I agree with @Marcus Plutowski’s answer, with an addition: If we unpause too late, we also die.
I’m not sure of that statement, but Andrej Karpathy has asserted that he believes a ‘cognitive core’ model, stripped of memorized knowledge but capable of all forms of reasoning, should be possible in ~1-10B parameters. I’ve heard others throw around the idea that in principle there should be a python file <1M tokens in size sufficient to kick off RSI. EY said, a long time ago, that someday we may find it is possible to run an AGI on the equivalent of a 1995 laptop. How Full Butlerian Jihad do we imagine a Pause could be, in practice? Because if the answer is not 100%, then we’re buying time and talking price.
I’m not sure of that statement, but Andrej Karpathy has asserted that he believes a ‘cognitive core’ model, stripped of memorized knowledge but capable of all forms of reasoning, should be possible in ~1-10B parameters.
Two decades ago, I would have guessed you needed at least the equivalent of an RTX Pro 6000 to run AGI at near human speed. But now we have models like Qwen3.8 27B, which are highly capable and persistent coding agents, and it has remarkably wide knowledge of the world. And you don’t even need an RTX Pro 6000. It’s obviously not a fully general human replacement, but it’s a surprisingly large chunk of one. So if you’re willing to drop a bunch of world knowledge and add the missing capabilities, yeah, there’s probably some smallish constant multiplier of 27B that gets you there.
So I wouldn’t be entirely surprised if there exists a recipe for AGI (or even weakly superhuman AI) that would fit in a 2030 homelab. If this exists, it would be a recipe for ruin: a potentially world-ending technology that could be hosted by a single hobbyist.
This is the other reason[1] I would like to try to halt AI progress: If it’s possible to explain the key ideas needed to built an human-replacement AGI with 1 or 2 scientific papers, I would prefer to delay the writing of those papers for as many years as possible.
I agree with @Marcus Plutowski’s answer, with an addition: If we unpause too late, we also die.
I’m not sure of that statement, but Andrej Karpathy has asserted that he believes a ‘cognitive core’ model, stripped of memorized knowledge but capable of all forms of reasoning, should be possible in ~1-10B parameters. I’ve heard others throw around the idea that in principle there should be a python file <1M tokens in size sufficient to kick off RSI. EY said, a long time ago, that someday we may find it is possible to run an AGI on the equivalent of a 1995 laptop. How Full Butlerian Jihad do we imagine a Pause could be, in practice? Because if the answer is not 100%, then we’re buying time and talking price.
Two decades ago, I would have guessed you needed at least the equivalent of an RTX Pro 6000 to run AGI at near human speed. But now we have models like Qwen3.8 27B, which are highly capable and persistent coding agents, and it has remarkably wide knowledge of the world. And you don’t even need an RTX Pro 6000. It’s obviously not a fully general human replacement, but it’s a surprisingly large chunk of one. So if you’re willing to drop a bunch of world knowledge and add the missing capabilities, yeah, there’s probably some smallish constant multiplier of 27B that gets you there.
So I wouldn’t be entirely surprised if there exists a recipe for AGI (or even weakly superhuman AI) that would fit in a 2030 homelab. If this exists, it would be a recipe for ruin: a potentially world-ending technology that could be hosted by a single hobbyist.
This is the other reason [1] I would like to try to halt AI progress: If it’s possible to explain the key ideas needed to built an human-replacement AGI with 1 or 2 scientific papers, I would prefer to delay the writing of those papers for as many years as possible.
My first reason to halt is that recent frontier math results suggest we might be getting uncomfortably close to RSI.