Perhaps this is related to their new "invisible watermark" concept which would probably require rather contrived language patterns to make possible.
Perhaps this is related to their new "invisible watermark" concept which would probably require rather contrived language patterns to make possible.
I love this theory. "We've invented a new invisible watermark that can detect whether code is LLM written."
The watermark: counting instances of 'load-bearing seam', 'the hard truth', 'and that's the whole point'.
I think opus was released before they included it on model. Its hard to say, but from what Ive read it doesn’t seem like it would have that drastic of an effect.
I had the same thought though.
The watermarking is independent of the model. The model itself has the probability weights to determine the next token. The watermark is similar to things like temperature and top_p/top_k in that the watermark adjusts the probabilities in a deterministic way that changes over time to hide tells from word choices.
If it's using Aaronson's approach it shouldn't have any noticeable affect on generations. When it picks between options weighted by probability after the generation of logits, it still follows the probability mass, it just uses a known pseudorandom seed so that when you go back and look at the exact choices you can fingerprint it.