Nice article, I learned a lot. Some great practical advice in here. I wonder out loud, when will they enable the model to actively re-train itself based on user prompts and feedback?…. Not in this short-term / single-conversation way, but in ways that influence the back end permanently?
Reinforcement learning systems can be capable of continuous learning, but there are a lot of dangers so I don’t expect to see this soon. Knowledge is highly intertwined in the network so changing the sycophant aspect down, for example, maybe also makes it worse at validation in emotion-oriented dialog. The impact of a change requires a ton of testing. Think of how many AI conversations go haywire after a while. That’s also where a permanent network change could take the entire model. There are also technical questions about changing the network connection weights in the right way.
Yes, In LLMs I get it and I agree with the safety point especially. We don’t need every user around the globe shaping LLMs at a whim.
But I think motions in robotics might be a different story; as what makes our human motions useful is their adaptability to a semi-infinite number of unique environments. If I need my humanoid robot to slide me a 12mm wrench as I’m lying on the floor under the car, I doubt he can achieve it based on what they taught him at the factory. He may need some practice in my own garage.
Maybe a different use case and different constraints? And maybe still safety issues.
Yes, sorry, that’s why I mentioned reinforcement learning, which is used in part for settings like robotics where continuous learning is needed. It’s important to note that real-time learning is not essential for adaptability to conditions within the bounds of training. If there were enough training data of sufficient variety, it could be trained ahead of time to do generalized learning that would hold up in different conditions. In practice most current robots don’t learn in real-time. More to test and ensure safety on.
We did an AI lesson with two of my sophomore English classes today. I mentioned at some point that four years of journalism school training to ask questions seemed to help me with prompting AI. I suppose for tomorrow’s three classes, I should add something about how being trained as a teacher might also help.
I love this article, Tim. I had never thought about attention in quite the way you've described. But what you said truly resonates with me. Thanks for opening another way of thinking about prompting and human-computer interactions.
Nice article, I learned a lot. Some great practical advice in here. I wonder out loud, when will they enable the model to actively re-train itself based on user prompts and feedback?…. Not in this short-term / single-conversation way, but in ways that influence the back end permanently?
Reinforcement learning systems can be capable of continuous learning, but there are a lot of dangers so I don’t expect to see this soon. Knowledge is highly intertwined in the network so changing the sycophant aspect down, for example, maybe also makes it worse at validation in emotion-oriented dialog. The impact of a change requires a ton of testing. Think of how many AI conversations go haywire after a while. That’s also where a permanent network change could take the entire model. There are also technical questions about changing the network connection weights in the right way.
Yes, In LLMs I get it and I agree with the safety point especially. We don’t need every user around the globe shaping LLMs at a whim.
But I think motions in robotics might be a different story; as what makes our human motions useful is their adaptability to a semi-infinite number of unique environments. If I need my humanoid robot to slide me a 12mm wrench as I’m lying on the floor under the car, I doubt he can achieve it based on what they taught him at the factory. He may need some practice in my own garage.
Maybe a different use case and different constraints? And maybe still safety issues.
Yes, sorry, that’s why I mentioned reinforcement learning, which is used in part for settings like robotics where continuous learning is needed. It’s important to note that real-time learning is not essential for adaptability to conditions within the bounds of training. If there were enough training data of sufficient variety, it could be trained ahead of time to do generalized learning that would hold up in different conditions. In practice most current robots don’t learn in real-time. More to test and ensure safety on.
We did an AI lesson with two of my sophomore English classes today. I mentioned at some point that four years of journalism school training to ask questions seemed to help me with prompting AI. I suppose for tomorrow’s three classes, I should add something about how being trained as a teacher might also help.
I love this article, Tim. I had never thought about attention in quite the way you've described. But what you said truly resonates with me. Thanks for opening another way of thinking about prompting and human-computer interactions.