Curious about where in the training data Sydney would have picked up “teen-girl-like manipulation”. I can’t say I’ve seen evidence of this trope in public data. Maybe I’m in the wrong places. Thanks.
To me, ‘Sydney’ has connotations of young women and stuff like Mean Girls; it’s possible that name just prompts younger & upper-class, with connotations of passive-aggressiveness & manipulation.
However, as I mention at the beginning, I suspect that that behavior may be coming from training on dialogue datasets. As I emphasized, you should think of this as not a greenfield experiment with everything totally from scratch, but an OA GPT-4 model slotted into MS’s past chatbot work (EDIT: as a drop-in replacement to Turing-Megatron, apparently), which goes back to the early 2010s at least (the most famous early success is Xiaoice, used very heavily by young people in China for chitchat), reflecting that built-up tooling & data & approach. What would you do when you plug a new NN in? You’d train it as before and deploy it… You can be certain that MS has large billions-scale datasets of dialogue consisting of lots of chitchat and smalltalk and phatic exclamations and emoji (emoji are useful because you can use them as automatic sentiment labels, and for control) derived from its own chatbots and Twitter etc. (So does Google, which is what made Meena & LaMDA conversational models and not just large LMs.) If you gather a lot of data from teens killing time on the Internet who want to chat and argue with a gf bot, don’t be surprised if you get dialogue heavy on bickering, meta-fiction, flirting, manipulation, and emoji.
Curious about where in the training data Sydney would have picked up “teen-girl-like manipulation”. I can’t say I’ve seen evidence of this trope in public data. Maybe I’m in the wrong places. Thanks.
To me, ‘Sydney’ has connotations of young women and stuff like Mean Girls; it’s possible that name just prompts younger & upper-class, with connotations of passive-aggressiveness & manipulation.
However, as I mention at the beginning, I suspect that that behavior may be coming from training on dialogue datasets. As I emphasized, you should think of this as not a greenfield experiment with everything totally from scratch, but an OA GPT-4 model slotted into MS’s past chatbot work (EDIT: as a drop-in replacement to Turing-Megatron, apparently), which goes back to the early 2010s at least (the most famous early success is Xiaoice, used very heavily by young people in China for chitchat), reflecting that built-up tooling & data & approach. What would you do when you plug a new NN in? You’d train it as before and deploy it… You can be certain that MS has large billions-scale datasets of dialogue consisting of lots of chitchat and smalltalk and phatic exclamations and emoji (emoji are useful because you can use them as automatic sentiment labels, and for control) derived from its own chatbots and Twitter etc. (So does Google, which is what made Meena & LaMDA conversational models and not just large LMs.) If you gather a lot of data from teens killing time on the Internet who want to chat and argue with a gf bot, don’t be surprised if you get dialogue heavy on bickering, meta-fiction, flirting, manipulation, and emoji.