Eliciting Latent Knowledge (ELK) - Distillation/Summary - AI Alignment Forum
This post was inspired by the AI safety distillation contest. It turned out to be more of a summary than a distillation for two reasons. Firstly, I think that the main idea behind ELK is simple and c…
x Eliciting Latent Knowledge (ELK) - Distillation/Summary — AI Alignment Forum Eliciting Latent Knowledge Research Agendas AI Frontpage 28 Eliciting Latent Knowledge (ELK) - Distillation/Summary by Marius Hobbhahn 8th Jun 2022 26 min read 2 28 This post was inspired by the AI safety distillation contest . It turned out to be more of a summary than a distillation for two reasons. Firstly, I think that the main idea behind ELK is simple and can be explained in less than 2 minutes (see next section). Therefore, the main value comes from understanding the specific approaches and how they interact
related reading
- Mediumai-alignment.com
- Alignment remains a hard, unsolved problem — LessWronglesswrong.com
- Towards a better circuit prior: Improving on ELK state-of-the-art — LessWronglesswrong.com
- Mechanistically Eliciting Latent Behaviors in Language Models — AI Alignment Forumalignmentforum.org
- Self-Fulfilling Misalignment Data Might Be Poisoning Our AI Modelsturntrout.com
- Eliciting bad contexts — LessWronglesswrong.com
- Mechanistic anomaly detection and ELK — LessWronglesswrong.com
- AI Control: Improving Safety Despite Intentional Subversion — LessWronglesswrong.com
- Incriminating misaligned AI models via distillation — LessWronglesswrong.com
- Evaluation-Conditioned Training: Teaching Models to Generalize to Stronger Oversight Regimesarxiv.org
- Auditing language models for hidden objectives — LessWronglesswrong.com
- Bitter Lessons from Distillation Robustifies Unlearningbrucewlee.com