Metadata

This recent paper from the Berkeley Sky Computing Lab has been making waves in systems community. Of course, Aleksey and I did our live blind read of it, which you can watch below. My annotated copy of the paper is also available here. This is a fascinating and timely paper. It raises deep questions about how LLMs will shape the research process, and how that could look like. Below, I start with a short technical review, then move to the broader discussion topics. Technical review The paper introduces AI-Driven Research for Systems (ADRS) framework. By leveraging the OpenEvolve framework ,  ADRS integrates LLMs directly into the systems research workflow to automate much of the solution-tweaking and evaluation process. As shown in Figure 3, ADRS operates as a closed feedback loop in which the LLM ensemble iteratively proposes, tests, and refines solutions to a given systems problem. This automation targets the two most labor-intensive stages of the research cycle, solution tweaking...

This week, Aleksey and I met not to dissect a research paper, but to chat about "the process of PhD". I had recently wrote a post titled "The Invisible Curriculum of Research" , where I framed research as an iceberg, with the small visible parts (papers, conferences) resting on the hidden 5 Cs: Curiosity/Taste: what problems are worth solving. Clarity: how to ask precise and abstracting questions. Craft: writing, experimentation, presentation. Community: collaboration and contribution. Courage: resilience through setbacks. Above is the video of our chat, with a lot of personal anecdotes and a few rants. But if you want to cut to the chase, the highlight reel is below. What a PhD Really Produces The real product of a PhD is not the thesis, but you, the researcher! The thesis is just the residue of this long internal transformation. Like martial arts, the training breaks you and rebuilds you into someone who sees and thinks differently. This transformation cannot be ...

This paper (to appear at SOSP'25) is one of the latest efforts exploring the dream of a one-round commit for geo-replicated databases. TAPIR tried to fuse concurrency control and consensus into one layer. Tempo and Detock went further using dependency graphs.  Aleksey and I did our usual thing. We recorded our first blind read of the paper. I also annotated a copy while reading, which you can access here . We liked the paper overall. This is a thoughtful piece of engineering, not a conceptual breakthrough. It uses  future timestamps to align replicas  in a slightly new way, and the results are solid. But the presentation needs refinement and stronger formalization. (See our livereading video about how these problems manifested themselves.)  Another study to add to my survey , showing how, with modern clocks, time itself is becoming a coordination primitive. The Big Idea Tiga claims to do strictly serializable one-shot (multi-shot ok with reconnaissance queries t...

Courses, textbooks, and papers provide the formal curriculum of research. But there is also an invisible curriculum. Unwritten rules and skills separate the best researchers from the rest. I did get an early education on this thanks to my advisor, Anish . He kept mentioning "taste", calling some of my observations and algorithms "cute", and encouring me to be more curious and creative and to develop my "taste".  Slowly, I realized that what really shapes a research career isn't written in any textbook or taught in any course. You learn it by osmosis from mentors, and through missteps: working on the wrong problem, asking shallow questions, botching a project, giving up too soon. But if you can absorb these lessons faster, you will find research more fulfilling. The visible curriculum teaches you how to build a car. The invisible curriculum teaches you where to go, who to ride with, and how to keep going when the road turns uphill. After 25 years of exp...