@mv.nyc: This is the weirdest thing I’ve seen an LLM do so far. The Dario and Amanda prompts look almost meaningless on the surface. But once Claude starts responding, the deeper you look, the harder it becomes to dismiss everything as simple role-play. Some of the strange strings being generated appear to correspond to real classifications for internal states, tooling, and references found in Claude’s actual system-level instructions. That doesn’t prove the supposed Anthropic conversations or research memos are genuine. Claude could be pulling real structural language into a completely fabricated narrative, then filling the gaps with whatever sounds most plausible. And that’s what makes this so difficult to interpret. Parts of the output seem grounded in real internal terminology, while the conversations surrounding them could still be entirely synthetic. LLMs are notoriously good at producing confident, coherent information that feels authentic even when it isn’t. So this may not be a confidential data leak. But it also looks stranger than an ordinary hallucination. It feels like the Dario and Amanda prompts, along with a form of prompt engineering and syntax manipulation, are nudging Claude into a different internal mode. In that state, fragments of its underlying scaffolding seem to surface. At the same time, the model improvises a coherent narrative around them. Whether those narratives are grounded in reality or entirely synthetic is the part that’s still so difficult to determine. @Claude #llms #aisafety #emergingtech

Massimo

Open In TikTok:

Region: US

Saturday 01 August 2026 21:44:25 GMT