I'd guess the training data would be a combination of baseline personality behaviors (turn-taking conversation, basic social skills, education level) with data specific to an individual. Recordings, letters, perhaps online posts but that's potentially not representative of the person since online is typically 'impersonal' by design.
Some of us may have too little to 'go on' to make a convincing avatar. Others may have biased samples, perhaps from lectures or recordings of performances.
Probably it may require an interview with those than knew them well. There would be memory bias, and the natural tendency to speak kindly of the dead. That might even serve the purpose better - it's a memorial after all, meant to be presentable and maybe not reflect the worst aspects of a personality.
Passing the “Is It Dad” test probably needs less public biography and more mundane private data—texts, voice notes, recurring jokes, and how he handled uncertainty.
I'd guess the training data would be a combination of baseline personality behaviors (turn-taking conversation, basic social skills, education level) with data specific to an individual. Recordings, letters, perhaps online posts but that's potentially not representative of the person since online is typically 'impersonal' by design.
Some of us may have too little to 'go on' to make a convincing avatar. Others may have biased samples, perhaps from lectures or recordings of performances.
Probably it may require an interview with those than knew them well. There would be memory bias, and the natural tendency to speak kindly of the dead. That might even serve the purpose better - it's a memorial after all, meant to be presentable and maybe not reflect the worst aspects of a personality.
Passing the “Is It Dad” test probably needs less public biography and more mundane private data—texts, voice notes, recurring jokes, and how he handled uncertainty.
Yeah, this feels like the kind of thing that is bad for the soul, or at least the abstract concept of what a soul represents.
I might be wrong, but fine-tuning a model based on a particular human’s behavioural output should do the job to some extent, right?