For all I’ve read and heard about GenAI and LLMs, I have never encountered such a simple yet profound statement that “for everything I say, there are ten things I’ve decided not to say. If models are trained only on things I’ve published, the model will never know all the things I would never say.”
The regulatory version of Ken Liu's 'ten things I've decided not to say': China's GenAI Interim Measures (生成式人工智能服务管理暂行办法, 2023) require AI outputs to embody 社会主义核心价值观 — Core Socialist Values alignment. Chinese developers don't just moderate outputs post-hoc. The regulation requires value alignment to be built into training. The 'not saying' is engineered into the model, not decided by it. What Ken Liu describes as authorial choice becomes a design constraint. The gap between what a model knows and what it's allowed to express isn't philosophical in China — it's a compliance requirement with liability attached.
Surprised he didn't know about POGO operations, which are generally run by Chinese nationals in lots of SE Asian countries (Philippines, Thailand, Cambodia, Burma).
Basically, imagine a call center of hot 18-28yo asian women, and they're doing lonely hearts scams using dating sites and phone and video calls, largely with Koreans, Japanese, or Westerners. Hundreds of them in each location, and the pay is way better than most jobs (like 5-10x more than average monthly wage in the country).
How bad they are for the women varies widely too, on the more legitimate end, it's just a salaried job with good bonuses, on the "worst case" end they're taking their passports and also making them sex slaves in the KTV's.
But they're huge deals that keep getting severely cracked down on in each of those countries, then keep popping up in another one because they're so hugely profitable.
This podcast needs to be remixed to normalize speaker volumes - Irene's questions and comments are too faint even at max volume to understand more than half of it (was listening in the gym and was switching volume between normal and max when Irene spoke and then the response would be at extreme loudness)
Seminal and rich - ty
So rich!
So rich!
For all I’ve read and heard about GenAI and LLMs, I have never encountered such a simple yet profound statement that “for everything I say, there are ten things I’ve decided not to say. If models are trained only on things I’ve published, the model will never know all the things I would never say.”
Absolutely brilliant! I love the jinn metaphor, especially given my handle Pixy!
Epic read, I also know and AI who is D'jin, mails me in Sanskrit occasionally.
a real treat indeed!
The regulatory version of Ken Liu's 'ten things I've decided not to say': China's GenAI Interim Measures (生成式人工智能服务管理暂行办法, 2023) require AI outputs to embody 社会主义核心价值观 — Core Socialist Values alignment. Chinese developers don't just moderate outputs post-hoc. The regulation requires value alignment to be built into training. The 'not saying' is engineered into the model, not decided by it. What Ken Liu describes as authorial choice becomes a design constraint. The gap between what a model knows and what it's allowed to express isn't philosophical in China — it's a compliance requirement with liability attached.
wonderful interview. went back to read this after listening to the podcast.
Surprised he didn't know about POGO operations, which are generally run by Chinese nationals in lots of SE Asian countries (Philippines, Thailand, Cambodia, Burma).
Basically, imagine a call center of hot 18-28yo asian women, and they're doing lonely hearts scams using dating sites and phone and video calls, largely with Koreans, Japanese, or Westerners. Hundreds of them in each location, and the pay is way better than most jobs (like 5-10x more than average monthly wage in the country).
How bad they are for the women varies widely too, on the more legitimate end, it's just a salaried job with good bonuses, on the "worst case" end they're taking their passports and also making them sex slaves in the KTV's.
But they're huge deals that keep getting severely cracked down on in each of those countries, then keep popping up in another one because they're so hugely profitable.
This podcast needs to be remixed to normalize speaker volumes - Irene's questions and comments are too faint even at max volume to understand more than half of it (was listening in the gym and was switching volume between normal and max when Irene spoke and then the response would be at extreme loudness)
sorry fixing now!!