New international consortium formed to create trustworthy and reliable generative AI models for science

A global consortium of scientists from federal laboratories, research institutes, academia, and industry has formed to address the challenges of building large-scale artificial intelligence (AI) systems and advancing trustworthy and reliable AI for scientific discovery.

The Trillion Parameter Consortium (TPC) brings together teams of researchers engaged in creating large- scale generative AI models to address key challenges in advancing AI for science. These challenges include developing scalable model architectures and training strategies, organizing, and curating scientific data for training models; optimizing AI libraries for current and future exascale computing platforms; and developing deep evaluation platforms to assess progress on scientific task learning and reliability and trust.

Toward these ends, TPC will:

• Build an open community of researchers interested in creating state-of-the-art large-scale generative AI models aimed broadly at advancing progress on scientific and engineering problems by sharing methods, approaches, tools, insights, and workflows.
• Incubate, launch, and coordinate projects voluntarily to avoid duplication of effort and to maximize the impact of the projects in the broader AI and scientific community.
• Create a global network of resources and expertise to facilitate the next generation of AI and bring together researchers interested in developing and using large-scale AI for science and engineering.

The consortium has formed a dynamic set of foundational work areas addressing three facets of the complexities of building large-scale AI models:

• Identifying and preparing high-quality training data, with teams organized around the unique complexities of various scientific domains and data sources.
• Designing and evaluating model architectures, performance, training, and downstream applications.
• Developing crosscutting and foundational capabilities such as innovations in model evaluation strategies with respect to bias, trustworthiness, and goal alignment, among others.

TPC aims to provide the community with a venue in which multiple large model-building initiatives can collaborate to leverage global efforts, with flexibility to accommodate the diverse goals of individual initiatives. TPC includes teams that are undertaking initiatives to leverage emerging exascale computing platforms to train LLMs—or alternative model architectures—on scientific research including papers, scientific codes, and observational and experimental data to advance innovation and discoveries.

Trillion parameter models represent the frontier of large-scale AI with only the largest commercial AI systems currently approaching this scale.

Training LLMs with this many parameters requires exascale class computing resources, such as those being deployed at several U.S. Department of Energy (DOE) national laboratories and multiple TPC founding partners in Japan, Europe, and elsewhere. Even with such resources, training a state-of-the-art one trillion parameter model will require months of dedicated time—intractable on all but the largest systems. Consequently, such efforts will involve large, multi-disciplinary, multi-institutional teams. TPC is envisioned as a vehicle to support collaboration and cooperative efforts among and within such teams.
 

“At our laboratory and at a growing number of partner institutions around the world, teams are beginning to develop frontier AI models for scientific use and are preparing enormous collections of previously untapped scientific data for training,” said Rick Stevens, associate laboratory director of computing, environment and life sciences at DOE’s Argonne National Laboratory and professor of computer science at the University of Chicago. “We collaboratively created TPC to accelerate these initiatives and to rapidly create the knowledge and tools necessary for creating AI models with the ability to not only answer domain-specific questions but to synthesize knowledge across scientific disciplines.”

The founding partners of TPC are from the following organizations (listed in organizational alphabetical order, with a point-of-contact):

AI Singapore: Leslie Teo

Allen Institute For AI: Noah Smith

AMD: Michael Schulte

Argonne National Laboratory: Ian Foster

Barcelona Supercomputing Center: Mateo Valero

Brookhaven National Laboratory: Shantenu Jha

CalTech: Anima Anandkumar

CEA: Christoph Calvin

Cerebras Systems: Andy Hock

CINECA: Laura Morselli

CSC - IT Center for Science: Per Öster

CSIRO: Aaron Quigley

ETH Zürich: Torsten Hoefler

Fermilab National Accelerator Laboratory: Jim Amundson

Flinders University: Rob Edwards

Fujitsu: Koichi Shirahata HPE: Nic Dube

Intel: Koichi Yamada

Jeülich Supercomputing Center: Jenia Jitsev

Kotoba Technologies, Inc.: Jungo Kasai

LAION: Jenia Jitsev

Lawrence Berkeley National Laboratory: Stefan Wild

Lawrence Livermore National Laboratory: Brian Van Essen

Leibniz Supercomputing Centre: Dieter Kranzlmüller

Los Alamos National Laboratory: Jason Pruet

Microsoft: Shuaiwen Leon Song

National Center for Supercomputing Applications: Bill Gropp

National Renewable Energy Laboratory: Juliane Mueller

National Supercomputing Centre, Singapore: Tin Wee Tan

NCI Australia: Jingbo Wang

New Zealand eScience Infrastructure: Nick Jones

Northwestern University: Pete Beckman

NVIDIA: Giri Chukkapalli

Oak Ridge National Laboratory: Prasanna Balaprakash

Pacific Northwest National Laboratory: Neeraj Kumar

Pawsey Institute: Mark Stickells

Princeton Plasma Physics Laboratory: William Tang

RIKEN Center for Biosystems Dynamics Research: Makoto Taiji

Rutgers University: Shantenu Jha

SambaNova: Marshall Choy

Sandia National Laboratories: John Feddema Seoul

National University, South Korea: Jiook Cha

SLAC National Accelerator Laboratory: Daniel Ratner

Stanford University: Sanmi Koyejo

STFC Rutherford Appleton Laboratory, UKRI: Jeyan Thiyagalingam

Texas Advanced Computing Center: Dan Stanzione

Thomas Jefferson National Accelerator Facility: David Dean

Together AI: Ce Zhang

Tokyo Institute of Technology: Rio Yokota