Responsible Deployment: Staged Release, Access Control
Deploy language models responsibly through staged rollouts, tiered access control, content filtering, and production monitoring systems. Read full notebook
Michael Brenndoerfer | Data & AI, Private Equity, Technology
Personal website of Michael Brenndoerfer - Data & AI, Software Engineering, Private Equity, Technology, and more
Latest posts
Long-Form Generation: Outline, Hierarchy, and Coherence
Covers long-form text generation with outline-based planning, hierarchical decomposition, entity tracking. Read full notebook
Test-Time Compute: Sampling, Refinement, Optimal Inference
Covers test-time compute strategies: multiple sampling, iterative refinement, compute-optimal inference, and inference-time scaling laws for language models. Read full notebook
Learning Rate Warmup: Linear Warmup and Large Batch Training
Explains how learning rate warmup stabilizes early training by gradually increasing the learning rate, with theory, linear warmup. Read full notebook
Deduplication: Exact, Near-Duplicate, and Substring Methods
Explains how deduplication removes exact copies and near-duplicates from training corpora using SHA-256 hashing, Jaccard similarity over character shingles. Read full notebook
Hallucination Mitigation: RAG, Decoding, and Training
Reduce LLM hallucination using retrieval augmentation, self-consistency decoding, DPO training, and calibrated uncertainty expression. Read full notebook
Mechanistic Interpretability: Circuits, Induction Heads
Reverse-engineer transformer networks into human-understandable algorithms by identifying circuits, induction heads, and mechanistic discoveries. Read full notebook
Research Directions: Open Problems, Benchmarks
Examines the open problems, promising research areas, benchmark gaps, and community priorities shaping the future of language AI. Read full notebook
Societal Implications of AI: Labor, Equity, Regulation
Explains how language AI reshapes labor markets, widens or narrows access gaps, drives regulatory frameworks. Read full notebook
Alignment Challenges: Scalable Oversight, Goal Specification
Examines the core alignment challenges facing modern LLMs: scalable oversight, alignment tax, goal mis-specification, reward hacking. Read full notebook
Hallucination Detection: NLI, Self-Consistency
Covers four methods for detecting LLM hallucinations: entailment-based scoring, knowledge base verification, self-consistency checks. Read full notebook
Probing Classifiers: Decoding What Language Models Learn
Explains how probing classifiers reveal what linguistic information is encoded in neural network representations. Read full notebook
Hallucination Types in Language Models
Explains how language models hallucinate: intrinsic and extrinsic hallucination, factual errors, fabrication, and inconsistency with NLI-based detection. Read full notebook
Hallucination Causes: Why Language Models Fabricate Facts
Examines the structural causes of LLM hallucinations: training data noise, exposure bias, knowledge gaps, and generation pressure in language models. Read full notebook
Representation Harms: Stereotyping, Erasure, and Bias
Explains how language models cause harm through stereotyping, erasure, and demeaning associations, with measurement methods and concrete examples. Read full notebook
Capability Frontiers: Emerging AI, World Models, Planning
Examines emerging language model capabilities including in-context learning, chain-of-thought reasoning, world models, planning and agency. Read full notebook
Fairness Metrics: Demographic Parity, Equalized Odds
Covers the key mathematical definitions of algorithmic fairness, from demographic parity to equalized odds. Read full notebook
Efficiency Frontiers: Architecture, Hardware, and Inference
Examines efficient transformer architectures, hardware co-design, inference optimizations like speculative decoding. Read full notebook
Scaling Frontiers: Limits
Examine the physical, statistical, and economic limits of LLM scaling, the data wall crisis, and architectural innovations like MoE and inference-time compute. Read full notebook
Bias Mitigation: Debiasing, CDA, and Fair Fine-tuning
Practical techniques for reducing demographic bias in language models: data balancing, embedding debiasing, adversarial training. Read full notebook
Bias Measurement: WEAT, Stereotype Scores, Fairness Metrics
Measure bias in language models using embedding association tests, generation metrics, classification fairness measures, and standard benchmarks. Read full notebook
Model Cards: Documentation, Intended Use, and Limitations
Write model cards that communicate intended use, training data, evaluation results, and limitations for responsible AI deployment. Read full notebook
Bias in Language Models: Sources, Types and Amplification
Explains how language models inherit demographic, cultural, and occupational bias from training data, and why they amplify these biases beyond the data. Read full notebook
LLM Watermarking: Schemes, Detection, and Robustness
Explains how token-level watermarking embeds hidden statistical signals into LLM outputs, enabling cryptographically verifiable attribution and AI provenance. Read full notebook
Evaluation Prompt Engineering: Designing Reliable LLM Judges
Design reliable LLM judge prompts using explicit criteria, few-shot examples, and chain-of-thought formatting to maximize evaluation accuracy. Read full notebook
Memorization and Privacy in Language Models
How language models memorize training data, methods for measuring extractable memorization, PII risks in web-scale corpora, and practical privacy mitigations. Read full notebook
Position Bias in LLM Judges: Measurement and Mitigation
Explains how position bias, verbosity bias, and sycophancy distort LLM evaluation. Measure swap consistency, detect length effects. Read full notebook
Retrieval-Augmented Training: RETRO Architecture
How RETRO trains language models with retrieval from scratch, using chunked cross-attention to integrate a 2T-token database and cut parameter needs 25x. Read full notebook
LLM-as-Judge: Scalable AI Evaluation with Language Models
Build LLM-as-Judge evaluation pipelines: prompt design, judge model selection, calibration against human annotations, and bias mitigation. Read full notebook
Process Reward Models: PRM Training, Math Reasoning
Process reward models score individual reasoning steps instead of final answers alone. Covers training data, credit assignment, math tasks, and limitations. Read full notebook
Constitutional AI: Principles, Critique, Revision
Explains how Constitutional AI trains safer LLMs using constitutional principles, AI-driven critique and revision, and RLAIF preference labeling. Read full notebook
LLM Preference Evaluation: Pairwise Comparisons and Elo
Evaluate language models with pairwise comparisons and Elo ratings, including preference aggregation, confidence intervals, bias, and significance tests. Read full notebook
Cohen, Fleiss & Krippendorff: IAA Metrics & Implementation
Covers chance-corrected agreement metrics for NLP annotation reliability. Calculate Cohen's kappa, Fleiss' kappa, and Krippendorff's alpha with Python examples. Read full notebook
Reasoning Frontiers: o1 and Test-Time Compute
Examines o1-style reasoning models, test-time compute scaling, process reward models, and open research questions shaping the frontier of AI reasoning. Read full notebook
Human Evaluation Design: Interface, Instructions & Cost
Design rigorous human evaluations for language models. Topics include interface design, annotation paradigms, annotator selection. Read full notebook
Benchmark Saturation: AI Evaluation Metrics, Ceiling Effects
Covers benchmark saturation in AI evaluation. Explains why static metrics hit ceiling effects, lose statistical power, and how dynamic benchmarks solve this. Read full notebook
Reasoning Limitations: Failures, Shortcuts
Examines systematic reasoning failures in LLMs including spurious correlations, reasoning shortcuts, negation failures. Read full notebook
Benchmark Contamination in LLMs: Detection
Explains how benchmark contamination affects LLM evaluation accuracy. Topics include detection methods including n-gram matching, MinHash. Read full notebook
Mathematical Reasoning in LLMs: Benchmarks, Training, Limits
Explains how LLMs solve math problems, from grade-school word problems to competition math. Topics include chain-of-thought, process reward models, GRPO. Read full notebook
TruthfulQA: Testing AI Truthfulness and Imitative Falsehoods
Explains how TruthfulQA tests AI truthfulness by identifying imitative falsehoods. Read full notebook
Reasoning Verification: Process Reward Models, Guided Search
Explains how process reward models score each reasoning step, how verification-guided search selects correct chains. Read full notebook
Reasoning Strategies: Self-Consistency, Tree of Thought
Explains how self-consistency, tree of thought, least-to-most prompting, and decomposition strategies improve language model reasoning accuracy and reliability. Read full notebook
MBPP: Python Code Generation Benchmark
Examines the MBPP benchmark for Python code generation evaluation. Topics include crowd-sourced programming tasks, pass@k metrics, execution-based testing. Read full notebook
Chain-of-Thought Prompting: Zero-Shot, Fine-Tuning
Explains how chain-of-thought prompting enables language models to reason step by step. Topics include few-shot CoT, zero-shot CoT, self-consistency. Read full notebook
HumanEval: Functional Code Generation Evaluation with Pass@k
Explains how HumanEval evaluates LLM code generation using functional correctness and the pass@k metric. Read full notebook
Text Generation Applications: Use Cases and Quality
Examines LLM text generation for content creation, writing assistance, and code. Topics include quality dimensions, constraint verification, prompt design. Read full notebook
Reasoning Foundations: Types, LLMs, Failure Modes
Examines deductive, inductive, abductive, and causal reasoning in LLMs, including how transformers support inference chains and where reasoning breaks down. Read full notebook
GSM8K: Evaluating Mathematical Reasoning in Language Models
GSM8K tests grade-school mathematical reasoning with multi-step word problems. Covers dataset structure, answer scoring, and known benchmark limitations. Read full notebook
Conversational AI: Dialogue Systems and Chatbot Design
Build intelligent dialogue systems with LLMs, covering conversation management, slot filling, memory strategies, and chatbot evaluation techniques. Read full notebook
Model Merging Applications: Multi-Task, Style
Apply model merging to combine task fine-tunes, blend styles, compose capabilities, and evaluate merged models using normalized scores and Pareto analysis. Read full notebook