Because utility likely scales a lot quicker as AI becomes able to do research autonomously. Pointing millions of GPUs at a research problem, when each of those GPUs can do better work than the best AI researchers... things could get fast fast. And AI doesn't need to be superhuman for this to be true. They don't need to be generally better than the best AI researchers in order to do better work than the besst AI researchers. The frontier of AI capabilities is jagged. They just need one important spike to extend beyond that of the best human researchers, and that will open up a domain in problemspace over which AI can problem-solve efficeintly and human experts cannot,

that will open up more areas of problemsapce chain reaction

10 million advanced GPUs running 24/7

vs maybe 10,000 top AI researchers at goog/anth/xai/oa/msft/amzn

1 gpu = 1 person

10,000,000 vs 10,000 is 1,000x bigger

we could be looking at liek a 100x speedup

What I learned about writing from working on my previous post

• The ending was rushed as I left it too late and got tired, leading to an unpolished post
• i liked it tho
• I guess this is my daily blog now, so I could have sections like ‘interesting stuff a saw today’ ‘main article body’
• it’s pretty much me chronicling the advent of superintelligence and/or how we paused and/or how i was wrong about everything ai timelines related
• it’s pretty important just to sit down and write without distractions nor diversions. I should set a timer for 1 hour and just do nothing but write for that hour, then take a short break, then do another hour, until the post is complete
• maybe every 7th post should be a larger-scale post sythesising my previous 6 posts and that could be posted on my substack
• OK, so let’s do the first hour now

First Hour

<aside> ⏱️ One hour timer started at 4:14 PM UTC - ends at 5:14 PM UTC

</aside>

I asked the Notion AI to set a timer for one hour. I have no idea whether it will function. The emoji clock does not inspire confidence.

So, the topic of this post is that AI research automation arrives earlier than full-replacement of AI researchers.

I’d be interested to see AI2027’s take on this… do they mention it, or is this jimsight?