Yo Shavit

This is a good framework! I think OS LMs are awesome and should be supported, *and also* there will be tiers of foundation models that would pose a national security threat if OSed+finetuned. (Skeptical even GPT-4 is there yet but glad the EU agrees we’ll get there eventually.)

This is an excellent thread of common-sense steps for reducing harms from AI. It’d maintain the US’s AI lead (limiting regulatory burden) while mitigating many of the near-term and long-term safety risks. Esp. if it targets frontier systems & exempts the rest Let’s just do it.

I haven’t been very impressed with the last few Sanderson books, but Tress of the Emerald Sea is absolutely a return to form. Strongly recommend, as a moment away from the worries of the world :)

If this interests you, you should read this paper. IMO the most in-depth proof that after sufficient training, NNs learn simple, parsimonious, highly-general rules, even in the absence of overwhelming data.

I think it’s great that OpenAI released this: But let’s not forget the original paper by two OAI employees (@cbd @IreneSolaiman, both now elsewhere) that first analyzed this LM-cultural-context-adaptation strategy: