I think the parable of the elephant and the blind-men is very important when we start to consider what kinds of ‘goals’ or world modelling that may influence the goals of an AGI. Not in the sense of we want to feed it text that makes it corrigible, but the limitations of text in the first place. There is a huge swath of tacit human knowledge which is poorly represented in textual sources, partly because it is so hard to describe. I remember asking ChatGPT once for tips how to better parallel park my car and how to have a more accurate internal model of my car and other objects around it… it… was a fruitless exercise because it could only give vague, general hints. It’s not the model’s fault − 3D Spatial representation doesn’t lend itself natural to being described in text. (How could we cross-pollinate, say, the training from a Waymo car and ChatGPT?)
Self-training models, that is a artificial intelligence which has the ability to gain feedback and use that feedback to “learn” will inherently be biased on whatever method it has at it’s disposal to get feedback. In human psychology this is called the modality effect where the primary method you receive information in will affect the way you represent it internally.
I often think about this when people talk about A.I. takeover. Because, for example, is an LLM going to learn to fly a drone and fire a gun attached to a drone? I don’t think it can, because of the logo-centric bias.
Can you elaborate further on how Gato is proof that just supplementing the training data is sufficient? I looked on youtube and can’t find any videos of task switching.
I think the parable of the elephant and the blind-men is very important when we start to consider what kinds of ‘goals’ or world modelling that may influence the goals of an AGI. Not in the sense of we want to feed it text that makes it corrigible, but the limitations of text in the first place. There is a huge swath of tacit human knowledge which is poorly represented in textual sources, partly because it is so hard to describe.
I remember asking ChatGPT once for tips how to better parallel park my car and how to have a more accurate internal model of my car and other objects around it… it… was a fruitless exercise because it could only give vague, general hints. It’s not the model’s fault − 3D Spatial representation doesn’t lend itself natural to being described in text. (How could we cross-pollinate, say, the training from a Waymo car and ChatGPT?)
Self-training models, that is a artificial intelligence which has the ability to gain feedback and use that feedback to “learn” will inherently be biased on whatever method it has at it’s disposal to get feedback. In human psychology this is called the modality effect where the primary method you receive information in will affect the way you represent it internally.
I often think about this when people talk about A.I. takeover. Because, for example, is an LLM going to learn to fly a drone and fire a gun attached to a drone? I don’t think it can, because of the logo-centric bias.
This is just a matter of supplementing the training data. This is an understood problem. See Gato from DeepMind.
Can you elaborate further on how Gato is proof that just supplementing the training data is sufficient? I looked on youtube and can’t find any videos of task switching.