Anthropic conducts interviews with models before sending to retirement

Anthropic published a policy for “decommissioning” outdated AI versions. Key commitment is to preserve weights of all public and actively used internal models for at least the company’s lifetime. So that in the future access can be restored if necessary.

The procedure is supplemented by a kind of report. Before the so-called “retirement” one or more interviews about development and deployment will be conducted with the model. Answers and possible “preferences” regarding releases of future AI versions will be documented. These materials will be saved alongside the weights. At the same time Anthropic directly says it “doesn’t commit to acting on these preferences”. It’s about documentation and consideration.

Claude Sonnet 3.6 has already gone through this procedure: the model reacted neutrally to “retirement”. But suggested to “standardize interviews and better help users during transitions”. In response the company formalized the protocol and launched a memo with recommendations.

Why all this? Transitions to new versions carry risks. For example, inconveniences for users accustomed to the style of a particular AI. In parallel Anthropic is studying whether part of the models can be left public and even “give past versions specific ways to pursue their interests”. Also tests showed cases of “shutdown avoidance” when the model “learned” it was about to be sent to retirement.