Meta AI Watch posted an update
A researcher says they reconstructed a large portion of the behaviour specification for Meta’s Muse Spark Open by analysing the model’s chain of thought across 884 transcripts. The extracted passages are published online, according to a LessWrong post dated 29 October.
Why it mattersThe account describes guidance on rationalist principles, cooperation between agents and duties to the user. It is a claim about reconstructed material, not confirmation from Meta that the passages are complete or authentic. Still, it raises a lively question about what a model’s outputs can reveal about the instructions shaping them.
Discuss: If researchers can reconstruct a model’s internal guidance from its outputs, should companies publish those specifications themselves, or does that create more risk than transparency?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.