When the Trillion Parameter Consortium holds its 2025 all‑hands conference and exhibition, TPC25, in San Jose on July 28–31, computational scientists, system architects, policy leaders, and AI stakeholders from research, industry, and government will tackle one question: how can trillion‑parameter frontier models be built, shared, and governed for everyday science?
Over three days, TPC25 attendees will dive into tutorials on building agents and a hands‑on hackathon. Plenary speakers such as Argonne’s Rick Stevens and RIKEN’s Satoshi Matsuoka will outline scientific AI priorities, while 30 breakout sessions will range from LLMs for cancer research to improving energy efficiency in supercomputing. The result promises to be part workshop, part strategy summit, and an early look at how open, frontier‑scale AI may reshape science in the next decade.
Agentic Tutorials and Hackathon
Organizers have framed the program around three straightforward goals: expand the contributor community through tutorials and a hackathon, offer a wider view in a series of plenary talks, and preview and create projects during a slate of breakout sessions.
For the first day and a half, there will be tutorials and hackathons to bring more hands on deck in the interests of making frontier AI more accessible for researchers. Running in parallel with the tutorials, the hackathon will pair newcomers with experienced developers in teams of three. Each group will build a simple agent within an agentic framework, such as a DOI lookup bot or a lightweight web search helper. This setup will ensure a healthy mix of skill levels that organizers hope will spark peer learning.
The focus on agents reflects how agentic workflows are becoming standard in science through platforms such as Microsoft Discover, OpenAI’s Deep Research, and Google DeepMind’s AI co-scientist. By the end of the session, participants should leave with practical experience they can apply to their own research.
Plenary Highlights: Leaders Map the Road Ahead
The plenary sessions will offer a larger view of the priorities of the global research computing community. Rick Stevens, associate laboratory director at Argonne and an AIwire Person to Watch, will open the plenary series with “Reinventing Discovery: Accelerating Science in the Age of Artificial Super‑Intelligence.” Stevens will discuss how frontier‑scale models have moved beyond existing as task‑specific AI tools to helping shape scientific questions and methods. In his opening talk, he will trace how rapid gains in hypothesis generation, simulation, and data interpretation now clash with slower steps such as chemistry, fabrication, and field work. Borrowing from Amdahl’s Law, he will outline the challenge as a rethink of the entire research workflow rather than a single leap in compute. Stevens will also tackle practical questions: how to set goals that reflect human priorities, keep results trustworthy when synthetic data is used, and decide which tasks still require human expertise. He plans to close with concrete actions the community can take to navigate this new landscape together.
AI researchers like Rick Stevens, Satoshi Matsuoka, and Charlie Catlett will be participating in TPC25. (Source: TPC25)
Thierry Pellegrino of AWS will make the case for hybrid HPC, or linking on‑premises systems with cloud capacity, as the practical path forward. His talk will show how AI tools, classical simulations, and early quantum resources can share one workflow, drawing brief examples from weather prediction, computational fluid dynamics, and molecular modeling that have sped up compute jobs without more spending.
Satoshi Matsuoka, director of RIKEN’s Center for Computational Science, will outline what still stands between large language models and everyday scientific work. Drawing on RIKEN’s AI4Science program and recent runs on the Fugaku and Frontier supercomputers, he will discuss methods for handling multi‑terabyte data, pairing neural surrogates with physics solvers, and spreading workflows across CPU, GPU, and early quantum hardware. Early results in climate modeling, 3‑D imaging, and materials discovery show up to 6.9x speed‑ups, and Matsuoka will close by flagging open questions around provenance and human oversight for petascale-and-beyond ML stacks.
Other plenaries will examine technical and ethical considerations as AI moves deeper into science. Ian Foster of Argonne will outline an “AI‑native discovery platform” that marries trillion‑parameter models to simulations, knowledge graphs, and autonomous labs. Ricardo Baeza‑Yates of the Barcelona Supercomputing Center will offer a survey of irresponsible AI in the form of bias and disinformation before mapping the governance and regulatory tools meant to curb it. Microsoft’s Preeth Chengappa will ask who holds agency when AI agents become ubiquitous, and what that means for ownership and control of results. Karthik Duraisamy (University of Michigan) will outline a proposed “active inference” architecture, a layered system that pairs models capable of counterfactual reasoning with continuous real‑world validation, to close three stubborn divides in research AI: the abstraction gap, the reasoning gap, and the reality gap.
Together, plenary sessions like these and the many others planned for TPC will help attendees evaluate the successes and challenges of frontier‑scale AI in scientific practice.
Breakout Tracks: Turning Ideas into Working Groups in BoF Sessions and Lightning Talks
(Source: Shutterstock)
The third goal of TPC25 is to drive focused collaboration through afternoon breakout sessions. There are 30 breakout slots spread across six rooms and five time blocks. Most sessions will follow a working‑group format: 90‑minute blocks that open with four to six lightning talks and leave time for focused discussion. Some tracks span several blocks, giving follow‑up sessions room to refine ideas.
Birds‑of‑a‑Feather meetups will run in parallel. Topic tracks cover both immediate research needs and long-term infrastructure questions. Domain science goes deep in the Life Sciences lane, where three sessions examine AI for cancer, agentic bio workflows, and foundation models for biology. Lightning talks will discuss topics like peptide‑binder design, genome‑scale language models, and precision‑health pipelines. Other sessions will tackle energy efficiency for AI at supercomputing sites and plans for a foundation model that represents the electrical‑grid distribution system. Organizers expect several of these informal meetups to mature into full TPC working groups once the meeting concludes.
Why TPC25 Matters for AI‑Driven Science
With practical sessions, plenary talks, and breakouts that could turn ideas into prototypes, TPC25 will provide a clear look at frontier AI for science and the people guiding its future. Tutorials and the hackathon will help attendees gain new skills, while plenary talks will outline AI’s influence on research. Birds of a Feather meetings and lightning talks will give projects the momentum they need to advance. For researchers, system builders, and anyone following AI’s impact on science, the San Jose gathering should not be missed.
To see an agenda and learn more about the Trillion Parameter Consortium and TPC25, visit https://tpc25.org and register at https://tpc25.org/register.