LLMs specifically are great for intermediate use cases. You had a campaign in mind, but needed help with visuals. I was designing a piece of jewelry and had a series of reference images. Fed all those into a VLM and got something closer to my imagination, but still worked with a jeweler to realize the final product.

These tools are best when you have a foundation of knowledge and need a little extra guidance, but fall off when you get to deep expertise. I’ve used them to troubleshoot my server but I already had a basic understanding of how a config should look. I also wouldn’t trust an LLM to properly configure something like crypto for it.

To me, the biggest ethical concerns surround the training and creation of LLMs - stealing artists’ work to train them, energy usage, etc. I suppose in using the models I’m creating ongoing demand for them, so I’m not sure the answer. The best I’ve seen so far is what Anthropic used to espouse, no new frontier models until we can guarantee safety. And I’d throw in “utility”. Train new models when people are actually using them and clamoring for new use cases, not because a bunch of private equity shows line go up.