Both approaches - watermarking and 'requiring testing' seem pretty pointless. Bad actors won't watermark and tools will quickly emerge to remove them. The 'me

Both approaches - watermarking and 'requiring testing' seem pretty pointless. Bad actors won't watermark and tools will quickly emerge to remove them. The 'megasyn' AI that generated bioweapon molecules wasn't even an LLM and doesn't need insane amounts of compute.