Good vs Evil: Chosing the Right Moral Compass for you LLM
summary
Argues that moral alignment must be region-specific and actionable: train with local curated datasets, map community values into modular value boxes that translate into concrete loss functions, and apply fine-tuning (RLHF/LoRA) plus prompt-tuning and API gating. Shows a workflow: collect messy labeled data, run automated red-teaming and human safety reviewers, deploy with monitoring and rollback. Claims those safety reviewers and static measures often fail after model drift, so continuous tuning and provenance tracking are required for reliable alignment.
AI-written summary · about
chapters (8)
- 0:00Introduction and context
- 3:11Cultural identity and upbringing
- 6:31Identity crisis and self-perception
- 9:27Influence of social media and external pressures
- 12:49Philosophy and self-colonization
- 15:36Question and answer session
- 17:29Model alignment and ethical dilemmas
- 20:03Model behavior and user expectations
related talks
29mCustomizing and Evaluating LLMs, an Ops Perspective
01h 29mLLM Safety, Alignment, and Generalization
47mEthical Machine Learning
38mMaking LLMs reliable - A practical framework
29mLocal LLM fine-tuning: a practical example
29mGenerative-AI: Usecase-Specific Evaluation of LLM-powered Applications
28mTrustworthy LLMs: Vibe checks are not all you need
48mPJ Hagerty: AI and Ethics - What We SHOULD Be Pondering