Here are some hints to get you started:

https://x.com/therealkrantz/status/1892551363084574802?t=gGY7P-r-4ZvZA3iTclYwzw&s=19

All of them sound vaguely like talking points I've heard you say a lot but the original piece is so long and hard to understand that I'm not sure if the exact wordings of the options have been slightly changed so that they're narrowly not true anymore 🤔

@TheAllMemeingEye Happy to clarify any questions in the comments.

@Krantz could you write a 1-3 paragraph layman language explanation of the Krantz mechanism?

i.e. don't use any phrases you wouldn't expect an average English speaking person to already understand, certainly don't use any phrases you've invented or redefined, do a Toki-Pona-ing / Yudkowskian-tabooing ( https://www.lesswrong.com/posts/WBdvyyHLdxZSAMmoz/taboo-your-words ) if need be. Existing explanations you've shared seem to be either extremely long or include a bunch of opaque invented/redefined language e.g. krantz-x, constitution, ledger, collective intelligence etc.

@TheAllMemeingEye

It is a mechanism that assigns points to people for checking facts. It's not complicated. I'm trying to explain that "getting points for checking facts" (a sort of decentralized Metaculus for propositions) is the same thing as paying people to demonstrate publicly that they understand something.

This is how analytic philosophers argue.

It also produces mechanistically interpretable alignment data. We will need a lot of alignment data. It will be a valuable resource and I'd like future generations to be fairly reimbursed for it in a way they can understand.

What's confusing?

@Krantz thanks for explaining :)

A couple points of clarification:

• I know elsewhere you mentioned offering free bets on prediction markets that depend on a given piece of info as a concrete way to pay to incentivise someone to understand such info, and that you've suggested using your mechanism to teach the world about AI x-risk so they'll be motivated to prevent it, but do you have calculations that prove that this is a more cost efficient and robust way to prevent AI x-risk than the research, governance, and outreach work that Effective Altruism organisations are currently funding? Intuitively I would expect it to be many orders of magnitude more expensive than is practical, and that a significant portion of the population are unknowledgeable / inexperienced / uncaring enough that they would still reach false conclusions even if financially incentivised not to.

• You mention this mechanism producing mechanistically interpretable alignment data. Am I right in thinking this applies to the specific case where you're paying someone to understand what their own values are, and where presumably there's also a rule that they have to explicitly and truthfully write out their reasoning before they can get paid? If so, how do you incentivise the people to be truthful, how do you create a coherent set of objectives from the heavily conflicting and often diametrically opposed values of different people, and perhaps most crucially how do you make a mechanistically-uninterpretable machine-learning-based ASI remain truthfully aligned to the objectives you give it?