If ASI Is Coming Soon, A Superhuman Alignment Researcher Is Coming Sooner

OpenAI Researcher Roon posted this today: Plan A has always been to solve alignment, not by waiting around for human enhancement or whatever other capeshit, but in a bitterlesson-complementarity way, how everything else in modern AI has gone https://x.com/tszzl/status/2117377997166817550 This is good news for a positive singularity. There is a contradiction in a lot of the doomer arguments. Most of them treat it as a given that recursive self-improvement is close, and that it gets us AGI and then ASI rapidly after. But if RSI can produce a general superintelligence, why couldn't it produce something much narrower first, an AI that's superhuman at alignment research? It is significantly more difficult for a swarm of AI agents to create an ASI that poses existential risks than it is to create a narrow AI that has superhuman intelligence at alignment research. If RSI can be used for create the former, it can definitely create the latter as well. Dario's definition of powerful AI is a Nobel-level expert in basically every field. Being superhuman in one field is an easier target than being superhuman in all of them. So if ASI is coming soon, an automated alignment researcher should show up before it, not after. You can't have it both ways. Either AI progress is fast enough to be dangerous, in which case it's also fast enough to help solve the problem, or it's slow, in which case we have lots of time to prepare. The usual reply is "alignment research isn't easily verifiable." But neither is most frontier work in Biology, Chemistry or even Computer Science in general. If unverifiability stops AI from getting scary good at alignment, it should slow it down in those fields too, which would make existential risks significant lower than the doomers claim. How do people expect for AI to recursively self improve into an ASI that is intelligent enough to outsmart the most intelligent researchers in biology, chemistry, medicine, physics, but, for some reason, the job of an alignment researcher is somehow safely limited to human researchers? It doesn't make sense. If we are truly on the way to ASI, the alignment problem will be solved on the road between AGI and ASI. The doom case needs AI to be superhuman at everything except the one problem that would make it safe. I haven't seen a good explanation for why that would be the case. If you have one, I'd genuinely like to hear it.