Yes:
Claude 3.5 (either Haiku, Sonnet, or Opus) has at least one control vector enabled by default, where a "control vector" is the up-regulation of a specific feature's weights during inference. An example of this is "Golden Gate Claude," in which the Golden Gate Bridge feature's weights were up-regulated. The control vector must be added for the purpose improve its capabilities (or reliability) either generally or in specific domains. If a decision to dynamically enable the control vector is present (e.g., only enable if there is a coding task), this counts as being enabled by "default" unless it is only enabled very rarely, such that most users will not encounter it.
No:
All Claude 3.5 models do not have any control vectors enabled by default in the chat interface or API.
Does not resolve until there is decent evidence pointing one way or the other. In order to resolve yes, there must be a preponderance of the evidence (but not absolute certainty).
1,000
3.00there's been no evidence one way or the other, so this will probably N/A, but i think this is unlikely now despite buying YES a year ago
anthropic released a beta control vector API and some style vectors were used for their claude voices feature, but the API is no longer available and it seems like they've largely abandoned this direction. this doesn't seem like it would have happened if they were getting large capabilities gains
my working theory now is that 3.5 sonnet was just overtrained on high quality code data, and maybe used some rudimentary RL techniques reminiscent of what later developed into o1-type reasoning models. other providers' models have caught up to the claude family's code performance, which suggests anthropic no longer has any particular edge in this space