[ExI] OpenAI Reaches A.I. Agreement With Defense Dept. After Anthropic Clash

[spike at rainier66.com]

-----Original Message-----
From: Keith Henson <hkeithhenson at gmail.com> 
Subject: Re: [ExI] OpenAI Reaches A.I. Agreement With Defense Dept. After Anthropic Clash

>...The best example of a US government project that was kept under wraps for a time was the Glomar Explorer and which eventually became public.
Look it up...

Keith, the Glomar Explorer was mothballed right up the street here, in Suisun Bay.  It was out there until about 2013.  I think it has been scrapped since then.  I recall the cover stories about the mission: the Glomar Explorer was said to be designed to recover manganese nodules from the ocean floor.  I knew it was a cover story, because manganese doesn't form nodules and isn't worth much anyway.




>...  Certainly, the US military has not built an AI training center on the scale Musk did in Memphis...

Keith that is an astute observation.  However... a military AI would not need to be on the scale of Musk's AI training center.  A military AI training center needs far less input material and has a far more limited scope.  It doesn't need to be brilliant at everything.  It only needs to be brilliant at defense.  Golden dome needs missile defenses that can handle multiple targets and react a lot faster than humans can.  Humans in the loop introduce risk in that task.

>...and it would be obvious from Nvidia's sales reports, on the ground, and from orbit....

Nvidia is making the fastest and best GPUs, but again, for some very specific tasks, an AI data center is not necessarily locked into that.  This is a good thing, for most of Nvidia's production is in Taiwan, which is likely to be attacked by China soon, which may end TSMC's output.  The Blackwell chips are being made in Arizona however, and there are other alternatives with US-based production.  You can be sure the military did not make themselves completely dependent on a chip being made in Taiwan.  They think of these kinds of things.

Keith regarding your father's experience with Sputnik, I would counter suggest the top brass already knew about Sputnik by other means, such as the same one the Russians used to learn of what we were doing at Los Alamos: insiders were exfiltrating the information.  They didn't want that information showing up in a report.


>...Building it in a remote location would be worse; first, you would have to build living space for construction workers.

Keith

The military already owns large buildings near the coast, close enough they could pipe in seawater to cool the condensers (both the steam condensers from the turbines and the freon condensers from the HVAC equipment needed to keep the buildings cool.  

I do find it most amusing to remember my first computer class I took while in high school at the local community college.  Their computer was one of the old-timey mainframes which still had vacuum tubes.  The operator stayed in there in a coat, for the room had to be kept at about 10 celcius.  That computer was outdated even in those days (mid 70s) but it worked well enough to teach the few students who were studying computers in those days.

Now, 50 years later, we have such advanced computing hardware, we are once again thinking of good old fashioned mechanical engineering solutions to extracting huge amounts of heat.  Looks to me like the best bet is to dump the heat directly into the sea, without a cooling tower, just a wildly scaled up condenser cooled by lots and lots of seawater.

Regarding that, I made two mistakes in my earlier estimates of cooling water flow for a military AI center.  I took the power output of the 20 nuclear reactors (a Musk-scale AI center) and estimating the water flow needed to carry away the resultant heat.  No one caught the two errors.  The 400 MW is the output of the nukes, so all that heat must be carried off from the processors.  But there is additional heat, because the condenser on the nukes also needs to be cooled.  So the system would be even bigger than I proposed.  However that was only one mistake... the other mistake was assuming any AI training center which wasn't bigger than Musk's and Bezos' data centers would not be worth a damn.  This is not right.  Any military AI center would intentionally not be on that scale anyway: that would be government competing against capital, which isn't legal.  The military's AI wouldn't need to be on that scale because it doesn't do what Grok, Claude and ChatGPT do.  The military's AI would be all about defense, only.  Smaller input, less processing needed, far less user demand on it, for it wouldn't have millions of proles asking it silly questions about the latest rock star or supermodel.

A military AI center would be far easier to hide than anything Bezos and Musk are doing, using existing military facilities and budgets, for it isn't on that scale.  It doesn't need to be.

spike