saw a theory today that anthropic's models collapse if they overtrain on their own output, so the watermarks tell it not to re-ingest its own shit.
saw a theory today that anthropic's models collapse if they overtrain on their own output, so the watermarks tell it not to re-ingest its own shit.