J-Lens
Can you read a model's next words from its middle layers?
We're testing whether a model's upcoming words can be read from its middle layers, with every pass mark written down first. In the runs so far, the Jacobian and logit lenses did not beat a plain word-frequency guess, and a trained lens beat it only for the very next token. A bigger 54-document map is now being built, to test whether more data improves the Jacobian lens.
Read the full breakdown heldContinuous Cognition
Can a memory-decay policy choose what to keep inside a fixed budget, without leaking what it should have dropped?
We're testing whether a memory policy can boil an agent's older conversation down to a fixed budget, keep the decisions that matter, and repeat no secret. In the latest run, three policies were tested on three made-up conversations each, and all three failed. The work is paused until the next candidate's pass marks are written.
Read the full breakdownWhen something we said turns out wrong, the correction is kept on the Corrections page. Earlier receipts are in the lab notebook.