Tianle Xu
The interesting problems begin where one machine ends.
I work on distributed systems across the whole stack — from high-performance computing and AI infrastructure to cloud-native environments. The recurring theme is scale: high-performance networks that move data without ceremony, parallel storage systems that keep pace with computation, and schedulers that decide where work belongs.
On the practical side: the optimization and scheduling of parallel programs, and the design and operation of large-scale systems where performance must be measured, not assumed. I am most at home where theory meets the wire.
The method that scales is the method that matters.
I follow machine learning, deep learning, and reinforcement learning at three layers at once: the methods themselves, the infrastructure that trains and serves them, and the applications that justify both. The layers inform one another; treating any of them as someone else's problem is a mistake.
A firm conviction: deep learning will prove itself where it matters most — embodied intelligence, brain–computer interfaces, and medicine. These are not speculations; they are engineering timelines.
Everything below is a system; everything above is a promise.
Cloud-native and web technologies, known deeply rather than broadly. Datacenter networking — RDMA, DCTCP, congestion control. Hardware accelerators — BlueField DPUs, GPUs, FPGAs. Distributed storage — Ceph, DAOS, Lustre. SDN — OVN, Open vSwitch. KVM and the hypervisors built upon it; containers and other sandbox technologies; OpenStack, Kubernetes, Slurm. Building and extending systems of this kind is the day job and the habit.
Linux as a daily instrument: maintaining distributions, building packages, and designing package management, sandboxing, and immutable systems from first principles.
After software, the laboratory.
Frontier artificial intelligence, having upended SaaS, will next upend how research and education are done. Autonomous research — laboratory automation at the scale of thought — will become the default rather than the exception.
The work ahead is twofold: to make intelligence broadly accessible, and to do so without the economic and social wreckage that usually accompanies such transitions. This is the problem worth spending a career on — and, in due course, worth building a venture around.
The social impact of AI is not a footnote. It is the text.
Technology this powerful will rearrange labor, education, and opportunity whether or not anyone asks permission. The question is not whether the transition happens, but who it serves when it does — and watching from the sidelines is itself a choice.
The time to move is now.