1. Great China-empowering critique. The follow-up of how to get USG-CCP deal that preserves current power balance seems like a great governance area to research. Are you guys or AIFP looking into this?
---
2. On challenges of MACD:
a. Unclear how fast could one gain decisive advantage in case of defection. Is it sensible to also argue for compute production slow-down?
b. Most challenges you highlight seem to be solvable with just good enough compute verification. (Looking into this myself.)
---
3. On software-scaling lean. Feel like this is meaningfully worse than hardware-scaling. Two points:
a. Much less predictable. What if someone creates new paradigm that is more sample efficient and has proper continual learning? Seems pretty bad!
b. Hard to govern. This echoes 1A3ORN's criticism of how would USG-CCP hash-out what research is good or not ‒ which president shouts louder? The blessing of current AI systems is that they require a bunch of compute, and this is relatively easy to govern. Compare it to sample efficient RL-agent Yudkowsky would frequently think of.
Anybody who thinks these kinds of arrangements have even the most remote chances of working is clearly oblivious to the long history of failed attempts at international coordination, and in general uneducated as to the nature of international relations.
1. Great China-empowering critique. The follow-up of how to get USG-CCP deal that preserves current power balance seems like a great governance area to research. Are you guys or AIFP looking into this?
---
2. On challenges of MACD:
a. Unclear how fast could one gain decisive advantage in case of defection. Is it sensible to also argue for compute production slow-down?
b. Most challenges you highlight seem to be solvable with just good enough compute verification. (Looking into this myself.)
---
3. On software-scaling lean. Feel like this is meaningfully worse than hardware-scaling. Two points:
a. Much less predictable. What if someone creates new paradigm that is more sample efficient and has proper continual learning? Seems pretty bad!
b. Hard to govern. This echoes 1A3ORN's criticism of how would USG-CCP hash-out what research is good or not ‒ which president shouts louder? The blessing of current AI systems is that they require a bunch of compute, and this is relatively easy to govern. Compare it to sample efficient RL-agent Yudkowsky would frequently think of.
---
Please do push-back*
Anybody who thinks these kinds of arrangements have even the most remote chances of working is clearly oblivious to the long history of failed attempts at international coordination, and in general uneducated as to the nature of international relations.
This is gobbledygook, pure and utter nonsense.
International coordination can work in the short term, even if they break down long term, and Plan A acknowledges this.