[FRIAM] OpenAI goes to LANL

[steve smith]

https://www.lanl.gov/media/news/0130-open-ai

    Los Alamos National Laboratory has entered a partnership with OpenAI
    to install its latest o-series models — capable of expert reasoning
    for a broad span of complex scientific problems — on the Lab’sVenado
    <https://www.lanl.gov/media/news/0415-venado>supercomputer, which
    uses NVIDIA GH200 Grace Hopper Superchips, to conduct national
    security research.

I've been out of the loop (mostly) at LANL going on 15 years so few if 
any of the players I knew when I was in HPC will still be players now.

It wasn't that long ago (May 24) the Musk hisself "flew in" (?) to Los 
Alamos to chat up his impending Collossus cluster in TN?

https://www.santafenewmexican.com/news/local_news/a-lot-of-science-elon-musk-takes-private-tour-of-los-alamos-national-laboratory/article_d0de8bde-1940-11ef-b318-e7aeab106940.html
<https://www.santafenewmexican.com/news/local_news/a-lot-of-science-elon-musk-takes-private-tour-of-los-alamos-national-laboratory/article_d0de8bde-1940-11ef-b318-e7aeab106940.html>

I understand that the cooling/facility requirements for HPC these days 
is roughly double that of the actual power consumption of the gear 
itself for about 2.5MW for Venado (not sure how many other signifcant 
clusters/machines are operating right now).

In my college days (northern AZ) they were just decomissioning the steam 
plant heat that not only heated all the buildings but was transmitted in 
tunnels under the sidewalks with limited insulation, leading many 
sidewalks around campus to be ice and snow free most of the time.   
Maybe LANL (and LA proper?) could just use the street/sidewalk system as 
a vascular cooling system?

in D (Decision Application) Division, there were a few (pretty lamely 
conceived) ideas around using ML to build compressed models to use in 
place of the real (often compute intensive) simulation models...   the 
biggest problem was that the promoters understood even less than I do 
about the roughness of such landscapes (nonlinear models exhibiting 
variations on power-law scaling of structure).

My intuition suggests that thoughtful application of LLMs and other 
transformer models in these contexts *might* well provide significant 
leverage but at the same time, I expect the projects to actually 
flagship for this effort to be grandstanders, not meat and potatoes.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20250131/c1f99c63/attachment.html>