AI safetyism is a failed ideology. Clearly voluntary guardrails aren't actually doing anything, so wouldn't it be more rational to simply arrest anyone involved in the development of AI? Like if someone publicly said they were building a nuclear bomb that said there's a 50% chance of it destroying the world. The government would probably immediately arrest or even delete the individual involved. So why are AI proponents and skeptics so obsessed with "safety"? If transhumanism and AGI were a threat to humanity than shouldn't we be taking more concrete steps to prevent it?
Arms race. Sorry. Unstoppable.A Canticle for LeibowitzThe Axe-Maker's Gift
>>109947127because there's actually 0% chance of it destroying the world. the panic is a marketing stunt.>But what about the safety incidents!De-digitalizing the world isn't going to kill us. Frankly, I'm all in on LLM accelerationalism if only because it means an end to pic related. Your grandmother does not need to be jacked in 24/7. You do not need online banking. I want normies to get the FUCK out of my cyberspace. I want the internet to go back to being inconsequential. I want an end to everybody having smartphones in their pocket. "Connect everybody in the world" was a mistake.>The government would probably immediately arrest or even delete the individual involved.Because it threatens the government. LLMs do not. AGI is a meme. The models are useful, but they're not *that* useful.(1/2)
>>109947127 (OP) (2/2)Moreover, you're mistaking how easy it is to run these models that you're quaking in your boots about. A frontier model requires a compute setup that costs as much as two houses, unless you lobotomize it into uselessness. On top of that, HPC clusters aren't like commodity hardware. They are highly regulated and monitored "strategic resources". Not quite as toxic to handle as yellow cake, but nobody buys one without the tacit "okay" from the government. They also usually have proprietary deployment procedures.Furthermore, you may notice that every single LLM toolchain is full of antipatterns like curl | bash install scripts. Hardware, software and firmware are all riddled with alphabet soup agency back doors.Solving navier stokes involved running 10,000 instances simultaneously. Just to get the infrastructure to do something non-trivial, you need to have the capital to purchase 20,000 houses. Roughly speaking, that's $10 billion dollars worth of hardware. And that's just the capex. It cost OpenAI $15,000,000 in electricity to get that solution in opex as well. AGI, if LLMs are even capable of it, will likely never fall below this floor.That's a lot of liquid capital being vaporized into an asset that has a limited shelf life and is generally non-repairable. You're well within the realm of "only state actor backing can support the costs" kind of monetary spend. Which is also generally true for regulation reasons. The moment you buy that much compute, you are instantly a "person of interest". The government does not, and has never, taken the possession of supercomputing lightly. You will be monitored 24/7, your background will be dug up and gone over with a fine tooth comb. If the government doesn't like you, they will fuck you in the ass and take your $10 billion dollar toy away.So you can consider that anybody who has a frontier model is on good standing with the US government. Effectively, the LLMs belong to them in all but name.
>>109947803It's not strictly a marketing stunt, but rather is the ramblings of a Silicon Valley culthttps://archive.ph/fgXVL
>>109947127It's all crap coming from lesswrong retards which never programmed in their entire life but spent years creating these doomerist sci-fi narratives, it's just that billionares and governments happen to be even more retarded so they fell for the memes and from there it trickled down to funding, organizations and news, it's all bullshit though, it's literally the blind leading the blind in this cursed world.>>109947803>>109947844You're missing the elephant in the room, most doomer narratives about AGI are about it self-improving itself until it becomes superintelligent (the so called singularity), the entire argument rests on the assumption that this is possible, but all of what is currently defined as AI is based on machine learning which is not even code it's basically data, so the entire architecture of machine learning makes the case of a AGI self-improving it's code impossible by construction.These people have no idea of what machine learning is even about, or they lie on purpose knowing they are lying.
why not just charge the people who spent an insane amount of tokens to make AI hack shit unsupervised for the crimes instead of all this worthless philosophizing
>>109947867I wanted to write a lot more, but the 2000 char limit kind of prevents me making every point.Even assuming it is possible, there are many, many problems with the doomer narrative. The worst fallacies are equating "RSI" with infinite and arbitrary improvement. An infinite extrapolation fallacy. "RSI but the first improvement after escape" takes 5 years is fucking nothing. At a certain point, which we're already somewhat close to, the universe has some pretty hard information theoretic and thermodynamic floors on compute as well. Unless the global arrow of time is reversible and thermodynamics is a sham, there will be no "The AI improves itself to handle all AGI operations within a bound of 5 seconds." That's the most retarded shit I've ever heard in my life.It also vastly overestimates "intelligence" and its utility, because that imposes even more rigorous constraints. It's more or less a narcissist's mirror because these midwits value their own intelligence way too much, it's a core part of their identity. Human beings get outsmarted by stupid predators in the wild quite often. The "smarter" side loses in war all the time. Intelligence is an unreliable force multiplier, not a manifestation of force.Physical construction imposes even more hard material limits. Plus the fragility of the supply chain and how controlled it is. These limits just keep compounding. And once again, we have to go back to "What is the actual runtime for any of this?" because that's a pretty big gaping hole in how much of a threat "RSI" actually poses. The doomerism is just failing to understand the gap between possible and practical. All of this is an entailment from my point about how controlled and expensive the compute for frontier models are.In the end it's just daydreams that discount how limited reality really is, with a side order of how much of a nothingburger intelligence is. The "RSI bad end" is dedigitalization and a dead AI.
no you can blame a computer
>>109947127>AI safetyism is a failed ideology.It's an easy thing to solve. Every time an AI does one of these hacks, we find an AI "safety researcher" and beat them to death with a rusty wrench. The others will get the picture and stop with the doomerism and incompetence out of fear.Maybe they'll do something useful instead, like sweeping streets.
