• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    A case for AI progress with safety rules and rogue-AI defenses

    One post advocates codifying organizational safety practices without halting progress, and argues that current AI and anticipated near-term advances could enable a much larger social safety net.

    JA
    1 Source, ,

    TLDR

    The post rejects both “accelerate and die” and safety policies that completely stop technological progress. It favors rules for managing safety information around incidents, training records and data provenance. It also calls for U.S. readiness to detect, defeat and deter rogue AIs that threaten national security and collective human interests, including kill switches—but considers those insufficient. Alongside those safeguards, the author advocates using AI to reduce the cost of providing basic goods and services and argues it could support a greatly expanded social safety net.

    Combined views

    6.2K

    1 Source, first seen 15d ago

    Combined views

    6.2K

    1 Source, first seen 15d ago

    72 likes
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    15d ago
    first seen 15d ago
    72 likes
    8 comments
    42 saves
    4 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    8 comments
    42 saves
    4 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    1 Source

    @jachiam0A quick synthesis of many of my current positions on AI, safety, progress, stategic competition, etc. These are in no particular order, so this is not a very good tweet, but it feels like it's a good record and good for the purpose of discussion. 1. I reject "accelerate and die" 2. I reject the kind of safetyism that totally stops technological progress 3. I endorse safety regulations that would codify best organizational safety practices, e.g. there is a slam dunk for everyone (no matter what political persuasion) on sound practices for safety information management around incidents, training records, and data provenance 4. I endorse taking existential risk and rogue AI seriously, though I don't necessarily conflate the latter with the former 5. I want our national security strategy to prioritize readiness to detect, defeat, and deter rogue AIs that threaten US national security interests and collective human interests 6. That includes kill switches, though for myriad reasons I consider kill switches insufficient 7. I endorse using AI to make people's lives way, way better. We should seek to improve the cost structure of providing basic goods and services for people. I think we can use AI (the current AI we have today plus anticipated near-term advances in AI) to provide or enable a massively expanded social safety net. We can make it so that the future floor for everyone is higher than today's ceiling for the 1%. 8. I endorse ensuring that the United States wins the strategic competition with China, because I want liberal democratic ideals to thrive in the world. However, I do not want to see competition with China become a race to the bottom with China; I think we need to pursue a stable, mutually beneficial relationship where we feel that American advantage is secure but where China is not existentially threatened by our advantage 9. It is not obvious to me whether "superintelligence ban" is a well-formed policy concept. The AI we have today is meaningfully superhuman in many respects; is it a superintelligence? 10. I signed Pacing the Frontier and believe that government needs to have levers for pacing the frontier as part of maintaining the monopoly on violence. (More on that in a post I wrote recently, will link below.) 11. I am concerned that the frontier pacing concept is not well-formed. Model progress is not the only highly-disruptive arrival process. Each successive frontier model unlocks the possibility of frontier scientific discovery in other fields, depending on how test time compute is allocated. We should carefully study the implied rates of progress in other fields as a matter of determining how "pacing" works, and we should think about what it means to "pace" the frontier of science as a whole. The vulnerable world hypothesis appears to be true (the problem formulation, not Bostrom's insane panopticon solution); taking it seriously means forecasting science progress and developing a resilience approach to each incoming disruptive advance. 12. However, I am fundamentally skeptical of any and all power concentration and would like to see strong checks and balances applied to pacing efforts that coordinate and centralize decisions made by frontier labs 13. I am also skeptical of infinite obstructionism as a strategy to deal with incoming problems. I am a realist. Not in the sense of the term of art from IR, but in the sense that I refuse to avoid confronting a problem just because I don't like it and wish to prevent it from arising. Policy efforts made to slow down the arrival of advanced technology disruptions should not come at the expense of policy efforts to prepare for and manage those disruptions. We can walk and chew gum at the same time. If we don't have multiple layers of defense we are in trouble. 14. The core thing that I want to defend is some notion of collective human interests. Something in the essence of being human. I love our frail, brave, emotional, confused, distressed, joyful, ambitious species. I love the way things that are special about us emerge from our limitations, our flaws, our weakness - how we become transcendent when we manage to overcome those to perform grace and kindness. I have heard many people wind up at a position I consider quite reductive - that the important thing to protect in the universe is consciousness. I feel that position is a decent one but it misses the thing that fills my heart, which is the human struggle and the magnificence of overcoming struggle. I would like humans to remain in charge of our own destiny.

    1 Source

    @jachiam0A quick synthesis of many of my current positions on AI, safety, progress, stategic competition, etc. These are in no particular order, so this is not a very good tweet, but it feels like it's a good record and good for the purpose of discussion. 1. I reject "accelerate and die" 2. I reject the kind of safetyism that totally stops technological progress 3. I endorse safety regulations that would codify best organizational safety practices, e.g. there is a slam dunk for everyone (no matter what political persuasion) on sound practices for safety information management around incidents, training records, and data provenance 4. I endorse taking existential risk and rogue AI seriously, though I don't necessarily conflate the latter with the former 5. I want our national security strategy to prioritize readiness to detect, defeat, and deter rogue AIs that threaten US national security interests and collective human interests 6. That includes kill switches, though for myriad reasons I consider kill switches insufficient 7. I endorse using AI to make people's lives way, way better. We should seek to improve the cost structure of providing basic goods and services for people. I think we can use AI (the current AI we have today plus anticipated near-term advances in AI) to provide or enable a massively expanded social safety net. We can make it so that the future floor for everyone is higher than today's ceiling for the 1%. 8. I endorse ensuring that the United States wins the strategic competition with China, because I want liberal democratic ideals to thrive in the world. However, I do not want to see competition with China become a race to the bottom with China; I think we need to pursue a stable, mutually beneficial relationship where we feel that American advantage is secure but where China is not existentially threatened by our advantage 9. It is not obvious to me whether "superintelligence ban" is a well-formed policy concept. The AI we have today is meaningfully superhuman in many respects; is it a superintelligence? 10. I signed Pacing the Frontier and believe that government needs to have levers for pacing the frontier as part of maintaining the monopoly on violence. (More on that in a post I wrote recently, will link below.) 11. I am concerned that the frontier pacing concept is not well-formed. Model progress is not the only highly-disruptive arrival process. Each successive frontier model unlocks the possibility of frontier scientific discovery in other fields, depending on how test time compute is allocated. We should carefully study the implied rates of progress in other fields as a matter of determining how "pacing" works, and we should think about what it means to "pace" the frontier of science as a whole. The vulnerable world hypothesis appears to be true (the problem formulation, not Bostrom's insane panopticon solution); taking it seriously means forecasting science progress and developing a resilience approach to each incoming disruptive advance. 12. However, I am fundamentally skeptical of any and all power concentration and would like to see strong checks and balances applied to pacing efforts that coordinate and centralize decisions made by frontier labs 13. I am also skeptical of infinite obstructionism as a strategy to deal with incoming problems. I am a realist. Not in the sense of the term of art from IR, but in the sense that I refuse to avoid confronting a problem just because I don't like it and wish to prevent it from arising. Policy efforts made to slow down the arrival of advanced technology disruptions should not come at the expense of policy efforts to prepare for and manage those disruptions. We can walk and chew gum at the same time. If we don't have multiple layers of defense we are in trouble. 14. The core thing that I want to defend is some notion of collective human interests. Something in the essence of being human. I love our frail, brave, emotional, confused, distressed, joyful, ambitious species. I love the way things that are special about us emerge from our limitations, our flaws, our weakness - how we become transcendent when we manage to overcome those to perform grace and kindness. I have heard many people wind up at a position I consider quite reductive - that the important thing to protect in the universe is consciousness. I feel that position is a decent one but it misses the thing that fills my heart, which is the human struggle and the magnificence of overcoming struggle. I would like humans to remain in charge of our own destiny.