Thoughts on AI safety in this era of increasingly powerful open source LLMs