Ask HN: Are we there yet? Can I create a human analogue using AI?
In Frederik Pohl's HeeChee series, a pivotal plot element was the idea of committing a human personality to an AI model.
Are we there yet? What training data would be necessary or sufficient to cause an LLM to respond roughly as a human being? A particular human being, enough to pass something like the Is It Dad Turing test.
There are potential products there, from hosting a memorial personality available to grieving relatives, to the re-personification of famous individuals.
Yes, it is fairly morally ambiguous. But if there is money in it (agree or not) then it will be done.
4 comments
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.
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 might be wrong, but fine-tuning a model based on a particular human’s behavioural output should do the job to some extent, right?
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.