The other night I went on chat.lmsys.org and repeatedly had random models write funny letters following specific instructions. Claude and Llama were completely

The other night I went on chat.lmsys.org and repeatedly had random models write funny letters following specific instructions. Claude and Llama were completely useless and refused to do any of it, OpenAI's models sometimes complied and sometimes refused (it appeared that the newer the model, the worse it was), and everything else happily did more or less as instructed with varying levels of toning down the humor. The last thing the pearl-clutching pieces of crap need is more "safety."