Birds of a Feather Archive

How Must HPC Education Change for a World Where AI Generates Our Code?


Authors: Julia Mullen (MIT Lincoln Laboratory), Tim Mattson (Human Learning Group), Albert Reuther (MIT Lincoln Laboratory), Weronika Filinger (EPCC at the University of Edinburgh)

Abstract: In the near future (if not already) HPC software will be generated by AI. The question is not “if” we must adapt, but “how” we will adapt. Everything about our software will be impacted. In this BoF, we will focus on training: How does what we teach have to change to accommodate the necessary shift in skills in the AI era? We need ideas from across the community and people who will dedicate the time to debate these ideas, digest them, and produce a position paper to drive the change we need in HPC. It all starts with this BoF.

Long Description: Through academic courses, HPC summer schools, and conference workshops, the HPC education and training community has a long history of preparing people for careers in HPC. The workshops, tutorials and summer schools are designed to fill a gap between academia, where HPC skills are rarely taught, and professional research positions, where HPC skills are essential. A key learning objective is the ability to balance the scientific application concerns with knowledge of computer architecture to create performant, scalable programs. For decades this has translated to learning how to use new libraries, new languages, and new software tools effectively on new hardware. As AI code generation becomes more widespread, how much of this training remains relevant and what new training is required?

Beyond the hype of faster code development and shiny objects, what will AI-generated software mean for the HPC community? Consider the software development cycle; design, implement, validate. How does this change if AI takes over the implementation stage? Will the process speed up or will more time be required in design and validation, leading to a longer development process? How does this impact programmer skill development and HPC training? Do we know what AI generated code does well versus what it is bad at? How do we as educators develop recommendations and best practices to guide our students? As experts in the field we know how to evaluate and select the best algorithm for a given computer architecture and application, yet older algorithms that do not align well with modern architectures are still being taught in universities and included in AI training. The use of AI generated code expands the potential for performance issues due to poor choices. How do we educate users to discern whether or not the AI has chosen the correct algorithm and created efficient scalable code?

This BoF gathers the community together to discuss the necessary changes required to train HPC users in light of AI code generation tools. The BoF will be structured around discussion that focuses on several key questions: - What will AI never do for us? - What skills do humans need to validate the quality of AI generated code? - How will humans guide AI to map scientific problems onto the right algorithms for the actual hardware to be used?

For each key question, a speaker will describe the question and provide context. Then that speaker will lead a discussion of the question with the audience and a notetaker will record salient points from the discussion. The BoF will close with a call-for-action to form teams to work together to produce a position paper on necessary changes to HPC training in response to AI in software development. We will present this paper at SC27 with specific guidelines and recommendations for updating HPC education and training to address the challenges of generative AI.


Back to Birds of a Feather Archive Listing