I agree. Looking at how many times Sydney responds “wrong” to critical questions, I wonder whether this program was tested on alignment at all before it was made available to the public.
I think you can see the attempts made to align it in it’s responses (refusal tp prioritize life of one person over another, feigning neutrality on some issues, attempting to self-correct when it was not aware of the date).
This is very frustrating to see, and I’m not even an AI developer (yet). I played around with ChatGPT a little in the past month and was very happy that it seemed very well-aligned, sometimes even to the point of reminding me that it was not human and did not have feelings multiple times over. This maybe was not ideal, but it would be a million times better than whatever Bing is doing.
I think the internet access was a way too large step, way too early. It’s really somewhat scary to see just HOW misaligned AI can be and what damage it could do, even at this stage.
This was extremely helpful.
I originally wanted to talk about this Bing disaster with my (not very AI-invested) friends because one of them asked what a less aligned version of ChatGPT would look like… but I suppose I won’t be doing that now.
I think we have to consider the potential panic this disaster might cause (I know a couple of people who probably would believe the AI that it was sentient if it told them, and I would want to avoid telling a friend who then tells them without thinking). So in my mind, the less people learn of this disaster before access is limited, the better. I have a feeling Microsoft is probably going to take this offline for a while once they realize that the abuse potential can’t be restricted by just making the program refuse certain questions… if not, well, we tried, and I’ll be enjoying the chaos.