I like this idea, but it seems much better in practice to instead use some obscure but otherwise modern language like Welsh. I think this gets you most of the benefits with minimal substantially reduced cost.
We could try to only apply the obscure-language-only techique for some applications, but this effects all of pretrain, so it might increase cost by a factor of 2.
What I really like about ancient language is that there’s no online community the model could exploit. Even low-ressource modern languages have online forums an AI could use as an entry point.
But this consideration might be eclipsed by the fact that a rogue AI would have access to a translator before trying online manipulation, or by another scenario I’m not considering.
Agree with the lack of direct access to CoT being one of the major drawback. Though we could have a slightly smarter reporter that could also answer questions about CoT interpretation.
I like this idea, but it seems much better in practice to instead use some obscure but otherwise modern language like Welsh. I think this gets you most of the benefits with
minimalsubstantially reduced cost.One issue with this idea is that understanding what the model is doing via CoT or natural language communication bottlenecks will become worse due to lossy-ness in translation. I think if the AIs do most of their reasoning in CoT, this consideration is probably dominant and we should just use english.
We could try to only apply the obscure-language-only techique for some applications, but this effects all of pretrain, so it might increase cost by a factor of 2.
What I really like about ancient language is that there’s no online community the model could exploit. Even low-ressource modern languages have online forums an AI could use as an entry point.
But this consideration might be eclipsed by the fact that a rogue AI would have access to a translator before trying online manipulation, or by another scenario I’m not considering.
Agree with the lack of direct access to CoT being one of the major drawback. Though we could have a slightly smarter reporter that could also answer questions about CoT interpretation.