Jack Lindsey on X: "@pgodfreysmith Assorted thoughts: -- The "injected thought" detection task consists of two sub-tasks, the anomaly detection step (is there an injection or no) and the identification step (what was injected). It's now fairly clear that these are handled by (mostly, at least) distinct mechanisms" / X
Assorted thoughts: -- The "injected thought" detection task consists of two sub-tasks, the anomaly detection step (is there an injection or no) and the identification step (what was injected). It's now fairly clear that these are handled by (mostly, at least) distinct mechanisms -- see e.g. https:…
Assorted thoughts: -- The "injected thought" detection task consists of two sub-tasks, the anomaly detection step (is there an injection or no) and the identification step (what was injected). It's now fairly clear that these are handled by (mostly, at least) distinct mechanisms -- see e.g. arxiv.org/abs/2603.05414 and arxiv.org/abs/2603.21396 -- the former residing in somewhat earlier layers than the latter. -- The identification step of the task is closely connected to our recent global workspace paper transformer-circuits.pub/2026/workspace…. Our current understanding is that some (not…
saved by
related reading
- the j-lens: finding an llm's unspoken concepts / chiragctxnn.github.io
- A global workspace in language models \ Anthropicanthropic.com
- Verbalizable Representations Form a Global Workspace in Language Modelstransformer-circuits.pub
- Introspection Adapters: Training LLMs to Report Their Learned Behaviorsarxiv.org
- Emergent Introspection in AI is Content-Agnosticarxiv.org
- A Mechanistic Explanation of Prompt Injection (and why you should study roles) — LessWronglesswrong.com
- Circuit Tracing: Revealing Computational Graphs in Language Modelstransformer-circuits.pub
- Emergent introspective awareness in large language models \ Anthropicanthropic.com
- Tracing the thoughts of a large language model \ Anthropicanthropic.com
- In-context Learning and Induction Headstransformer-circuits.pub
- [2604.16812] Introspection Adapters: Training LLMs to Report Their Learned Behaviorsarxiv.org
- A Review of Anthropic's Global Workspace Paper — LessWronglesswrong.com