NPPS Leadership Team

[Cedric Serfon <cedric.serfon AT cern.ch>]

Hi,

>Develop a prototype HENP LLM with current data (e.g. ATLAS, RHIC) and first generation feature extraction tools

Belle II will also start to evaluate LLM : we just proposed 2 tasks related to LLM to some Indian students (Michel can give you more details). Having a bunch of CPUs to train these models can be useful.

Regards,

Cedric

---

De : Phys-npps-mgmt-l <phys-npps-mgmt-l-bounces AT lists.bnl.gov> de la part de Brett Viren via Phys-npps-mgmt-l <phys-npps-mgmt-l AT lists.bnl.gov>
Envoyé : mercredi 20 mars 2024 20:48
À : Torre Wenaus <wenaus AT gmail.com>
Cc : Torre Wenaus via Phys-npps-mgmt-l <phys-npps-mgmt-l AT lists.bnl.gov>
Objet : Re: [Phys-npps-mgmt-l] Dollars on the floor and GenAI

 

I see.

>From my view there is a problem how to absorb the money fast enough.

Just having some cash won't help the LS4GAN part as everyone is
time-poor (and we lost Dmitrii Torbunov to CSI).

But, maybe there is a story involving buying a significant number of
GPUs to be an efficient shared resource for NPP AI/ML.  If that
direction sounds good I can try to give some more verbiage.


-Brett.

Torre Wenaus <wenaus AT gmail.com> writes:

> Money is in the hundreds of k I believe. Timescale is 'yesterday', not sure it's compatible with a fresh hire. We have a
> meeting this afternoon on DUNE spending ideas, should have a better picture after that. 
>
> On Wed, Mar 20, 2024 at 3:30 PM Brett Viren <bv AT bnl.gov> wrote:
>
>     Generally, this sounds really good.  But, I do wonder what is the amount
>     of money and what is the time scale.  Is this at the scale of hiring a
>     post-doc?
>   
>     -Brett.
>
>     Torre Wenaus via Phys-npps-mgmt-l <phys-npps-mgmt-l AT lists.bnl.gov>
>     writes:
>   
>     > Hi all,
>     > It seems there's FY24 money on the floor that the management is looking for ways to sweep up quickly into projects.
>     I did
>     > some opinionating with Hong that I'll share with you. You'll recall our GenAI boosterism slide we prepared for a DOE
>     > request
>     >
>     > https://docs.google.com/presentation/d/1dSNSrkrPBjEmQDmmfOgmObiIX9ZGVc3YSmojBzLFcZY/edit#slide=id.p1
>     >
>     > (I just shared the slide with our AI/ML mattermost channel, I should have done that earlier)
>     >
>     > My comments to Hong...
>     >
>     > I think injecting some effort on that slide's work program would be a great way to use some available money. It has
>     1-3
>     > year goals
>     >
>     >   • Near term: 1-3 years
>     >   
>     >       □ Complete R&D on using AI to draw trusted, quantified inferences from real data
>     >       
>     >       □ Develop a prototype HENP LLM with current data (e.g. ATLAS, RHIC) and first generation feature extraction
>     tools
>     >
>     > where the first one is building on Brett's LS4GAN work. His LDRD ends circa June. The work has been successful but
>     in some
>     > ways frustrating for Brett. Physics turned down the LDRD, it went through as CSI. So working on his idea has been in
>     the
>     > hands of -- good, no criticism -- CSI people who are more concerned about extracting CS papers than physics-directed
>     > development and application. It is work in an area that rightly is very high profile in scientific application of
>     AI, we
>     > should continue our activity, with a physics focus.
>     >
>     > And some support for the second bullet could get us into the LLMs for HENP game reflected in the Apr 9 meeting
>     > contribution!
>     >
>     > ...of course there'll be no shortage of proposals for spending money quickly, e.g. DUNE and EIC, but wanted to get
>     this in
>     > his ear. Comments welcome. DOE is now making noises there may be tens of millions to spend on AI *for experiments*,
>     not
>     > just ASCR, in FY25, it would be good for us to have ideas and plans-for-a-plan ready.
>     >
>     >  Torre
>     >
>     > --
>     > -- Torre Wenaus, BNL NPPS Group, ATLAS Experiment
>     > -- BNL 510A 1-222 | 631-681-7892
>     > _______________________________________________
>     > Phys-npps-mgmt-l mailing list
>     > Phys-npps-mgmt-l AT lists.bnl.gov
>     > https://lists.bnl.gov/mailman/listinfo/phys-npps-mgmt-l
>
> --
> -- Torre Wenaus, BNL NPPS Group, ATLAS Experiment
> -- BNL 510A 1-222 | 631-681-7892