4 ms·
Based on what they published about it [0], LaMDA is trained on a gigantic corpus of text to do basic text prediction, and then fine-tuned to generate responses
by tsimionescu 4y ago
Based on what they published about it [0], LaMDA is trained on a gigantic corpus of text to do basic text prediction, and then fine-tuned to generate responses to dialog, with additional care to (1) evaluate the safety/quality of responses it gives (e.g. to avoid recommending someone commits suicide, as previous ML chatbots sometimes did) and (2) to retrieve facts directly from external systems, so it is less likely to produce random realistic-sounding information.
Re-training a model of this size probably takes years of CPU-time, it's not a task that can be done often, so I very much doubt re-training it on the dialog produced so far would be a part of operating it.
Even if such re-training were done, I very much doubt it would result in the model knowing who a particular user it talked to before was, once the context was lost. This is precisely why they need to rely on external fact data-bases to get correctness: even though the model is trained on thousands of history books, it doesn't remember them in the way we do - to it, the phrase "Napoleon Bonaparte fought at Waterloo" is probably only mildly more likely than "Abraham Lincoln fought at Waterloo" is.
[0] https://ai.googleblog.com/2022/01/lamda-towards-safe-grounded-and-high.html https://ai.googleblog.com/2022/01/lamda-towards-safe-grounde...