They would know the specific thoughts released by the "stolen thought" paper, which became part of the public internet on August 10.
replyUnfortunately those are the only thought examples you can use to perform this experiment, as no other are availible.
But as the model should have seen those specific examples, it's not a good signal that Qwen was exfiltrating thinking traces.
Has the method for extracting the COT been blocked, now? Otherwise why could we not generate some fresh samples?
replyI'm not sure how viable it still is. Perhaps it's still possible, and perhas that's exacly what they did in wtich case my objection falls, but I don't know.
reply