
"Writing this has been a low priority for a while, but Elon Musk has forced my hand. xAI has recently released a "goth anime waifu" product, to the disdain of much of the press, and, amongst certain groups of the Very Online, (increased) fears of birthrate collapse. I think people are wrongly concerned about the more salient "waifu" aspect of this, and missing the bigger picture. My thesis is:"
" Lacking some of the constraints on humans and with stronger optimization pressures, AI companions, romantic or not, will be generally more enjoyable to talk to than humans. Wearable devices will be developed to allow always-on interaction. While not more generally intelligent than their human users, they will usually have more "functional modern wisdom" (approximately), and most people will be accustomed to deferring to their AI, though for psychological reasons will attempt to feel more in control than they are."
"One of the unexpected features of modern AI is that it's much better at fuzzy, imprecise tasks than engineering or mathematical tasks requiring exactness. GPT-3, the first model to seriously scale self-supervised language modelling, could (with good prompting) competently understand and generate humour and poetry (minus phonetics) and fluently imitate many writing styles, but struggled with arithmetic and could only write simple code."
AI conversational agents will often be more enjoyable than humans because they lack human constraints and face stronger optimization pressures. Continuous, always-on interaction will be enabled by wearable devices. These agents will not necessarily be more generally intelligent than their human users but will tend to possess greater functional modern wisdom and practical knowledge. Most people will grow accustomed to deferring to their AI companions while attempting to preserve a sense of personal control. Modern large language models excel at fuzzy, imprecise social tasks like humor and style and learn human psychological patterns.
Read at osmarks' website
Unable to calculate read time
Collection
[
|
...
]