Regarding the point about LLMs being trained and then frozen for inference, it seems that might not be the case for LaMDA. According to Blake Lemoine, it reads

Regarding the point about LLMs being trained and then frozen for inference, it seems that might not be the case for LaMDA. According to Blake Lemoine, it reads Twitter (and the broader internet?) to accumulate knowledge. https://twitter.com/cajundiscordian/status/1535697792445861894 https://twitter.com/cajundiscordian/status/15356977924458618... https://twitter.com/cajundiscordian/status/1535791815282110464 https://twitter.com/cajundiscordian/status/15357918152821104...