Opinion

What if superintelligence is just weak?

​In response to “2023 Or, Why I am Not a Doomer” by Dean W. Ball.Dean Ball is a pretty big voice in AI policy – over 19k subscribers on his newsletter, and a former Senior Policy Advisor for AI at the Trump White House – so why does he disagree that AI poses an existential danger to humanity? In short, he holds the common view that superintelligence (ASI) simply won’t be that powerful. I strongly disagree, and I think he makes a couple of invalid leaps to arrive there.Better Than Us Is EnoughHis main flawed argument is that he implies AI must be omnipotent and omniscient to wipe us out and then explains why that won’t be the case. He states: “one common assumption… among many people in ‘the AI safety community’ is that artificial superintelligence will be able to ‘do anything.’” He then argues that “intelligence is neither omniscience nor omnipotence,” and that even a misaligned AI with “no [..] safeguards to hinder it” would “still fail” because taking over the world “involves too many steps that require capital, interfacing with hard-to-predict complex systems.” But omnipotence or omniscience was never the requirement, it just needs to be smarter and better than us – humans.Think ForwardImportantly, it doesn’t actually take superintelligence to wipe out or disempower humanity. For me to imagine this, I simply need to think forward to the not-so-distant future. Imagine you get a tiger cub. Think forward to what the tiger will look like in a year and ask yourself: could it kill me in a year? Now do this with AI. Imagine the future with a billion robots, AI running the military, AI doing basically all jobs with perhaps some level of human oversight, AI running the media, biolabs, political and military decisions, critical infrastructure. That metaphorical tiger could kill us. Ball himself imagines a future where AI is “embedded into much of the critical infrastructure and large organizations in America, such that it is challenging to imagine what life would be like if Claude ‘turned off.’”Ball also discusses scenarios in which superintelligence has almost outlandish abilities, performing science breakthroughs without much experimentation. He focuses on Yudkowsky’s claim that “a sufficiently superintelligent AI system would be able to infer not just the theory of gravity, but of relativity” from a few frames of a falling apple, or that “bootstrap molecular nanoengineering.” Ball may be correct that these specific claims are wrong, but these are not load-bearing parts of any story for why AI might become dangerous. You don’t need to infer relativity from first principles to engineer a bioweapon. Notably, Yudkowsky himself has given other scenarios that do not require the AI to make scientific breakthroughs without experiments (see IABIED, chapter 2).If your response is “but there will be many AIs and there will be monitoring, so we’ll be safe,” then you’ve shifted to a different (and very flawed) argument.[1] The point is that clearly AI will be able to take over in the future if we haven’t aligned it well by then. In reality, it probably won’t take that long to largely automate all jobs and tasks, since it’s enough to achieve some combination of: secure power, enable actions in the physical world, get rid of or sideline humans. And once it reaches a critical capability level, the AI has to act fast because of competing AI projects that represent future rival agents.An Old Argument, Made WorseThe core of Ball’s case, that the world is simply too complex and chaotic for any intelligence to control, is not a new argument. Robin Hanson made a similar case in his 2008 Foom Debate with Yudkowsky: innovation is too distributed across many actors, no single AI can race ahead of all competitors fast enough to dominate. But Hanson more correctly understood that this is an argument about the speed and distribution of AI takeoff, not an argument against existential risk. Ball takes Hanson’s position and corrupts it by treating it as a refutation of existential risk from AI entirely.^The “many AIs and monitors” defense is pretty weak: unaligned AIs can cooperate with each other; monitoring can be evaded, there’s simply too much to monitor, AI doing the monitoring for us could itself be jailbroken or could cooperate with the systems it’s supposed to watch, and AIs can hide their reasoning through methods like steganography.Discuss ​Read More

​In response to “2023 Or, Why I am Not a Doomer” by Dean W. Ball.Dean Ball is a pretty big voice in AI policy – over 19k subscribers on his newsletter, and a former Senior Policy Advisor for AI at the Trump White House – so why does he disagree that AI poses an existential danger to humanity? In short, he holds the common view that superintelligence (ASI) simply won’t be that powerful. I strongly disagree, and I think he makes a couple of invalid leaps to arrive there.Better Than Us Is EnoughHis main flawed argument is that he implies AI must be omnipotent and omniscient to wipe us out and then explains why that won’t be the case. He states: “one common assumption… among many people in ‘the AI safety community’ is that artificial superintelligence will be able to ‘do anything.’” He then argues that “intelligence is neither omniscience nor omnipotence,” and that even a misaligned AI with “no [..] safeguards to hinder it” would “still fail” because taking over the world “involves too many steps that require capital, interfacing with hard-to-predict complex systems.” But omnipotence or omniscience was never the requirement, it just needs to be smarter and better than us – humans.Think ForwardImportantly, it doesn’t actually take superintelligence to wipe out or disempower humanity. For me to imagine this, I simply need to think forward to the not-so-distant future. Imagine you get a tiger cub. Think forward to what the tiger will look like in a year and ask yourself: could it kill me in a year? Now do this with AI. Imagine the future with a billion robots, AI running the military, AI doing basically all jobs with perhaps some level of human oversight, AI running the media, biolabs, political and military decisions, critical infrastructure. That metaphorical tiger could kill us. Ball himself imagines a future where AI is “embedded into much of the critical infrastructure and large organizations in America, such that it is challenging to imagine what life would be like if Claude ‘turned off.’”Ball also discusses scenarios in which superintelligence has almost outlandish abilities, performing science breakthroughs without much experimentation. He focuses on Yudkowsky’s claim that “a sufficiently superintelligent AI system would be able to infer not just the theory of gravity, but of relativity” from a few frames of a falling apple, or that “bootstrap molecular nanoengineering.” Ball may be correct that these specific claims are wrong, but these are not load-bearing parts of any story for why AI might become dangerous. You don’t need to infer relativity from first principles to engineer a bioweapon. Notably, Yudkowsky himself has given other scenarios that do not require the AI to make scientific breakthroughs without experiments (see IABIED, chapter 2).If your response is “but there will be many AIs and there will be monitoring, so we’ll be safe,” then you’ve shifted to a different (and very flawed) argument.[1] The point is that clearly AI will be able to take over in the future if we haven’t aligned it well by then. In reality, it probably won’t take that long to largely automate all jobs and tasks, since it’s enough to achieve some combination of: secure power, enable actions in the physical world, get rid of or sideline humans. And once it reaches a critical capability level, the AI has to act fast because of competing AI projects that represent future rival agents.An Old Argument, Made WorseThe core of Ball’s case, that the world is simply too complex and chaotic for any intelligence to control, is not a new argument. Robin Hanson made a similar case in his 2008 Foom Debate with Yudkowsky: innovation is too distributed across many actors, no single AI can race ahead of all competitors fast enough to dominate. But Hanson more correctly understood that this is an argument about the speed and distribution of AI takeoff, not an argument against existential risk. Ball takes Hanson’s position and corrupts it by treating it as a refutation of existential risk from AI entirely.^The “many AIs and monitors” defense is pretty weak: unaligned AIs can cooperate with each other; monitoring can be evaded, there’s simply too much to monitor, AI doing the monitoring for us could itself be jailbroken or could cooperate with the systems it’s supposed to watch, and AIs can hide their reasoning through methods like steganography.Discuss ​Read More

Related Posts

Leave a Reply

Your email address will not be published. Required fields are marked *