Plan A’s verification system is uniquely good, but Appendix L shows a different weakness: the evidence can appear and the people in charge can still explain it away. I wrote a response about what a behavioural standard for those decision-makers might add, along with how Appendix V places AI welfare inside the larger control system: [What Plan A Cannot Verify] (https://forum.effectivealtruism.org/posts/z2D3nWrJDv4FpkHsS/what-plan-a-cannot-verify)
AI 2040: Plan A is the new scenario from the team behind AI 2027[1]. Instead of a straightforward prediction of the future, it provides a best guess positive vision for how humanity should navigate the creation of super-intelligence.
I've only read a bit of it, so I'm not sharing any takes yet. I'm posting this so that there is a place on the EA Forum to discuss the essay/micro-site/project as we read it.
Read AI 2040: Plan ASpecifically, the named authors are: Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt, Daniel Kokotajlo
Always struggled to understand who these are supposed to be for. Politicians and policymakers don’t know or care about what an acausal trade is, and I suspect a lot of this piece is just too ‘out there’ to be persuasive to someone who spends most of their day doing ordinary politics. Is it supposed to be for the people who will then persuade the politicians? Something else? Very confused
My guess is that politician's staffers read it, and also that it elevates the status of the authors such that they are likelier to be taken on to advise. I presume it's less likely Daniel Kokatajlo would have been interviewed by Bernie Sanders if he hadn't written AI 2027.
The smart thing in the framing here is that politicians will soon (hopefully) be reaching for more comprehensive plans on what to do about AI. And this is called 'Plan A'. So it's not crazy to think that something like this may end up influencing legislation.