I've really enjoyed reading these over the past while. Thanks so much for all your hard work.
I had a, probably not so significant question, about the "natural deterioration" thing.
In another part of the update, in the Europe part, talking about undersea cables rupturing, somewhere between 100-200 a year, would this rupture fit into one of these "natural deterioration" breaks? and was the 100-200 a figure for Europe, or globally?
Not super important question, just curious. If you had any places I could read more if you didn't wanna answer feel free to link them, too.
> As well as the risks associated with open sourcing powerful AI systems, DeepSeek poses a problem for those calling for the US to race ahead on AI against China, in a dash to superintelligence. Those calling for this, such as Leopold Aschenbrenner, have argued that the safest scenario is where the US has, maintains, and grows a “healthy lead” over China. This now looks significantly less tenable.
Not sure how you quantify "significantly less tenable", but
1.) isn't the difference in compute available still the largest factor, even assuming faster algorithmic convergence
2.) o1/r1 are far from superintelligent systems
3.) I still expect it to make a decisive difference if you can deploy 10 vs. 1000 vs. 1000000 instances of your superintelligent system
1) Differences in compute still matter, but if China is able to do things 100x cheaper, with 100x less compute, with labor that is earning much less than $1M/year, they can still match America. E.g., if you think about a possible Stargate project with $500B over four years, given algorithmic differences China could match this at $5B. In this story America has become fat and wasteful.
I think in a world where Biden's export controls really worked, maybe compute could still be the overriding factor, but it's unclear that we're still in that world.
2) This is not clear to me; they are clearly superhuman across some dimensions.
3) I don't think American labs have 1M x more compute than Chinese ones?
1. Presumably American labs will ~immediately match and soon likely exceed the algorithmic improvements (because the US has presumably way more high-quality human capital available to it)
2. I was thinking "superhuman across most dimensions relevant for surpassing each other in an economic competition". In my thinking that is still ~1-2 years to go
3. But in a few years? I understand the export controls to only really start biting in a few years. Maybe Epoch has projections of the compute gap?
I've really enjoyed reading these over the past while. Thanks so much for all your hard work.
I had a, probably not so significant question, about the "natural deterioration" thing.
In another part of the update, in the Europe part, talking about undersea cables rupturing, somewhere between 100-200 a year, would this rupture fit into one of these "natural deterioration" breaks? and was the 100-200 a figure for Europe, or globally?
Not super important question, just curious. If you had any places I could read more if you didn't wanna answer feel free to link them, too.
Thanks again!
Worldwide, accidents, see https://yle.fi/a/74-20125339 to start digging more.
> As well as the risks associated with open sourcing powerful AI systems, DeepSeek poses a problem for those calling for the US to race ahead on AI against China, in a dash to superintelligence. Those calling for this, such as Leopold Aschenbrenner, have argued that the safest scenario is where the US has, maintains, and grows a “healthy lead” over China. This now looks significantly less tenable.
Not sure how you quantify "significantly less tenable", but
1.) isn't the difference in compute available still the largest factor, even assuming faster algorithmic convergence
2.) o1/r1 are far from superintelligent systems
3.) I still expect it to make a decisive difference if you can deploy 10 vs. 1000 vs. 1000000 instances of your superintelligent system
1) Differences in compute still matter, but if China is able to do things 100x cheaper, with 100x less compute, with labor that is earning much less than $1M/year, they can still match America. E.g., if you think about a possible Stargate project with $500B over four years, given algorithmic differences China could match this at $5B. In this story America has become fat and wasteful.
I think in a world where Biden's export controls really worked, maybe compute could still be the overriding factor, but it's unclear that we're still in that world.
2) This is not clear to me; they are clearly superhuman across some dimensions.
3) I don't think American labs have 1M x more compute than Chinese ones?
1. Presumably American labs will ~immediately match and soon likely exceed the algorithmic improvements (because the US has presumably way more high-quality human capital available to it)
2. I was thinking "superhuman across most dimensions relevant for surpassing each other in an economic competition". In my thinking that is still ~1-2 years to go
3. But in a few years? I understand the export controls to only really start biting in a few years. Maybe Epoch has projections of the compute gap?
One of the few things I read pretty much immediately, thanks for your work!
Small error, executive order "reforming the federal government hiring process" links to a tweet about Deepseek
Thanks Max, fixed