Where does systems and networking go in the age of AI?

Bogdan Stoica, a postdoc at Grainger College of Engineering’s Siebel School of Computing and Data Science, was halfway through his PhD when “AI happened.” The AI boom changed the focus of his research on analytical tools for software testing. He now uses a combination of traditional software testing and monitoring techniques with AI approaches.  

Stoica is hardly the only person trying to understand what the AI revolution means for academics entering the job market. To help figure it out, the Siebel School’s Systems and Networking Group held a retreat at Discovery Partners Institute (DPI) in April devoted to AI+Systems. Stoica’s remark was part of his lightning talk on his research.

Organized by Siebel School Associate Professor Tianyin Xu, one highlight of the day was SysNet’s first industry panel, giving students a chance to hear directly from professionals about the challenges they face in industry and the future they see for students in this fast-moving age.

Panel highlights
“Computing Systems: Today and Tomorrow” featured DPI Director of R&D Klara Nahrstedt as moderator, with panelists from both industry and academia:

• Manoo Assa, AIOps director at CCC Intelligent Solutions

• Derek Ferguson, chief software officer for Fitch Group

• Siebel Professor Brighten Godfrey, also a cofounder of Veriflow (acquired by VMware in 2019)

• Siebel Assistant Professor Minjia Zhang, who worked as a principal researcher at Microsoft before joining Illinois

The most pressing questions from the student audience focused on the future of work in the age of AI. All four panelists agreed that in this job market, the fundamentals are still important.

“There’s a tremendous value in understanding the mathematics and what goes on inside the models,” Ferguson said. In good news for recent graduates, he added that younger employees adapt much better to AI technology, to the point that Fitch Group tries to embed recent university grads in each of their teams. 

Expanding on that idea, Assa said his team is eager to see how recent graduates can bring in their research and academic expertise.

Another student asked if LLMs really improve productivity in any current workflows.  Ferguson said the productivity gains can be “super dramatic” with LLMs, but organizations will need to change their workflow to fully take advantage of them, with close collaboration between the business and the tech sides. 

In response to a question about what changes faculty plan to make in response to the growth of AI including in pedagogy and subject matter, Godfrey said the situation is still evolving. This semester, he tried something different: interviewing every student about their solution to a problem, which not only encourages students to think through problems more fully, but also prevents them from just using Claude. 

“We will be using AI as an assistant, which may accelerate the learning in a lot of places,” he said. He also plans to give students more opportunities with what might be considered a more client-facing role: working with teams to determine “what do you need this thing to do?” before building a system.

Zhang developed a machine learning systems class, reemphasizing that while technologies are moving fast, “there are principles behind all the technology.”

“It just puts more pressure and challenge on the instructor to determine what principles will endure,” he said.

Nahrstedt asked what questions academia can help industry solve. Assa cited AI governance, especially important with the growth of agentic AI. Ensuring data quality, model training, bias detection, and audit trails are only some of the areas where more research would help.

Ferguson suggested building a tooling framework so “mere mortals” can understand what’s going on within the LLM.

Zhang cited the three Es: efficiency, effectiveness, and ease of use. 

Godfrey cited the need for more applications that could be tailored to specific industries.

Since interesting research requires interesting data, Nahrstedt asked the panel how companies can give academics access to their data sets, testbeds, computer clusters, and other resources.

Assa and Ferguson said their organizations try to embrace what academia can bring. 

Zhang said internships can provide individual students with an opportunity to work with interesting datasets. However, he agreed that there should be more cross-collaboration, as companies need expertise and academia needs partnerships. “Could be a win-win,” he said.

Assa also emphasized that while AI-generated research has shown potential, it still requires lots of humans in the loop. “It could get to a good amount of accuracy but not replace the experts,” he said.
 

Other highlights from the day

• University of Chicago Professor Junchen Jiang, co-founder of Tensormesh and faculty lead of LMCache, gave the event’s keynote address on AI-native data systems. e He discussed the LLM inference system and KV cache in both industry and research. Things are changing so quickly, he said, that he has “learned more in the last two years than in the previous 10.”

• Siebel School Assistant Professors Aishwarya Ganesan and Nishil Talati each gave faculty talks on systems for resource-disaggregated data centers and efficient agentic AI, respectively.

• Over lunch, the group celebrated student achievements in the past year.

Lightning Talks

At the end of the day but before the fun of operating system–related trivia, grad students and postdocs gave lightning talks on their research, spanning everything from network design to TikTok:

• Siddarth Agarwal: Just-in-time DMA

• Shreesha Bhat: AgileLog: A forkable shared log for agents on data streams 

• Soham Chakraborty: Drone self-localization and landing

• Jackson Clark and Yiming Su: SREGym, a high-fidelity training ground for Site Reliability Engineering

• Xiangyu Han: Near-optimal network design for ML clusters

• Kiran Hombal: Replicated storage

• Anna Karanika: RASC: Enhancing observability and programmability with smart device action

• Maleeha Masood: Quantifying TikTok using vision language models

• Chenghao Mo: GPU-based vector data management for filtered and multi-vector search

• Bogdan Stoica: AI-assisted systems operation

• Talha Waheed: Multi-party load balancing in the cloud

• Lingzhi Zhao: AquaScope: Reliable underwater image transmission on mobile devices  

Overall, the event was successful at connecting the bright minds of the Siebel School with the fast-paced, ever-changing Chicago tech ecosystem. “The retreat was an excellent opportunity to bring students together, not just to hear from industry at this rapidly evolving time, but to celebrate their own work,” Nahrstedt said. “We hope to do more industry panels in the future.” AI will most certainly play a role.