02 / AI / SI Problems
AI / SI Problems
50 practice problems. How modern AI is built, trained, evaluated and constrained, from gradient descent to the superintelligence question.
Warm-up6 problems
- 01→
Softmax Temperature Scaling
Derivationsoftmaxtemperatureprobability - 02→
Cross-Entropy Loss Derivation
Derivationloss functionscross-entropyclassification - 03→
Perplexity as Evaluation Metric
Derivationevaluationperplexitylanguage models - 04→
Softmax Temperature and Output Distributions
Derivationsoftmaxtemperaturesampling - 05→
Output Token Probabilities and Calibration
Derivationprobabilitycalibrationlanguage models - 06→
Position Encodings in Transformers
Derivationtransformersposition encodingattention
Interview29 problems
- 01→
Backpropagation Through a Network
Derivationbackpropagationchain rulegradients - 02→
Vanishing Gradient Problem
Conceptualgradientsdeep networksactivation functions - 03→
Self-Attention Computation
Derivationattentiontransformersmatrix operations - 04→
Sinusoidal Positional Encoding
Conceptualpositional encodingtransformerssequences - 05→
Layer Normalization vs Batch Normalization
Conceptualnormalizationtransformerstraining - 06→
Neural Scaling Laws
Conceptualscalingcomputetraining - 07→
Compute-Optimal Training Budget
Derivationcomputescalingoptimization - 08→
RLHF Training Pipeline
ConceptualRLHFalignmentreward modeling - 09→
Linear Probing for Interpretability
Conceptualinterpretabilityprobingrepresentations - 10→
Jailbreak Attacks on LLMs
Conceptualsafetyadversarialred-teaming - 11→
Measuring Agent Reliability
System Designagentstool useevaluation - 12→
Implementing a Transformer Block
Codingtransformersimplementationarchitecture - 13→
Implementing Adam Optimizer
CodingoptimizationAdamtraining - 14→
Byte-Pair Encoding Tokenization
CodingtokenizationBPENLP - 15→
KV Cache for Efficient Inference
ConceptualinferenceKV cacheefficiency - 16→
Constitutional AI Principles
ConceptualalignmentConstitutional AIsafety - 17→
Reward Hacking and Specification Gaming
Conceptualreward hackingalignmentspecification - 18→
Capability Elicitation and Sandbagging
Conceptualcapabilitiesevaluationdeception - 19→
Loss Landscape and Optimization
Conceptualoptimizationloss landscapetraining - 20→
Emergent Capabilities in Language Models
Conceptualemergencescalingcapabilities - 21→
Distributional Shift in RLHF
ConceptualRLHFdistributional shiftreward modeling - 22→
Attention Mechanism Complexity
Derivationattentioncomplexityefficiency - 23→
Value Learning Problems
Conceptualvalue learningalignmentpreferences - 24→
Neural Scaling Laws
Derivationscalingpower lawscompute - 25→
Constitutional AI Methodology
System Designconstitutional AIRLAIFalignment - 26→
Neural Network Interpretability Techniques
Conceptualinterpretabilityfeaturescircuits - 27→
Safety Considerations for Tool-Using AI
System Designtool useagentssafety - 28→
Variance Reduction in Policy Gradients
Derivationpolicy gradientvarianceRL - 29→
Batch Normalization Mechanics
Derivationbatch normalizationtrainingoptimization
Hard15 problems
- 01→
Direct Preference Optimization
DerivationDPOalignmentoptimization - 02→
Superintelligence Threat Models
Conceptualsuperintelligencealignmentsafety - 03→
Emergent Abilities in Language Models
Conceptualemergencescalingcapabilities - 04→
Speculative Decoding
Conceptualinferencedecodingefficiency - 05→
Mixture of Experts Architecture
ConceptualMoEarchitectureefficiency - 06→
Extending Context Length
Conceptualcontext lengthattentionarchitecture - 07→
Watermarking LLM Outputs
Conceptualwatermarkingdetectionsafety - 08→
Mesa-Optimization and Inner Alignment
Conceptualmesa-optimizationinner alignmentdeception - 09→
Grokking in Neural Networks
Conceptualgrokkinggeneralizationtraining dynamics - 10→
Mode Connectivity in Loss Landscapes
Conceptualloss landscapemode connectivityoptimization - 11→
Multi-Agent AI Safety
System Designmulti-agentgame theorysafety - 12→
Superposition in Neural Networks
Conceptualsuperpositioninterpretabilityfeatures - 13→
AI Safety via Debate
System Designdebatescalable oversightsafety - 14→
Activation Patching for Circuit Analysis
Derivationactivation patchingcircuitsinterpretability - 15→
The Corrigibility Problem
Conceptualcorrigibilitycontrolalignment
Topics covered
AdamBPEConstitutional AIDPOKV cacheMoENLPRLRLAIFRLHFactivation functionsactivation patchingadversarialagentsalignmentarchitectureattentionbackpropagationbatch normalizationcalibrationcapabilitieschain rulecircuitsclassificationcomplexitycomputeconstitutional AIcontext lengthcontrolcorrigibilitycross-entropydebatedeceptiondecodingdeep networksdetectiondistributional shiftefficiencyemergenceevaluationfeaturesgame theorygeneralizationgradientsgrokkingimplementationinferenceinner alignmentinterpretabilitylanguage modelsloss functionsloss landscapematrix operationsmesa-optimizationmode connectivitymulti-agentnormalizationoptimizationperplexitypolicy gradientposition encodingpositional encodingpower lawspreferencesprobabilityprobingred-teamingrepresentationsreward hackingreward modelingsafetysamplingscalable oversightscalingsequencessoftmaxspecificationsuperintelligencesuperpositiontemperaturetokenizationtool usetrainingtraining dynamicstransformersvalue learningvariancewatermarking
Practise on Uncertain Systems
These problems test whether you understand the material. Uncertain Systems verifies that understanding—so you know what you actually know, not just what you read.