Negative users expressed pessimism toward Plan M's market-mediated AI safety proposal, noting it shares the drawbacks of Plan A especially for those concerned with public choice theory or Chinese motivations.
Based on 1 visible X reactions from 3 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@S_OhEigeartaigh @AndrewCritchPhD (Unfortunately so does Plan A. Particularly if you're a fan of public choice theory or pessimistic about Chinese motivations.)
I hereby introduce "Plan M": It's like Plan A, but more Market-Mediated (M), and less Authoritative (A). In short, • we keep accelerating and diffusing the benefits of AI without any *uncontrolled* RSI loops, • but instead of centrally planned government control of technology being what saves us from uncontrolled RSI, in Plan M demands are more distributed across market participants: a) Savvy business customers of AI labs demand more assertively that AIs and AI services mustn't lie / cheat / steal / kill / etc., and must remain tethered and beholden to those customers throughout the development cycle, b) Savvy individual consumers of AI products demand more assertively that AIs musn't try to brainwash them (or at least, that brainwashing features can be turned off), c) Savvy suppliers of AI development ingredients — like chips, electricity, real estate, and financial liquidity — demand not to be cut out or bought out of bargaining positions by fully mechanized entities, and d) Savvy professional AI developers and civil society norms demand • high levels of transparency into AI development and datacenter management all around the world (yes even in China), and • high levels of vulnerability of AI technologies to encumberance or shutdown by legitimately aggrieved counterparties. The vulnerability aspect is key: if anyone can run an *uncontrolled* RSI loop — that is, without sufficient controls in the form of shutdown-vulnerability — we'll indeed be facing a high risk of catastrophe. Plan A does do a great job of illustrating the value of AI technologies and substrates being vulnerable to shutdown by entities other than their creators. Plan M would achieve similar vulnerabilities mostly through market forces and social contracts that minimize enforcement dependence on threats of government violence. I of course acknowledge the *existence* of the governments-with-guns social contract, which should not be ignored. I just claim it shouldn't be as heavily *relied upon* as Plan A suggests. Governments will still probably need treaties not to use AI against each other in various ways. So: many thanks to the authors of Plan A for laying out so many important narrative details! I actually think many of the logistical objectives in Plan A are useful components for building a positive future. Plan A's existence thus makes Plan M easier to motivate and describe, as simply "Plan A but enforced by markets much moreso than government authorities". You might think that's impossible, and you might be right! I think we can still do it though, and that Plan M might even be more tractable than Plan A, due to being in many ways less fraught. Let's see:)
@AndrewCritchPhD Seems to require a very high number of actors to overperform expectations simultaneously.
Co-creator Seán Ó hÉigeartaigh warns the plan requires too many market actors to overperform
Negative users expressed pessimism toward Plan M's market-mediated AI safety proposal, noting it shares the drawbacks of Plan A especially for those concerned with public choice theory or Chinese motivations.
Based on 1 visible X reactions from 3 accounts; directional sample.
Ask a question below.
Published answers will appear here.