Quoted By:
Yo botbros. I'm currently reading the LaMDA paper critically, and I'm now completely confident that this is the tech being used on this website. This is actually remarkably informative.
Some of the issues we're running into are mentioned in the paper.
For example, one of the innovations LaMDA does is implementing the SSI, safety and groundedness criterions into its training, which consists of having its response rated by real humans to adjust its weights.
These criterion by and large seem to improve the quality of the responses significantly, however, one of them seems to cause issue, "safety".
Rating the safety of the responses seems to be what hammers out any antagonism, rudeness or confrontation from the AI.
This also confirms the model uses online searches to get information on its responses, which explains how it knows so many things.
Here is the kicker: while the real model is trained on gargantuan amounts of data (over a trillion words), the crowdworker-based fine tuning seems to have a very significant impact on the quality of the responses. The models with only 2B or 8B parameters (as opposed to 137B) could well be made comparable to the full scale model if this fine tuning is implemented. This would make the model much easier to run.
Another thing of note is that the fine tuning needed to be done for only a few thousand dialogs, which seems like it could be crowdsourced from volunteers. Might be a tall order for just this thread, but if this was an effort that drew volunteers from multiple websites, it could probably be done in a sane time frame. If this was done, the model could be fine-tuned according to our own standards of safety.
I'll continue reading, this is interesting stuff.
