SIGCOMM 2026 workshop: networking education should teach principles, not protocols

Larry Peterson and Bruce Davie, authors of the open source textbook Computer Networks: A Systems Approach, devoted their latest column in The Register to the SIGCOMM 2026 Education Workshop, provocatively titled „Networking Education in the Age of AI“. Peterson admits the title provokes him: he has repeatedly voiced scepticism about AI in education, and he studied during the previous „age of AI“ of the mid-1980s.

The workshop ran as a hybrid event, with professional camera operators and audio gear thanks to support from SIGCOMM. Peterson joined from eight time zones away, but the setup kept the large remote audience inside the discussion.

Design process, not artefacts

Peterson co-chaired an open mic session with Jim Kurose, whose textbook leads the field. Asked about the role of programming assignments, Kurose said exercises and Wireshark labs all matter, but the value of good animated slides should not be underestimated. Karen Sollins described modelling systems courses on law school case studies, where the professor assigns papers and students argue about which constraints shaped a design. The same instinct drives „Networking Unplugged“: pairs of students get a rope and design a way to send bits along it, discovering encodings such as NRZ and modulation.

What should remain at the core?

In a breakout on which topics to keep and which to rethink, Peterson raised QUIC as a motivating case: new and complex enough to be missing from many introductory courses. A student meeting QUIC for the first time should make sense of it by relating it to what they already understand — reliable transmission, congestion control, end-to-end security — which requires teaching the principles of networking and system architecture. The room came close to consensus: teach principles and the design process more than artefacts, and use a problem/solution approach rather than today's protocol set.

AI, abstraction and grounding

AI in teaching was discussed at length. One paper argued that AI coding agents let students take on bigger, more realistic system-building assignments, and suggested assessing learning through design discussions: test the learning, not the implementation. Peterson remains sceptical about LLM-generated code in production, but will suspend disbelief in the classroom and sees a case for open source models and shared research infrastructure.

His central claim is that educators must teach mental models. He quotes Rodney Brooks on the symbol-grounding problem: transformer-like LLMs manipulate symbols without grounding in the physical world, so they cannot reliably do what a robot does — and networks are no different. Abstraction, he adds, is worth teaching explicitly: hiding the details of the client-to-datacenter link lets students study a multipoint control unit (MCU) serving a large video conference, yet the same habit produced the well-known problems of HTTP over TCP.