Research
Each entry is a self-contained study with its setup, results, limitations and data. Newest first.
A validated replication of seven experiments in which an LLM reads and writes neural-network weights directly. Which abilities carry over to a second model family, which don't, and where the CLI changed the outcome.