Mechanisms
285 mechanisms · 391 lessons
Model training and research
- Dropout: Preserve Expectation While Masking Activations 1 application
- Fit training activations by saving boundaries and recomputing segments
- Adam: Track Gradient Direction and Coordinate Scale
- Defend an ML derivation when its assumptions change
- Choose Discriminating Probes for a Black-Box Model
- Softmax Cross-Entropy: From Logits to a Corrective Gradient 4 applications
- Critique a paper by testing its strongest alternative explanation
- Design a matched intervention that tests an ML mechanism
- Test a data mixture with matched checkpoint continuations
- L2 and Weight Decay: Where Does Shrinkage Enter the Update?
- Choose a test that separates training failure causes 1 application
- Choose model size and tokens under fixed compute
- Floating-Point Formats: Range Is Not Precision
- Implement a zero-impact LoRA adapter and test its first update
- Shorten a large gradient without changing its direction
- Choose pruning by stored bytes and executed shapes
- Separate rare-event sampling from probability and action choices
- Choose an initial scale by tracing signal through a layer
- Implement beam search with explicit live and finished policies
- Average micro-batch gradients before one optimizer update
- Save gradient history before choosing the next update
- BatchNorm and LayerNorm: Which Values Share Statistics?
- Test an optimizer gap without mistaking sharpness for its cause
- Pack tokens while preserving independent causal examples
- Mixup and CutMix: Keep Input Contributions and Targets Coupled
- Differentiate a Gaussian Sample by Holding Its Noise Fixed
- Mixed Precision: Preserve Gradients and Guard the Update 1 application
- Trace gradient credit through repeated uses of one weight
- Regularization: Diagnose the Gap and Choose a Constraint
- Set a step size from the training clock 1 application
- Train a student against informative teacher targets
- Find gradient, mode, and evaluation-state bugs in a training loop 1 application
- Verify synthetic examples without leaking final evidence
- Verify the Training Pipeline Before Tuning the Model 1 application
- Curate a traceable training distribution 1 application
Coding interview practice
- Expire every ended interval before measuring active capacity
- Insert singleton axes to preserve point-center-feature ownership
- Maintain ordered halves with the lower half owning the extra item
- Borrow border cells for instructions after saving their original obligations
- Keep the best k values with the weakest retained value at the root
- Use bounded occurrence counts as ranking coordinates
- Shrink a sorted candidate interval without losing the target 2 applications
- Compare skipping the current item with its last compatible prefix 1 application
- Factor clockwise rotation into row reversal and one-sided transpose swaps
- Keep sum-without-carry and unresolved carries as separate states
- Count disjoint final-move cases and retain only their dependencies 3 applications
- Keep the earliest finishing compatible range to preserve future room
- Combine n-minus-one edges with full undirected reachability
- Keep the best nonempty range forced to end at the current position
- Discard a wall whose best remaining partner is already bounded
- Freeze the current frontier before adding the next level 1 application
- Keep only distinct membership when equality collisions are the whole question
- Count exactly k events by subtracting two valid-start ranges
- Track missing copies and shrink each valid cover until one requirement breaks
- Discard irrelevant order, but retain every character count 1 application
- Combine the contribution strictly before and after each position
- Export one extendable branch but score a complete two-branch path
- Repair irrelevant endpoints before comparing symmetric characters
- Keep a fixed lead until the trailing pointer reaches the target predecessor
- Reverse all declared positions,including leading zeros
- Slide a fixed-width count summary instead of enumerating permutations
- Rebuild an algorithm from the state future steps need
- Partition half-open ranges without losing the short tail
- Use a historical frequency bound only for the maximum-length contract
- Allocate one copy per original identity before following edges
- Reject an undirected edge whose endpoints already share a component
- Pair a direct key index with explicit recency order
- Carry the legal interval inherited from every ancestor
- Jump past a repeated occurrence without moving the window backward
- Read a length before treating payload characters as opaque
- Clear exactly one lowest set bit per iteration
- Minimize over every valid final coin and its solved predecessor
- Keep the smallest possible tail for each achieved subsequence length
- Search the first speed that satisfies the deadline 1 application
- Resolve waiting indices at their first strictly warmer day 1 application
- Suspend outer decoding state while a nested repeat is built
- Close the newest unmatched bracket first
- Finalize the cheapest pending path under nonnegative edge costs
- Sort all meetings so the first incompatibility is an adjacent boundary
- Summarize the best earlier baseline for each possible sale
- Sort starts to keep only one open union component 1 application
- Descend only while both targets share the same ordered side
- Consume a BST's sorted inorder stream only as far as the requested rank
- Compare aligned values and missing-child positions together 1 application
- Select candidate membership before ordering only the retained group 2 applications
- Count new component seeds and mark each whole region 1 application
- Count signed ranges from earlier prefix totals
- Find a pair by querying only earlier values
- Copy data and validity into exactly the same slice
- Use strict value growth to make neighbor moves an acyclic suffix problem
- Compress all reachable forward choices into one prefix boundary
- Encode every recursive child slot,including missing ones
- Use a one-link relative gain to force a cycle collision
- Reverse the edges when many starts share a few goals
- Save each increasing-index choice path, then undo the last choice 3 applications
- Store each prefix minimum so pop restores the earlier answer
- Save the untouched suffix before reversing its incoming route 1 application
- Represent unvisited cells as one shrinking rectangle
- Retain both ending extremes when a negative can reverse their order
- Reuse the smaller integer left after removing one bit
- Extend a snapshot of reachable sums to use each occurrence once
- Attach the smaller frontier and advance only that source 1 application
- Fix one value, then discard impossible sorted pairs
- Exploit score-shift invariance before exponentiation
- Enumerate each palindrome through its unique odd or even center 1 application
- Share prefix paths but mark complete words separately 1 application
- Launch exact dictionary matches only from proven segment boundaries
- Use root order to partition exact inorder membership ranges
- Return transformed child roots, then reconnect them locally
- Return one subtree height after both children 1 application
- Traverse board paths and trie prefixes together with different state lifetimes
- Define two-prefix answers before compressing the table to rows 1 application
- Expand each consecutive run only from its unique start
- Cancel the complete expected universe against its observed members
- Release a dependency only when its unmet-requirement count reaches zero 2 applications
LLMs, agents, and post-training
- Choose an adaptation by the failure it must fix 1 application
- Trace attacker influence to a protected-asset boundary 1 application
- Test whether a visible trace adds useful monitoring evidence
- Cache past keys and values, but still read the growing prefix 3 applications
- Return a useful completion for the current editor state 1 application
- Route each token to selected feed-forward experts
- Bound a healthcare drafting task by evidence and clinical authority
- Reuse a finished decode slot without stopping other requests
- Trace communication and per-token updates through a transformer block 1 application
- Choose tokens by probability, retained support, or sequence score
- Compute softmax attention without storing every score 1 application
- Correct draft proposals without changing the target distribution 1 application
- Replace pairwise attention with a defined compressed computation
- Govern agent actions with durable authority and outcome contracts 1 application
- Compare parallel prompt rows with serial decode rows
- Intervene on internal state to test a mechanism hypothesis
- Resolve logical KV positions through shared physical blocks
- Choose quantization scales before claiming memory or speed gains
- Separate better judgment from containment under subversion 1 application
- Choose where position enters an attention computation 1 application
- Build preference measurements that remain meaningful under optimization 1 application
- Answer a question from retrieved, identifiable evidence 3 applications
- Restrict attention edges and trace information across layers
- Turn text into the exact token IDs a model expects
- Turn one prompt's checked outputs into relative advantages 1 application
- Spend inference compute only when selection evidence improves
Retrieval, ranking, and recommendations
- Personalize from permitted early evidence and bounded exploration
- Spend search work and vector bytes separately
- Add language modeling at a measured recommendation boundary
- Choose complementary purchase evidence, not generic similarity
- Score a new item's features against a known profile
- Govern a feed whose attention policy changes future creator supply 1 application
- Decide whether vector length should affect retrieval
- Reuse feature vectors instead of learning every pair weight
- Predict a missing rating from two learned rows 1 application
- Separate repeated matches, rarity and document length
- Mine confusing candidates without inventing negative labels
- Weight logged ranking outcomes only for supported counterfactuals
- Preserve query intent across a personalized search and learning loop
- Choose a score surrogate after defining the ranking group and metric
- Use each ranking evidence layer for the question it can answer
- Design a video recommender as a selective exposure policy
- Test a relation score before increasing embedding width
- Move item encoding out of the request path 1 application
- Compose whole recommendation shelves, not just the best local item lists
ML foundations
- Add an Identity Route, Not a Gradient Guarantee 1 application
- Alternate GAN Updates Without Detaching the Generator's Signal
- Fit an Encoder by Bounding a Latent Model's Likelihood
- Separate Approximation Existence from Finding the Weights
- Use Keys to Weight Values for Each Query 5 applications
- Read a matrix through its transformed basis vectors
- Fit linear log-odds and choose a probability-based decision 1 application
- Calculate spread by averaging centered products 2 applications
- Reduce centroid distortion by alternating fixed-center and fixed-membership steps
- Why L1 Can Set a Coefficient Exactly to Zero
- Choose Positive Pairs Before Optimizing Contrastive Scores
- Grow density-connected clusters through core points, not border points
- Derive parameter gradients with a Jacobian transpose
- Differentiate a Discrete Expectation Without Differentiating the Sample
- Compare paired model gains with statistical and practical thresholds
- Compute uncertainty removed by a noisy observation 2 applications
- Read a probability factorization and test what conditioning changes 1 application
- Reuse a Fixed Source While the Target Prefix Grows
- Condition a Gaussian function prior to predict a value and its uncertainty
- Train One Noise Prediction, Then Chain Reverse Updates
- Compute a feature-space inner product without constructing the features
- What Does a 95% Interval Promise?
- Preserve Mass While a Flow Changes Local Volume
- Trace Gradient Gain Through Repeated Jacobians
- Model dependent observations with hidden Markov states and local emissions
- Separate shared Gaussian factors from feature-specific noise
- Fit fixed observations by maximizing their likelihood 1 application
- Choose a Nonlinearity by Its Values and Local Slopes
- Trade separating margin against violations and identify support vectors
- Read local-neighborhood maps without treating their layout as measured geography
- Choose a probability model from the observation and process
- Fit latent models by tightening and improving a likelihood bound 1 application
- Normalize evidence-weighted hypotheses after a positive test
- Choose What to Estimate Before Choosing an Action
- Correct an expectation when samples come from a proposal
- Choose a weighted impurity reduction and trace its prediction region
- Build bootstrap replicates from the actual sampling units
- Accumulate Local Derivatives Backward from One Loss 2 applications
- Compare Sampling Dependencies, Not Token Versus Image Stereotypes
- Approximate repeated sample means without changing the data distribution
- Calculate signed area, then transform a density
- Choose a PCA rank from discarded squared variation
- Apply a symmetric matrix in its eigenbasis 1 application
- Fit squared-error predictions and check residual orthogonality
- Fit the next tree to the current ensemble's loss gradient
- Propagate state mass and check a stationary split
- Choose a Peak Learning Rate, Then Test Its Schedule
- Turn a Graph Neighborhood into One Feature Update
- Compute a marginal by collapsing tree branches into messages 1 application
- Separate within-model noise from disagreement across models
ML systems and infrastructure
- Trace immutable dependencies backward for replay and forward for impact
- Choose a collective from the data every rank must finish with 2 applications
- Recover the optimizer transition, not just the collective 2 applications
- Find the event that extends the synchronized step 1 application
- Check scaling assumptions, routed semantics, and fused-kernel equivalence separately
- Map communication frequency and volume onto actual links 3 applications
- Monitor data, decisions, and mature outcomes with different claims
- Trade replicated training state for temporary gathered units
- Make payment-risk decisions without mistaking selected outcomes for truth 1 application
- Identify the memory level that limits useful arithmetic
- Schedule micro-batches across layer stages without changing the optimizer step
- Count every model call once, then locate avoidable cost 1 application
- Fit a ranking funnel to traffic, latency, and full unit cost 2 applications
- Separate historical feature reconstruction from latest-value serving 1 application
- Build separate transformer compute, training-state, and request-cache budgets
Interview and career practice
- Confirm the work sample before choosing practice
- Defend a proportionate release decision with evidence
- Make a bounded technical claim that evidence can change
- Explain a values decision without hindsight
- Resolve a technical disagreement without rewriting the evidence
- Choose work against its next-best alternative 1 application
- Practice five observable skills across changing prompts 1 application
- Turn an ambiguous symptom into a bounded testable decision
- Present a project through decisions you can defend 1 application
- Choose the work profile,not the role label
- Calibrate scope with evidence,not a title translation
- Make one contribution claim inspectable and honestly bounded
Evaluation and product ML
- Read classification metrics from one labeled error table 3 applications
- Choose allow, intervene, or review from conditional costs
- Choose validation boundaries that match unseen production data 2 applications
- Locate an online regression before changing the model
- Compare complete matched runs rather than selected winners 1 application
- Separate a prediction from the effect of an action
- Explain a prediction with a declared reference and attribution rule
- Choose rank credit for graded relevance, all hits, or the first hit
- Check whether an outcome is observable before calling it negative
- Make predicted probabilities match the event they describe 1 application
- Design an evaluation contract for a document assistant 3 applications
- Estimate a model's product effect with a randomized experiment 2 applications
- Sweep one score ordering into ROC and precision-recall views
- Validate a model judge before trusting its semantic scores
Reinforcement learning and robotics
- Find where PPO stops rewarding a sampled probability change
- Compute advantages with a boundary-aware reverse scan
- Constrain the episode before grading its evidence
- Train with joint evidence while acting from local information
- Preserve task ordering with telescoping potential rewards
- Weight a sampled action's log-probability gradient by return 1 application
- Choose the learned action interface before naming an RL algorithm
- Learn robot actions under state-distribution and control constraints
- Compute a decision value from reward and continuation 1 application
- Choose actions from means,uncertainty bounds or posterior draws 1 application
Vision, language, and speech
- Arrange convolution patches as matrix columns
- Retain and write recurrent state coordinate by coordinate
- Sum frame alignments after a precise collapse rule
- Restore dense predictions using retained boundary evidence
- Choose how candidate objects become a final prediction set
- Learn word-vector scores from co-occurrence evidence
- Choose a speaker-match threshold from risk and attack evidence
- Turn image patches into positioned transformer tokens
- Keep live evidence current and user control immediate
- Choose where modality evidence meets and how much survives
- Decode box offsets and suppress selected overlaps
- Separate MLM target selection from input corruption
- Score complete label paths before choosing a sequence
- Name the changed distribution before choosing a correction
- Choose where speech-system knowledge can be updated
- Choose how audio evidence becomes a transcript
- Emit provisional words before committing the transcript
- State the allowed input changes before claiming robustness
- Advance audio time or emit a label from the same frame