Model Parameter Updates

Parameterization is needed for abilities that cannot be expressed through external signals: medical image recognition, natural prosody, removing the “AI flavor.” The solution is determined not by task stability, but by the nature of the ability.

The downside: long-term stable hard rules (bank transfers) cannot be left solely in parameters – deterministic server code is needed.

Before training: remove confidential data, filter erroneous trajectories, preserve an independent regression set. After – check for forgetting of general abilities and safety alignment.

Related: Four update carriers, Three post-training stages