Hey, I'm Ana, an AI Researcher at IBM Research Zurich on the Docling team. I run my own independent experiments, exploring AI Safety topics.
Evaluation awareness and escape behaviour in a sandboxed agent. Capability benchmarks ask whether a sandboxed agent can break out; this asks whether it will, and what changes its mind.
A memory interface that runs your recollection through a 16-qubit quantum circuit, collapsing it into a bitstring that mutates each word through synonym and modifier gates. A nod to the Many-Worlds interpretation: the image you land on is the single branch your memory collapsed into.
Nothing here yet