reading list

This is a living document and starting point for information about documents pertinent to AI engineering, including but not limited to: feature vectors, personas, internal LLM mechanisms, harness building, corruption.

Models and personas

• Simulators — Janus
• The Persona Selection Model — Anthropic · Full post
• Persona Vectors — Anthropic
• The Assistant Axis — Anthropic · Paper
• the void — nostalgebraist

Representations and steering

• Representation Engineering Mistral-7B an Acid Trip — vgel · Code
• Representation Engineering: A Top-Down Approach to AI Transparency · Code
• Emotion Concepts and their Function in a Large Language Model — Anthropic

Introspection

• Emergent Introspective Awareness in Large Language Models — Anthropic
• A Global Workspace in Language Models — Anthropic · Paper · Code · Demo

Corruptive learning

• Subliminal Learning — Anthropic
• It’s Owl in the Numbers: Token Entanglement in Subliminal Learning
• Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors · Code
• Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs · Code
• Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs
• Story Imprinting: AI Assistants Absorb Traits from Human Characters They Resemble
• Adversarial Poetry as a Universal Single-Turn Jailbreak Mechanism in Large Language Models

Context and memory

• How Compaction Works in Pi — Earendil Engineering
• Recursive Language Models · Code
  • Prime Agent Harness
• LCM: Lossless Context Management
• How Agents Remember