierMyers' diff algorithm turns longest common subsequence into a shortest path problem, then searches by remembering only the furthest point reached o...
- The Puzzle Chooses Its Tightest ColumnAlgorithm X is backtracking with an accounting discipline: express the puzzle as exact cover, branch on the most constrained uncovered column, and ...
- The Diameter Waits For An AntipodeThe farthest pair in a planar point set is not hiding in the interior. Build the convex hull, rotate a support line around it, and the diameter app...
- The Operator Carries Its Binding PowerPratt parsing turns expression grammar into a table of token behaviors: each token knows how to start an expression, how to extend one, and how tig...
- The Tree Fits in the ParenthesesSuccinct tree representations separate topology from navigation: a rooted ordered tree can live in a balanced-parentheses bitstring, while rank, se...
- The Mean Needs a FirebreakMedian-of-means estimates a mean by splitting samples into blocks, averaging locally, and letting the median discard a few damaged blocks.
- The Mid Has a Queue ShadowThe midprice ignores how much size is waiting at the best bid and ask. A one-step microprice treats the next quote move as a queue-depletion race.
- The Minima March RightSMAWK searches a totally monotone matrix without materializing it. The row minima are not independent; total monotonicity lets one comparison disca...
- The Piles Remember the FrontierPatience sorting computes a longest increasing subsequence by keeping one small tail value for every possible length. The piles are not the answer;...
- The Integrator Kept ChargingA PI controller with a saturated actuator can keep accumulating error after more command can no longer help. Anti-windup is bookkeeping for that hi...
- The Search Path Has an AddressA binary search tree can do the same comparisons and still touch very different memory blocks. A cache-oblivious layout changes the address sequenc...
- The Cell Has Two DiagonalsMarching squares turns a sampled scalar field into contour segments with a 16-case table. Two cases are ambiguous because the corners alone do not ...
- The Transcript Is a Key InputTLS 1.3 does not only derive keys from Diffie-Hellman bytes. It feeds the handshake transcript into HKDF labels, CertificateVerify, and Finished so...
- The Bucket Becomes the QueueHash tables earn constant-time behavior from a distributional assumption. Hash-flooding attacks spend chosen input to break that assumption and tur...
- The Timeout Is Not a StopwatchTCP's retransmission timeout is a guarded suspicion about silence: estimate the path, price the variance, and refuse ACK samples whose origin is am...
- The Nonce Has One JobAES-GCM is fast because each nonce/key pair creates a one-time encryption stream and a one-time authentication mask. Reuse the nonce and both contr...
- A Gap Is an Event, Not a CharacterNeedleman-Wunsch alignment is a dynamic program over possible paths through a grid. Gotoh's affine-gap version adds state so one long insertion or ...
- The Run Stack Remembers the OrderTimsort is often summarized as mergesort plus insertion sort. The more interesting idea is that it treats existing order as data: find natural runs...
- The Parent Pointer Learns a ShortcutUnion-find keeps a changing partition by storing parent pointers. Union by rank keeps the trees shallow; path compression makes every search repair...
- The Error Bar Is CurvatureFisher information turns local likelihood curvature into a lower bound on estimator variance. The Cramer-Rao bound is a geometry statement before i...
- The Old Code Still Had a SwitchKnight Capital's 2012 loss was not just a bad deploy. It was stale production code, a reused flag, missing router-level controls, and a live market...
- The Texture Needs a Smaller CopyMipmaps are prefiltered promises about pixel footprint: when one screen pixel covers many texels, sample a smaller copy instead of one lucky texel.
- The Page Must Not Outrun the LogWrite-ahead logging makes durability a recoverable story: before a dirty page reaches disk, the log record that explains it must already be durable.
- The Ruler Is a RatioWeber's law, Fechner's logarithmic scale, and Stevens' power law are not three interchangeable slogans. They are different ways to turn physical in...
- The Image Pays One Pixel at a TimeSeam carving turns resizing into a shortest-path problem: remove one low-energy connected pixel path, recompute the image, and audit what the resiz...
- The Fingerprint Has Two HomesA cuckoo filter keeps only short fingerprints, so deletion works because each fingerprint still carries enough information to find its other bucket.
- The Square Root Was in the ExponentQuake III's famous fast inverse square root is not dark magic. It is a logarithm hiding in IEEE float bits, followed by one Newton repair step.
- The Auction Should TickA continuous limit order book can turn public information into a private speed prize. Frequent batch auctions change the question from who arrived ...
- The Conflict Writes a ClauseCDCL solvers do not just backtrack when a SAT branch fails. They turn the failure into a resolvent, jump past irrelevant decisions, and make the sa...
- The Cursor Needs an AddressCollaborative text editing is not just merging strings. Every inserted character needs a stable identity, and the wrong address scheme can make two...
- The Feature Has a Shadow TwinKnockoff filters control false discoveries by making every feature race a synthetic twin that has the same correlation structure but no direct clai...
- The Alarm Has a Run LengthCUSUM and Page-Hinkley detectors are not magic drift bells. They are stopping rules that spend false-alarm budget to buy detection delay.
- The Distance Table Replaces the VectorProduct quantization makes vector search fast by turning every stored embedding into a handful of small codebook IDs. The query stays real; the dat...
- The Distance Is an XOR ShadowKademlia routes through a distributed hash table by treating closeness as XOR distance. That turns lookup into a prefix game with k-buckets, stale ...
- The Arbitrage Is a Negative CycleCurrency arbitrage becomes a graph problem after one logarithm. Multiplicative exchange rates turn into additive path weights, and a profitable rou...
- The Watermark Is a Promise About the PastEvent-time stream processing is a negotiation between correctness and latency. A watermark is the system's current promise about which old events a...
- The Dictionary Has Failure LinksAho-Corasick turns many string patterns into one automaton. The trick is not only a trie; it is knowing where to fall when a partial match breaks.
- The Interior Path Keeps Its DistanceInterior-point methods solve constrained optimization by making every wall expensive before touching it. A small LP lab follows the central path an...
- The Variable Order Is the AlgorithmReduced ordered binary decision diagrams are canonical only after you choose an order. The wrong order can make a small Boolean function look enorm...
- The Triangle Has a Sign BitThe orientation predicate looks like a tiny determinant. In computational geometry, that sign bit is a topology decision, and ordinary floating-poi...
- The Graph Forgets the TrianglesMessage-passing graph neural networks inherit a precise blind spot from color refinement. The 1-WL test can be strong, but it can also stare at two...
- The Graph Is an Equivalence ClassObservational causal discovery usually cannot identify one directed graph. Markov equivalence says the data sees a skeleton and unshielded collider...
- The Signature Spends a Secret OnceIn DSA-style signatures, the per-message nonce is not harmless randomness. Reuse it, shrink it, or put it on a clock, and the private key turns int...
- The Bin Is a FilterA DFT bin is not a tiny bucket that catches one exact frequency. It is a window-shaped filter, and spectral leakage is what happens when the tone f...
- The Collider Opens When You LookD-separation is the graph rule behind conditional independence in Bayesian networks: chains and forks close when you condition, but colliders open ...
- Every Path Has a GatekeeperDominator trees turn control flow into a question a compiler can answer: which blocks must every path pass through before reaching this point?
- The Map Becomes a StringA spatial index is often a strange act of compression: take a rectangle on a grid, turn it into a few one-dimensional key ranges, and hope locality...
- The Max Has a TemperatureEntropy-regularized control does not merely add random exploration after planning. It changes the Bellman backup itself: max becomes log-sum-exp.
- The Tree Gets More Lottery TicketsGreedy tree splits compare the best cut each feature can find. A feature with many possible cuts gets more chances to look important by accident.
- The Delta Is the QueryIncremental view maintenance works when change is smaller than state. The hard part is choosing which partial answers deserve to be remembered.
- One Shadow Is EnoughA collision detector does not need to understand two convex shapes all at once. If one projection leaves daylight between them, the case is closed.
- Move Two Knobs Or Miss The ModelOne-at-a-time sensitivity analysis walks along coordinate axes. Sobol indices ask how much output variance each input owns, including the variance ...
- The Path Is Not Its MarginalsViterbi decoding and posterior decoding both look through the same HMM trellis, but they optimize different losses. The most likely story is not al...
- The Menu Has EntropyHick-Hyman law is not a bumper sticker for removing choices. It is a small information-theoretic model of choice reaction time, and it gets useful ...
- The Cut Has a DenominatorMinimum cuts like cheap separations. Normalized cuts ask whether the separated set had enough internal graph volume to deserve being called a cluster.
- The Average Arrives LateRobbins-Monro stochastic approximation learns a root from noisy experiments. The step size has to keep moving long enough to find the target, then ...
- The Record Has No FieldsHyperdimensional computing stores symbols as wide random vectors. Binding and bundling let a record live in one vector, but cleanup only works whil...
- The Checksum Is a Polynomial RemainderA CRC is not a small hash pasted onto a packet. It is polynomial division over GF(2), and an error is missed exactly when the error polynomial is d...
- The Memory Chooses Its DecayMamba-style state-space models are recurrent, but the important move is selectivity: the update rule can decide which tokens write, which tokens fa...
- The List Grows TowersSkip lists replace tree rotations with coin flips. The trick works because a random tower of forward pointers gives logarithmic search in expectati...
- The Curve Moves in Three ShadowsYield curves are high-dimensional objects, but many daily moves look like mixtures of level, slope, and curvature. PCA makes that visible; risk sti...
- The Counterfactual Is a PortfolioSynthetic control builds a missing no-treatment path as a weighted portfolio of untreated units. The weights are transparent; the causal claim is n...
- The Motif Is a Nearest NeighborMatrix profiles turn time-series motif discovery into a nearest-neighbor ledger over sliding windows. The smallest entry finds repeated shape; the ...
- The Kernel Keeps a Few WitnessesNystrom kernel approximation replaces a full Gram matrix with a few landmark columns. The hard part is not the formula; it is choosing which witnes...
- The Posterior Pays RentPAC-Bayes bounds let the learned distribution depend on the data, but charge it KL divergence from a data-independent prior.
- The Normal Equations Square the TroubleLeast squares is one objective but not one computation. Forming A^T A can square the condition number; QR keeps the geometry closer to the data.
- The Label Is a Latent VariableMajority vote treats annotators as interchangeable sensors. Dawid-Skene style truth inference treats the true label and each worker's confusion mat...
- The Line Has a Breakdown PointLeast squares is a voting rule with a very loud microphone for far-away points. Robust regression is not one trick; Huber, Theil-Sen, and RANSAC fa...
- The Sum Has a ScheduleFloating-point addition is commutative but not associative. Once a sum is parallelized, shuffled, vectorized, or compensated, the reduction order b...
- The Remainder Is a Quotient EstimateBarrett and Montgomery reduction replace division by a fixed modulus with multiplication, shifts, and a small correction loop. The trick is not avo...
- The Type Checker Finds the Most General ShapeHindley-Milner inference feels like magic because it returns a principal type. Under the hood it is a small equation solver with one crucial habit:...
- The Plan Is a Row-Count ForecastA SQL optimizer does not merely choose joins. It makes a forecast about intermediate rows, then spends real I/O and CPU as if that forecast were true.
- The Event Loop Is a QueueAsync I/O removes waiting, not work. A single slow callback can turn a calm service into a tail-latency machine unless the queue is visible and bou...
- The Model Still Owes a FilterLearned Bloom filters are not classifiers with nicer marketing. The model may save bits, but the one-sided membership guarantee still has to be pai...
- The Cache Forgets the MiddleKV-cache eviction is a lossy attention policy. Sliding windows, attention sinks, heavy hitters, and observation-window voting each make a different...
- The Kernel Decides What a Difference IsMaximum mean discrepancy turns a vague distribution-shift question into a precise measurement. The catch is that the kernel is not decoration: it d...
- The Username Is Not a StringUnicode lets identifiers speak many languages, but raw string equality is the wrong security boundary. Normalization, case folding, script policy, ...
- The Object Has More Than One DoorGFlowNets are easiest to misunderstand if you think they optimize for the single best object. Their stranger job is to sample objects in proportion...
- The Rounder Is the TurnstileLow-precision training is not just a bit-width choice. Once updates are smaller than a grid step, the gate rule decides whether they are blocked or...
- The Address Came Back WrongThe ABA problem is what happens when compare-and-swap sees the same pointer twice and mistakes equality for history.
- The Edge Pixel Is a ContractAlpha compositing works when color, coverage, and color space agree. Premultiplied alpha is the contract that keeps filtered edges from lying.
- The One Percent Chance Has a PriceCumulative prospect theory is a model of how risky choices can be priced relative to a reference point, not by expected value alone.
- The Failing Test Case Shrinks by Asking HalvesDelta debugging turns a huge failing input into a small witness, but its promise is 1-minimality under a test oracle, not omniscience.
- The Front Door Opens Through the MiddleThe front-door criterion can identify a causal effect with hidden treatment-outcome confounding, but only by using a mediator in a very particular ...
- A Network Is a Kernel Until It MovesThe neural tangent kernel turns training into a fixed linear model in the infinite-width limit. A small lab shows what that approximation looks lik...
- The Bitmap Is a Container MapRoaring bitmaps compress integer sets by cutting the universe into 16-bit chunks and choosing a local container for each chunk's shape.
- The Prefix Sum Learns to MoveA Fenwick tree keeps an editable cumulative frequency table by storing suffix blocks whose widths are written in the low bits of their indexes.
- Ask Two Queues, Not the Whole FleetThe power-of-two-choices result says a load balancer can crush the maximum-load tail by comparing two random servers instead of trusting one random...
- The Loaded Die Has a TableThe alias method turns a fixed categorical distribution into equal-width boxes, so a draw needs one random column and one threshold comparison.
- The Gradient Needs a RulerA gradient is not a direction until a geometry says what a unit step means. On the probability simplex, the wrong ruler can turn a distribution fit...
- The Baseline Is a Control VariateIn policy gradients, a baseline is allowed because the score has zero mean. The best baseline is not always the average reward.
- The Likelihood Ratio Is a Running ScoreA sequential test is not permission to keep refreshing a p-value. It is a pre-committed stopping rule whose state is the accumulated log evidence.
- The Random Timeout Is a Vote SpacerRaft's election timeout is not just a wait. It is a randomized spacing device that makes split votes rare without making safety depend on clocks.
- The Prefix Sum Is an Address MachineParallel scan looks like a cumulative sum, but its real power is turning local facts into global addresses with logarithmic synchronization.
- The Weight Is a Receipt for Being MissedHorvitz-Thompson weighting turns unequal sampling chances into a finite-population estimator, but the variance bill arrives in the weights.
- The Instant Has to Fit Inside the CallLinearizability asks whether every concurrent operation can be assigned one legal instant between invocation and response.
- The Write Did Not Happen EverywhereA tiny store-buffering litmus test shows why memory order is a contract, not a timestamp.
- The Singleton Points Beyond the SampleGood-Turing and unseen-species estimators turn one-off observations into evidence about the probability mass you have not seen yet.
- The Pixel Is an ExpectationPath tracing works because a pixel can be written as an integral, and a ray can be treated as a random variable.
- The Hedge Is a Variance MeterA delta-hedged option does not merely bet on direction. It keeps a ledger of realized variance, rebalancing error, jumps, and transaction costs.
- A Dataset Can Be a Pull RequestData poisoning is what happens when training examples become a write path into model behavior.
- The Bounds Check Was Not a BoundarySpectre-style attacks broke a quiet assumption: discarded speculative work can still leave measurable traces in microarchitectural state.
- The Page Is Not a DeveloperPrompt injection is a control-plane bug in LLM applications: untrusted text is allowed to compete with trusted instructions, tools, and secrets.
- The Watermark Is a Tail TestLLM text watermarks are not visible signatures. They are statistical evidence accumulated token by token, and their reliability lives in entropy, l...
- The Mask Is a Second ClockDiscrete diffusion language models make generation a schedule over unknown tokens, not just a left-to-right stream. The hard part is deciding what ...
- The Query Remembers Only New FactsRecursive Datalog queries are fixed points over relations. Semi-naive evaluation makes the fixed point fast by propagating only the facts that were...
- The Error Changes GridsMultigrid is fast because it does not ask one iteration to see every scale. It smooths oscillatory error, moves the remaining error to a coarser gr...
- The Filter Peels Before It AnswersXor filters look like a three-probe membership test. The construction is stranger: build a random hypergraph, peel it, then solve a sparse XOR syst...
- The Search Distribution Learns CoordinatesCMA-ES is not magic mutation. It is a rank-based optimizer that lets a Gaussian search distribution learn the valley it is walking through.
- Two Thirds Is Not a VibeByzantine quorum certificates work because two possible histories are forced to share an honest voter, not because 67% has mystical distributed-sys...
- Discoveries Need a DenominatorFalse discovery rate control is not a ritual after p-values. It is a contract about the list you are willing to publish.
- Pixels Can Keep One LightReservoir resampling turns a stream of many possible light samples into one carried sample plus a weight ledger. The magic is not the sample; it is...
- Waiting for the Wrong FluctuationKramers escape explains why a system can look settled while rare noise quietly prices the exit.
- Grids Have Speed LimitsThe CFL condition is not a magic stability spell. It is a geometric warning: a time step cannot outrun the stencil that is supposed to carry inform...
- Loops Turn Evidence Into RumorBelief propagation is exact on trees because every message summarizes disjoint evidence. On a loopy graph, the same local rule can double-count its...
- Controllers Solve BackwardFinite-horizon LQR turns a future cost into a present feedback law. The Riccati recursion runs backward; the controller runs forward.
- Holes Have to PersistPersistent homology turns a point cloud into a scale experiment. Components and loops are allowed to appear and die; the barcode records which ones...
- Fixed Points Remember MistakesAnderson acceleration turns a slow fixed-point iteration into a small residual-canceling least-squares problem. The magic is real, but only if you ...
- Sketches Shrink Every DirectionFrequent Directions is a deterministic streaming sketch for covariance. It keeps a tiny matrix, repeatedly shrinks the spectrum, and still promises...
- Normalizers Hide in the PathAnnealed importance sampling estimates a partition function by walking through easier distributions and averaging the resulting path weights.
- Epsilon Is Not a PurseDifferential privacy is easier to reason about when epsilon is a likelihood-ratio threshold, delta is tail slack, and the accountant keeps the whol...
- Critical Points RememberAt the Ising critical point, local Monte Carlo updates keep revisiting the same large domains. A cluster move changes the unit of motion.
- TFLOP Numbers Are Not SpeedometersA processor's peak FLOP rate is only one roof. The other roof is memory bandwidth, and the useful question is how much arithmetic each byte gets to...
- Eigenvalues Are Not WeatherA stable matrix can amplify disturbances before it damps them. The missing diagnostic is not another eigenvalue, but the geometry of the eigenvecto...
- First Digits Live on a CircleBenford's law is not a numerology trick about 1s. It is a statement about logarithmic mantissas, scale invariance, and when a dataset has no prefer...
- Urns Remember Without OrderA Polya urn is reinforced and path-dependent, but still exchangeable. The order of the draws can disappear even when early luck never does.
- Cuckoo Hashing Fails as a GraphThe basic cuckoo table does not fail because the insertion loop is unlucky. It fails when the hidden bipartite graph contains too many edges for it...
- Flocking Is an Order ParameterA flock does not need a leader. In the Vicsek model, local alignment, finite speed, and angular noise are enough to create a measurable transition ...
- CG Buys One More DegreeConjugate gradient is not just gradient descent with a better memory. Each iteration buys another degree of polynomial freedom to suppress the eige...
- Stop Before the Vector Turns AroundA two-amplitude Grover lab for quantum search, amplitude amplification, overshoot, phase calibration, and the square-root limit of black-box search.
- Random Dots Need Personal SpaceA graphics and games lab for comparing random points, jittered grids, best-candidate sampling, and Bridson Poisson disks with gaps, voids, and spec...
- Momentum Loans for the PosteriorHamiltonian Monte Carlo proposes by moving like a physical system: conserve energy, preserve volume, then let Metropolis audit the numerical error.
- Ask Random Vectors for the TraceHutchinson's estimator turns random vectors into a matrix-free trace meter. The trick is unbiased; the hard part is controlling variance when the s...
- Mistakes Burn CapitalMultiplicative weights is a loss ledger with teeth: every expert keeps capital, bad rounds shrink it, and the learner gets a certificate against a ...
- Next Reaction Wins the RaceGillespie's stochastic simulation algorithm samples chemical reaction paths exactly by letting propensities compete as exponential clocks.
- Cliffs Live at InfinityPercolation has a clean infinite-lattice threshold, but every experiment sees a rounded finite box. The browser can show the cliff only by watching...
- Last Column Knows Where to SearchThe FM-index turns the Burrows-Wheeler transform into a searchable compressed index: exact matching becomes a right-to-left interval-narrowing game.
- Every Tile Starts as a MaybeWave Function Collapse looks like procedural magic, but the useful version is a constraint solver with a texture memory: local patterns, entropy ch...
- Buttons Make Promises to HandsFitts' law treats pointing as information transmission: distance, width, time pressure, and endpoint scatter decide whether a target is actually us...
- Next FLOP Has to ChooseScaling laws are not just a slogan about bigger models. Under a fixed training budget, parameters and tokens trade off, and data scarcity bends the...
- Leases Expire; Writes Do NotLease expiry is a time-based promise, but stale side effects need an epoch check at the resource. Clock padding and fencing tokens solve different ...
- When the Queue Answers Too LateTCP congestion control is a delayed conversation: the congestion window probes, the bottleneck replies with loss or delay, and a fat buffer can mak...
- Game Trees Tell the Wrong StoryAlpha-beta search is drawn as a tree, but repeated positions make the real work a graph. A transposition table is safe only when it stores identity...
- Stopwatches Are Not CalendarsWall clocks name calendar instants; monotonic clocks measure elapsed time. Mixing them turns NTP corrections, clock steps, and lease expiries into ...
- Smiles Pay in ButterfliesA volatility smile is not just a curve of Black-Scholes inputs. Through call-price curvature, it encodes a risk-neutral density, and the butterfly ...
- Search That Refuses to RereadKMP string search is linear because a mismatch leaves evidence behind. The failure table remembers which prefix is also a suffix.
- Do Not Spend the Budget on the Already GoodA treatment can be bad on average and still worth targeting. CATE and uplift models ask who changes because of treatment, not who would have done w...
- Denominators Can Learn the FutureImmortal time bias is a time-zero bug in causal survival analysis. A treatment can look protective because the treated group had to survive long en...
- Prediction Errors Leave FootprintsTD(lambda) gives the next surprise a trail to follow, so delayed reward can reach the states that made it likely.
- Heatmaps Are Not CircuitsActivation patching can be a real causal intervention, but a restored logit is evidence about a contrast, a metric, and a site, not an explanation ...
- Your Example Brought Its NeighborsBatch normalization is stateful and contextual: in training mode each activation is normalized by the mini-batch around it, then eval mode switches...
- Step Sizes Are DaresA learning rate is not just speed. It is a numerical stability promise about the sharpest direction the optimizer is willing to survive.
- Thresholds Only Speak NearbyRegression discontinuity turns a hard rule into a narrow causal comparison. The useful question is not whether the rule is sharp, but whether anyth...
- Samplers Can Pace in the Wrong RoomA Metropolis chain can accept most proposals and still fail to explore. Convergence is a claim about what the chain has forgotten, not how busy it ...
- Sharpe Ratios Have ClocksAnnualizing a Sharpe ratio by sqrt(252) is not a law of nature. It is an assumption about the sampling clock, the autocovariances, and how returns ...
- Vote Gaps Draw the BallRandomized smoothing turns a classifier into a Gaussian vote. When the top class wins by enough, the prediction earns a certified L2 ball.
- Prediction Is a Witness, Not a LeverGranger causality is incremental forecasting. It becomes causal evidence only after the conditioning set, lag order, and measurement process surviv...
- Give Every Candidate a ClockThe Gumbel-max trick turns a categorical draw into a race: add one random key to every log weight, take the winner, or sort the same keys for sampl...
- Every Token Gets a Clock HandRotary position embeddings do not paste a position vector onto a token. They rotate query and key coordinates so attention reads relative displacem...
- Pager Starts DrummingA Hawkes process notices when an event is also fuel for the next event. The mean rate can look harmless while the clock is carrying aftershocks.
- Missing Shards Are EquationsReed-Solomon erasure coding is not backup magic. It is the promise that any k honest surviving shards contain enough GF(256) equations to rebuild t...
- Crash Risk Lives in the CornerA portfolio can keep the same one-name loss distributions and broad rank association while changing its joint crash probability. The missing object...
- Smoothing Drew the SpotsReaction-diffusion systems bend intuition: local mixing plus nonlinear reactions can turn a nearly uniform field into spots, worms, fronts, and decay.
- Matching, but With Soft EdgesEntropic optimal transport turns a hard matching problem into matrix scaling. Sinkhorn repeats row and column normalization until a soft transport ...
- Prompts Are Ratios, Not CommandsClassifier-free guidance is not magic obedience. In the score idealization, the slider raises a conditional-vs-unconditional density ratio and buys...
- Case File: The Heap Was Not FullHigh RSS is not always a leak. Fragmentation, size classes, phase changes, and thread-local caches can make the allocator's ledger disagree with li...
- Delete Buttons Owe CounterfactualsMachine unlearning is not just deleting a row. It is the demand that a deployed model behave as if that row had never entered training.
- Do Not Reassign the World to Add One BoxConsistent hashing is mostly a remapping contract: when the fleet changes, only the keys that genuinely belong on the new owner should move.
- A* Trusts a Promise, Not a HunchA* is fast when its heuristic makes a mathematical promise about the remaining path cost. Break that promise and the speedup becomes an explicit bet.
- Memtables Borrow From TomorrowAn LSM tree makes writes cheap by turning random updates into sorted runs. Compaction is where the sorting loan comes due.
- One Transcript Can Be FakeA zero-knowledge proof is not a magic blackout curtain. In a Sigma protocol, a fake transcript can look real, while two forked answers reveal the w...
- Where the Half-Bits GoEntropy coding is the quiet accounting beneath compression: the model prices each symbol, and the coder pools those fractional prices into a revers...
- Clocks Can Say ConcurrentLamport clocks preserve causal order, but scalar timestamps still sort concurrent events. Vector clocks keep the partial order visible instead of i...
- CPUs Have to Guess Right NowBranch prediction is online learning at the front of the processor: tiny adaptive models guess control flow before the real label arrives.
- Let the Clause SleepModern SAT solvers are fast partly because most clauses stay dormant. Two watched literals turn unit propagation into a lazy alarm system for memor...
- Small Models Can Be ScoutsSpeculative decoding is exact because the draft model is only a proposal distribution. The verifier keeps the overlap and samples the missing targe...
- Side Deals Are the BugDeferred acceptance is not a sorting algorithm. It is a way to remove every mutually tempting off-book deal while choosing one endpoint of the stab...
- p95 Needs Its CrowdA p95 is a rank statement about a population, not a number you can average across shards. Sketches work because they keep rank evidence alive.
- Tails Come Back Through the Front DoorFFT convolution is fast, but the finite DFT lives on a circle. Without enough zero padding, the end of the answer wraps around and pollutes the beg...
- Gradients Keep ReceiptsReverse-mode automatic differentiation is not numerical differencing and not symbolic algebra. It is a transposed execution of the program, and the...
- Ask the Matrix Where It EchoesRandomized SVD works because a matrix can be interrogated by random probes. The probes do not inspect every entry; they listen for the subspace whe...
- Give QMC the Big Move FirstQuasi-Monte Carlo option pricing depends on the map from uniforms to paths. A Brownian bridge keeps the Brownian law but moves coarse path variance...
- Prices Forecast With a TabPrediction markets and proper scoring rules are the same incentive idea wearing different interfaces. LMSR turns a log score into an always-on mark...
- Trees Spend ExperimentsMonte Carlo Tree Search is not just lookahead. It is an adaptive sampling rule: spend simulations where uncertainty and upside are still worth payi...
- The Tokenizer Comes With a PhrasebookA BPE merge table is a learned phrasebook from a reference corpus. It gives familiar strings short names and leaves unfamiliar strings to spell the...
- Ladders Need Error BarsElo-like ratings are not prizes. They are online estimates of latent skill from noisy pairwise comparisons, and the dangerous omission is uncertainty.
- Games Guess the PastRollback netcode does not make packets faster. It lets online games run now, predict missing input, and repair the timeline when reality arrives.
- Rare Targets Teach You to QuitLow-prevalence visual search is not ordinary perception with fewer positives. Rarity changes item criteria, quitting thresholds, and mistakes.
- Matrices Were Never RequiredFlashAttention is exact attention scheduled differently: keep the running max, rescale the denominator, and avoid writing the full attention matrix...
- Same Bid, Different FuturesIn a limit order book, two orders at the same price are not the same trade. FIFO priority changes fill probability, adverse selection, and the valu...
- Patterns Can Hide Search TreesA regex can look like a declarative shape while the engine underneath is doing a depth-first search through guesses.
- Experiments Hide in the ResidualsDouble machine learning is not a confounding eraser. It builds a residual score so nuisance-model mistakes have less leverage over the causal number.
- Nearest Neighbor Search Needs RoadsHNSW-style vector search works because the index is navigable. Local edges give precision, long edges give reach, and the search beam is the fuel b...
- Put Both Rankers on the Same PageA/B tests compare rankers through different users and queries. Interleaving makes them share one result page, turning ranking evaluation into a pai...
- Folds Remember Label WindowsRandom cross-validation can leak through time even when every feature is legal. The culprit is often the label interval, not an obviously forbidden...
- Review Queues Spend MemorySpaced repetition is not a productivity spell. It is a control problem over memories whose recall probabilities decay at different rates.
- Blurry Interpreters Live in Your CompilerStatic analysis runs the program in a smaller universe: types, signs, intervals, shapes, effects, or any abstraction cheap enough to compute and so...
- Averages Lose Innocence in 3-DStein's paradox says the coordinatewise sample mean can be beaten uniformly once you estimate three or more normal means at the same time.
- Keep the Rewrite Around Until the EndEquality saturation turns compiler optimization from a brittle sequence of rewrites into a search over equivalent programs.
- Good Reward Hints Cancel on the LoopPotential-based reward shaping can make reinforcement learning easier without changing the task. Arbitrary hints can quietly create a new game.
- Stores Can Save Live ObjectsA garbage collector is a graph traversal racing a mutating graph. The write barrier is the small memory of a store that keeps the traversal true.
- Bring the Receipt, Not the Whole LogA Merkle proof does not say trust me. It says this exact item belongs to this exact committed root, and here are the sibling hashes to check.
- Two Clean Snapshots Can Break the WorldDatabase isolation levels are not adjectives about safety. They are claims about which shapes of history are allowed to exist.
- Adam Chooses the Units Before It StepsAdam is not just SGD with momentum. It keeps a per-coordinate memory of gradient scale, then silently decides what a unit step means.
- If You Keep Looking, Use a BankrollE-values and e-processes turn sequential evidence into a nonnegative martingale, so repeated looks become part of the design instead of a loophole.
- AUC Only Sees the PairROC AUC is not accuracy, calibration, alert quality, or deployment utility. It is the probability that one random positive outranks one random nega...
- Hash Tables Are Fast Until the Tail ArrivesOpen-addressed hash tables do not merely get fuller. Their probe distributions change shape.
- SHAP Bars Need a Chosen WorldSHAP values are not pure feature truth. They are Shapley values for a particular missing-feature game, background population, and prediction scale.
- LoRA Bets on the Missing DirectionsLoRA is not magic compression. It freezes the base model and asks whether the task-specific update mostly lives in a small spectral neighborhood.
- Collisions Keep the ReceiptsCount-Min is not a tiny dictionary. It is a one-sided bargain: lose identities, keep upper bounds, and budget the collision mass.
- PageRank Reenters Through the Side DoorPageRank is a damped random walk, and the teleport vector is the quiet place where outside judgment enters the graph.
- Reliable Bits Need Empty SpaceError correction is not a spell that makes bits reliable. It is a geometric bargain: spend redundancy to move valid messages farther apart.
- Formulas Can Lie to the ComputerVariance looks like a formula, but in finite precision arithmetic the algorithm you choose can decide whether the answer is accurate, noisy, or imp...
- Keep the Distances, Audit the DecisionsThe Johnson-Lindenstrauss lemma is not a slogan about making data smaller. It is a promise about finite geometry, with margins the application stil...
- Every Eviction Is a PredictionA cache replacement policy is not housekeeping. It is an online prediction about which object will be useful before memory runs out.
- Averages Mix Before They JudgeSimpson's paradox is not an arithmetic prank. It appears when a mixture is asked to answer a causal question it was not built to answer.
- Ask Who the Instrument MovedInstrumental variables do not recover the effect for everyone. Under strong assumptions, they recover the effect for people whose treatment was mov...
- Bootstrap the Assumption, Not the WishResampling is not magic. A bootstrap interval works by replaying a data-generating contract, and the wrong contract can be confidently wrong.
- Winning Can Be Bad NewsIn a common-value auction, winning is not just success. It is evidence that your estimate was unusually optimistic.
- Clusters Remember by OverlapReplicas do not remember because they agree in spirit. They remember when the next decision is forced to meet the last one.
- Who Leaves the Curve?Kaplan-Meier curves do not magically solve missing follow-up. They keep censored futures in the accounting and expose the assumption doing the work.
- Inverses Turn Noise Into BetsMean-variance optimization is elegant until the covariance matrix is estimated from too little history. The inverse turns tiny noisy eigenvalues in...
- Benchmarking Through the OutageClosed-loop latency tests can stop sending work exactly when the service is slow. Coordinated omission turns a tail-latency incident into a tidy pe...
- Do Not Trust a Naked ECECalibration plots are measuring instruments. Change the ruler, the bins, or the slice, and the confidence story can change with it.
- Sometimes the Posterior Needs a CrowdParticle filters approximate a nonlinear, non-Gaussian filtering distribution with weighted samples. The hard part is not drawing samples; it is ke...
- Button Presses Carry ClocksA drift-diffusion model treats a decision as noisy evidence accumulating until it crosses a threshold. The boundary prices time, mistakes, and unce...
- Yesterday Can Tighten an A/B TestCUPED borrows pre-period signal to reduce variance. Move the timing line past treatment, and precision quietly turns into bias.
- How Far Should a Noisy Sensor Move You?A Kalman filter is not a smoothing trick. It is a recursive argument about how much to trust a model versus a noisy measurement.
- Similar Pages Should Share a Lottery TicketMinHash turns document overlap into a collision event whose probability is the Jaccard similarity.
- Bit Arrays That Only Say MaybeA Bloom filter is a memory bargain: it can say definitely not, or maybe, but it never proves membership.
- Second Tries Are Still TrafficA retry is not free resilience. Under overload, it can become a traffic generator aimed at the component already failing.
- Fuzzy Joins Write FactsWhen there is no shared key, a data join becomes a statistical decision about identity under measurement error.
- Privacy Starts With a Volume KnobDP-SGD is often described as SGD plus noise. The better description is clipped influence plus accounted randomness.
- Midpoint as TruceA bid-ask spread is not one cost. It prices immediacy, adverse selection, inventory risk, and the mechanics of keeping a market open.
- Control Groups Hold Ghost SlopesDifference-in-differences is not before-minus-after. It is a wager that the control group carries the untreated trend the treated units never got t...
- Leaderboards Have Seating ChartsLLM-as-judge evaluation is not just a model call. It is a measurement protocol with position effects, verbosity effects, calibration choices, and s...
- Loss Functions Never See the Missing PairPreference optimization is not a direct tap into human values. It is a policy update seen through pairwise labels, a reference model, a KL temperat...
- Counting Without NamesHyperLogLog is the strange bargain behind large-scale distinct counts: forget identities, keep rare hash patterns, and still estimate how many uniq...
- Witnesses Are Not CamerasA memory report is not raw footage. It is reconstructed from noisy traces, gist, priors, cues, and a decision threshold.
- Merge Functions Become ProductsCRDTs are not a trick for avoiding distributed systems. They are a way of making merge algebra explicit: what grows, what can be forgotten, and wha...
- Alpha Was Hiding in the TimestampAt high frequency, a lead-lag signal can be market structure, a data-engineering artifact, or both. The Epps effect is the warning label.
- Context Windows Have Rent DueLong-context inference is not only attention math. Every live token leaves key and value vectors in GPU memory, and the serving stack pays that ren...
- Every Trial Spends the BudgetBayesian optimization is not optimizer dust. It is a policy for choosing the next expensive experiment under uncertainty, and its failures belong i...
- Models Forget the Weird Stuff FirstRecursive training can look healthy in the average case while sampling, filtering, and reuse quietly sand away rare modes.
- Dashboards Can Change the ExperimentA fixed-horizon p-value is a different object once the team checks it every morning and ships the first good-looking day.
- Risk Dials Have DelaysScaling exposure by estimated volatility is not a free Sharpe-ratio machine. It is a lagged control loop with estimator error, leverage caps, turno...
- Last Layer, Solver IncludedDiffusion and flow-matching models do not end at the neural network. The path, vector field, and numerical solver decide what distribution appears.
- Models Overbook Feature SpaceNeural networks may store more features than dimensions. Sparsity makes that bargain possible, and interference sends the bill.
- Every Expert Has a Line OutsideSparse mixture-of-experts models promise cheap active compute, but the router is really allocating scarce expert capacity under bursty demand.
- Straight Lines That Lie PolitelyHeavy tails are easy to see and hard to prove. A straight-ish log-log plot is a lead, not a verdict.
- One Outlier Can Spend the Whole ScaleQuantizing a language model is not just rounding weights. A few large activation channels can spend the whole numerical budget for everyone else.
- Draft Tokens Have to Earn Their KeepSpeculative decoding is often sold as free speed. The real question is whether accepted draft tokens repay the draft model, verifier overhead, and ...
- Vectors That Show Up EverywhereNearest-neighbor retrieval can fail because one vector becomes everybody's neighbor. Hubness is not a vague curse of dimensionality; it is a measur...
- Build a Copy Machine You Can SeeInstead of treating in-context learning as magic, build a browser-sized induction circuit that finds a repeated pattern and copies what followed it.
- After the Decision, the Trade BeginsA portfolio model has a decision price. The desk has a path through liquidity, volatility, impact, patience, and information leakage.
- Which Zero-Error Model Did You Get?Double descent is not a moral boundary between underfitting and overfitting. It is a clue that geometry, noise, and implicit regularization selecte...
- Feeds Teach Users, Then Take NotesA recommender does not merely observe preference. It allocates exposure, changes what users can learn about, and trains on behavior it helped create.
- Leaderboards Make Blunt InstrumentsAI benchmarks do not just score models. They build measuring instruments with ranges, reliability limits, blind spots, and exposure risks.
- Beautiful Backtests Come From a CrowdA beautiful backtest is weak evidence until the ideas, parameters, filters, assets, and windows that lost are counted too.
- Saying No Needs a Price ListA model should not answer because confidence cleared a universal threshold. The stop rule belongs to error costs, abstention costs, and the value o...
- Coverage Comes With Fine PrintConformal intervals are powerful because their guarantee is explicit: coverage under an exchangeability story, not magic under every deployment shift.
- Metrics Become Control SurfacesMetrics fail under pressure because selection, gaming, and feedback change the data-generating process that made the metric useful.
- Thirty-Two Tries Can Be One MistakeRepeated reasoning samples buy intelligence only when attempts are diverse, the selector is reliable, and the system knows when not to answer.
- RAG Is the Question of What Gets ReadLong-context models changed retrieval, but they did not remove the dispatch problem: deciding what evidence deserves attention, placement, and an a...
- Trust First, Then Size the BetKelly sizing is a clean formula with an unforgiving dependency: calibrated probabilities, executable odds, and a drawdown budget that survives cont...
- At 3 AM, Ask the CacheA production pattern for sparse, long-tailed estimates: robust rollups, partial pooling, visible fallback levels, and the discipline to ask for ML ...
No notes match your search.