Responsible, Secure, and Interpretable Language AI

Covers bias sources, bias types (demographic, cultural), bias in training data, bias amplification.

Covers embedding association tests, generation bias metrics, classification bias metrics, bias benchmarks.

Practical techniques for reducing demographic bias in language models: data balancing, embedding debiasing, adversarial training, and prompt-based interventions.

Learn the key mathematical definitions of algorithmic fairness, from demographic parity to equalized odds, and why satisfying all metrics simultaneously is impossible.

Covers stereotyping, erasure, demeaning associations, measuring representation harms.

Covers debugging, trust, safety, scientific understanding, interpretability approaches overview.

Covers attention weight extraction, attention head visualization, attention interpretation caveats, attention tools.

Covers attention vs importance, attention manipulation studies, gradient-based alternatives.

Covers linear probing methodology, probing task design, probing interpretation, control tasks.

Covers layer selection, representation evolution, task localization, layer probing patterns.

Covers patching methodology, locating information, patching experiments, causal tracing.

Covers logit lens concept, intermediate vocabulary projection, tuned lens, lens interpretation.

Covers SAE architecture, sparsity constraints, dictionary learning, SAE for LLMs.

Covers feature activation patterns, feature naming, automated interpretation, feature circuits.

Reverse-engineer transformer networks into human-understandable algorithms by identifying circuits, induction heads, and mechanistic discoveries.

Learn how steering vectors and activation addition let you modify language model behavior at inference time by injecting directional nudges into the residual stream.

Covers harmful content generation, misuse scenarios, unintended harms, safety threat models.

Learn how red teams systematically probe language models for safety failures, covering attack taxonomies, ASR metrics, RL-based automated attack generation, and real-world findings.

Covers jailbreak techniques, prompt injection, adversarial suffixes, jailbreak defenses.

Covers direct prompt injection, indirect prompt injection, injection in RAG, injection defenses.

Covers classification-based filtering, rule-based filtering, filter placement, filter evaluation.

Covers input guardrails, output guardrails, guardrail frameworks, guardrail design.

How language models memorize training data, methods for measuring extractable memorization, PII risks in web-scale corpora, and practical privacy mitigations.

Learn how differential privacy protects training data in language models, from the mathematical guarantee to DP-SGD, privacy budgets, and practical LLM fine-tuning.

Assets, actors, trust boundaries, control flow, data flow, and failure propagation across model-centered systems.

Malicious instructions in documents, web pages, messages, tool output, and retrieved context.

Schema attacks, compromised tools, poisoned indexes, persistent memory attacks, and integrity controls.

User and agent identity, scoped credentials, delegation, expiry, consent, and policy enforcement outside the model.

Credential isolation, capability sandboxes, network and filesystem boundaries, output validation, and containment.

Tool registries, dependency tampering, server discovery, trust-on-first-use, signing, and update risk.

Tamper-evident traces, anomaly detection, kill switches, investigation, rollback, and post-incident learning.

Task allocation, mixed initiative, handoffs, feedback, shared context, and preserving human agency.

Calibrated reliance, overtrust, undertrust, explanations, uncertainty, and behavior under time pressure.

Review queues, contestability, escalation paths, documentation, and accountability for consequential decisions.

Disability access, literacy, language inclusion, participatory design, and evaluating who benefits or is excluded.

Documenting complete systems, evaluating downstream use, independent audits, and monitoring material changes.

Ownership, risk tiers, review gates, incident reporting, change management, and retirement.