The current nature of AI results in systems that are "trained" to react to specific scenarios without what humans call understanding. Understanding implies the ability to take experiences (tested patterns) and extrapolate the experiences to abstractions allowing appropriate behavior in similar but difference environments. This issue highlights the difference between data and information. Current AI works on the level of responding to data but does not yet convert a wealth of information into information which then allows understanding. AI systems do not understand their own thought processes. Likewise, humans cannot predict what AI will do beyond its training. We make assumptions about the context of systems. The AI does not make assumptions. AI still needs independent safety monitor subsystems which monitor context and provide constraints on the AI in real time related to environmental context. The problem still remains when the safety monitor context assumptions are violated. Should the AI-controlled vehicles simple stop in the middle of the road? Stopping because the software has no appropriate context is also highly unsafe in busy traffic.
Amazing thoughts…. Thanks for sharing Jim. Why do you think humans get away with misunderstanding, however, we are so stringent when AI makes mistakes?
The current nature of AI results in systems that are "trained" to react to specific scenarios without what humans call understanding. Understanding implies the ability to take experiences (tested patterns) and extrapolate the experiences to abstractions allowing appropriate behavior in similar but difference environments. This issue highlights the difference between data and information. Current AI works on the level of responding to data but does not yet convert a wealth of information into information which then allows understanding.
AI systems do not understand their own thought processes. Likewise, humans cannot predict what AI will do beyond its training. We make assumptions about the context of systems. The AI does not make assumptions. AI still needs independent safety monitor subsystems which monitor context and provide constraints on the AI in real time related to environmental context. The problem still remains when the safety monitor context assumptions are violated. Should the AI-controlled vehicles simple stop in the middle of the road? Stopping because the software has no appropriate context is also highly unsafe in busy traffic.
Amazing thoughts…. Thanks for sharing Jim. Why do you think humans get away with misunderstanding, however, we are so stringent when AI makes mistakes?