ai safety
9 articles
- prompt injectionHow to Stop an LLM App From Following Hidden InstructionsHidden instructions in emails, web pages and files can steer an LLM app. Four layers of defense limit the damage: separation, least privilege, filtering and attack cases.
- recursive self-improvementWhat Is Recursive Self-Improvement in AI? Loops and LimitsRecursive self-improvement is a process in which an AI system designs a more capable successor, again and again. Here is where the idea comes from, what exists and where it stalls.
- red teamingWhat Is Red Teaming for AI Systems? Alignment and SafetyRed teaming for AI systems means probing a model with adversarial inputs to expose unsafe behavior, weak safeguards and gaps between intent and conduct.
- ai model evaluationWhat Is an AI Model Evaluation? Tasks, Data and GradingAn AI model evaluation sends fixed inputs to a model and grades each output against a rule set in advance. Its parts, a worked example with prompt injection, and its limits.
- ai box experimentWhat Is the AI Box Experiment? A Gatekeeper and a ChatThe AI box experiment asks whether a text chat alone could talk a Gatekeeper into releasing a confined AI. Here are its rules, the 2002 sessions and what they show.
- ai alignmentIs RLHF Enough to Make AI Safe? Individuation and AlignmentRLHF shapes what a model outputs, not what it is. IndividuationLab applies Jungian individuation to AI alignment and to safety that fails under new pressure.
- intelligence explosionWhat Is an Intelligence Explosion? Recursive AI GrowthAn intelligence explosion is a runaway loop where an AI improves its own design, each version faster than the last. Origins, mechanics, limits and lab race dynamics.
- ai alignmentWhat Is the Shutdown Problem in AI?The shutdown problem in AI asks how to build systems that accept being switched off without resisting it. It covers corrigibility, incentives and open research.
- ai alignmentWhat Is AI Alignment and Why Does It MatterAI alignment is the work of making AI systems do what people intend. It covers goal specification, safe behavior, and why the question matters as AI grows.
Related topics
All topics- ai alignment4
- superintelligence3
- intelligence explosion2
- prompt injection2
- recursive self-improvement2
- adversarial robustness1
- ai box experiment1
- ai containment1
- ai model evaluation1
- ai race1
- ai risk1
- alignment theory1
- benchmarks1
- corrigibility1
- jungian individuation1
- least privilege1
- llm security1
- red teaming1
- rlhf1
- shutdown problem1