Sentient AI
As machine learning technology continues to push the boundaries of what we expect computers to be able to do, I think we’ll see more conversation about what sentience means and whether we’ve built a “sentient AI.”
I continue to think we’re nowhere near this milestone, but I can understand why some people might look at LLMs (Large Language Models) like the one powering ChatGPT and say “hmmm that’s starting to look kinda close.” Here’s one way I think about those claims:
LLMs and image generating models (like DALL-E, Midjourney, and Stable Diffusion) use related approaches to 1. “predict” a way to complete a piece of text, in one case, or 2. derive an image that relates to a text prompt, in the other.
If ChatGPT appears to be sentient to you, but DALL-E does not, can you identify what the difference is?
To be clear, there are structural differences between LLMs and generative image models, and I’m not an expert on the specifics, though I’ve read a couple papers, but if someone is claiming one is sentient, I think they should be able to point at what differences make it so. So far I can’t and I don’t think “it generates words instead of pixels” is enough.