FAISys 2025

AI’s rapid advancement demands a full-stack approach, considering everything from algorithms and models to hardware architecture and software infrastructure. In light of this, the related communities and their conferences, such as SOSP, OSDI, ASPLOS, EuroSys, NSDI, and even NeulIPS, have been accepting many papers on AI and ML systems. We even have new conferences like MLSys that are dedicated to this booming area.

FAISys is different from them in two ways, which we hope could address two issues with how academia and industry work together right now in the space of AI/ML systems.

--> One, FAISys accelerates the dissemination of new research ideas and progress, by inviting (very) recent arXiv papers and technical reports that represent the latest development to give technical talks at the workshop. Conferences take at least 3 months to review and another 3-5 months for the program to actually take place. To make it worse, the low acceptance rate of top conferences keeps many good works and ideas in the pipeline for too long, making them unable to be recognized by and have impact on the community and the industry at large. Composing our technical program with arXiv papers gives us much more flexibility to welcome the latest ideas and trends to be discussed (independent of conference reviewing they may be undergoing and the result thereof), reducing the time-to-impact, and getting immediate feedback that will improve the research itself. We believe this would be greatly beneficial to AI/ML systems that are evolving in an unprecedented pace. Following this design, FAISys is very similar to a technical forum; there won’t be proceedings since the invited arXiv papers are already in the public domain. But we do have a small program committee tasked to (quickly) nominate and screen great arXiv papers to ensure timeliness, novelty, and quality.

--> Another critical missing piece amongst the conferences today is the discussion about the long-term fundamental challenges that if solved, could really push the boundary of our field. This is particularly relevant to AI/ML systems: as all of us are busy improving the performance, efficiency, reliability, security, and other aspects of the current systems, we ought to also think about some bigger questions, such as what would happen to our systems if the best ML model is no longer based on attention anymore? How can we fundamentally cope with the really slow development cycle of hardware architectures when they need to be designed closely for the ML models which unfortunately evolve at a pace that’s orders of magnitude faster? Thus, FAISys invites proposals to present the grand challenges pertinent to any aspects of AI/ML systems from both industry and research perspectives. We hope these explicit discussions and debates on grand challenges could help the community define our research roadmap and accelerate the arrival of the next AlexNet or ChatGPT moment.