The open source AI community thrives on collaboration, choice, and open contribution. As Joseph Groenenboom from Red Hat highlights, empowering developers to freely contribute and build options is essential for growing the entire AI ecosystem. PyTorch Conference North America 2026 is where that community comes together in person. Join... us to collaborate with open source pioneers, expand your deployment capabilities, and directly contribute to shaping the future of machine learning. Register for PyTorch Conference North America today: #PyTorchConshow more

@RedHat conference season is honestly the best way to tell which frameworks are still alive and which ones quietly died

Been doing some serious harness engineering around scientific AI. The basic architecture is: Python scientific core + Snakemake + uv + reproducible containers + pytest/Hypothesis + Ruff/mypy/Pydantic + HDF5 + RO-Crate + SLSA/Sigstore + read-only RAG + ParaView + OpenUSD + SALib/OpenMDAO + FEniCSx + selective Rust + FMI later. The point isn’t to throw a bunch of tools together. I’m separating responsibilities so no single part of the system—especially the AI—gets to both produce a scientific result and declare that result trustworthy. The Python layer contains the actual numerical physics. uv locks the environment, Pydantic governs scientific schemas and parameters, Ruff/mypy catch structural problems, and pytest/Hypothesis test both software behavior and physical invariants such as conservation, bounds, convergence and impossible states. HDF5 stores the actual scientific outputs, while Snakemake makes the computational dependency graph explicit rather than hiding the whole experiment inside one giant script. Above that is a separate trust/reproducibility layer. Containers capture the execution environment, RO-Crate records provenance, and SLSA/Sigstore plus detached hashes make it possible to verify where an artifact came from and whether it has been altered. The AI side is deliberately separated from scientific authority. AI/RAG can retrieve evidence, reason about failures, suggest parameter changes, propose models and generate candidate modifications—but it cannot silently change authoritative scientific state or certify its own result. Conceptually I’m aiming for: AI proposes → deterministic computation executes → independent verification decides. FEniCSx is also there as an independently implemented numerical benchmark rather than letting the primary solver effectively validate itself. SALib/OpenMDAO handle sensitivity and optimization, ParaView handles scientific fields, OpenUSD represents the machine/system, Rust is reserved for places where it actually buys something, and FMI comes later for external model coupling. I’ve had two different reactions to this architecture. One is that treating reproducibility as a first-class requirement is exactly what serious scientific AI needs. The other is that the minimum stack should stay closer to Python + uv + pytest + HDF5 initially, with workflow/provenance/supply-chain infrastructure added as complexity demands it. I think there’s probably a distinction between the right end-state architecture and the right implementation order. Curious how others building agent systems are handling this boundary: where do you draw the line between what an agent is allowed to propose and what it is allowed to treat as authoritative?

@RedHat

@RedHat Open collaboration is what keeps AI moving forward. Excited to see what comes from PyTorchCon.