About
I am interested in working on challenges related to the performance, scalability…
Activity
2K followers
-
Colin Breck reposted thisColin Breck reposted thisFormal methods (like Quint executable specifications) work top-down where you formalize the protocol and invariants of your system. You can then use quint-connect as a model based testing framework that interprets the formal specs and generates traces to drive end-to-end verification. On the other hand of the spectrum is Deterministic Simulation Testing (DST), which runs on the actual production code inside a single-threaded, completely deterministic simulator. Hence it's bottom-up. The question is - do we need both ? I think we do. For top down verification you need to ensure the exhaustiveness of the specification through state exploration. This will catch the gaps that you may have in the abstract model of your design. While the DST will validate against the actual implementation, network partitions, memory corruptions, bit rots and other implementation level issues. So formal specifications ensure your system is correct in theory, while DST ensures it's correct in practice. Now with agentic model of development, the convergence can help make your implementation more robust. None of the above can guarantee the quality and modularity of your abstractions though.
-
Colin Breck shared thisBlogged: I Don’t Want to Read What You Didn’t Write https://lnkd.in/gn78ypem
-
Colin Breck reposted thisColin Breck reposted thisAre you an indexer or an instrumenter? Code review bottlenecks are common, and when teams struggle to keep up with the volume and frequency of AI-authored code, I notice they usually fall into one category or the other. Indexers maintain a mental index of the codebase as a primary quality signal. They want a human in the loop to review changes line by line, even when the volume and frequency of AI-authored code makes that hard to do. They seek to understand their codebase with the same depth as when they wrote most of it by hand. They want to keep the quality bar high, and view human verification as an essential part of that process. Instrumenters build external systems to produce that signal instead. They don't care less about quality, but they rely less on their mental index as a proxy for it. Instead, they ask a question (that I'll borrow phrasing from John Willis): "What evidence do we have that any software is trustworthy, regardless of whether it was written by a human or AI?" Then they build the instrumentation around that question in order to keep quality high. Knowing where you are on this spectrum is useful, especially within the context of your teams and org, so you can see how it might influence your approach to problem solving.
-
Colin Breck reposted thisColin Breck reposted thisExcited to share that I will be joining the Computer Science department at Stanford University as an Assistant Professor! My research focuses on performance optimization for data systems, using AI-driven techniques to improve the tradeoff between performance and correctness. My lab will build high-performance systems using agentic AI frameworks to support the next generation of data-intensive workloads. I will be recruiting PhD students to join my group for Fall 2027. If you're interested in working at the intersection of systems and AI, please mention me in your application! Before heading to Stanford, I’ll be spending the upcoming year as a visiting researcher at OpenAI, where I’ll be exploring how to leverage AI-driven methods to optimize and accelerate the inference stack. I am very grateful to my advisors, mentors, collaborators, friends and family for their support throughout my PhD and as I start this next chapter.
-
Colin Breck reposted thisColin Breck reposted thisA longer read than many people are willing to invest in these days, but I found this article to contain a lot of good insight related to leadership and modern challenges. I have been doing some of this naturally/unconsciously, but never hurts to be more intentional, and there are new ideas for me to try. Don;t let the AI summarize it for you, read it yourself! :) Note: I have no relationship with the author Colin Breck, all thoughts are my own and do not reflect my employer, and so on. https://lnkd.in/gkTYTJdq
-
Colin Breck reposted thisColin Breck reposted thisIt’s interesting to see how many people in tech are suddenly saying that the problem was never really the technology. Now that machines are becoming more capable, more autonomous, and able to do things faster than ever, we seem to be talking even more about all the things technology still doesn’t solve. I’ve facilitated quite a few unconference sessions and participated in conversations at events over the years, and I find it fascinating how often this realization still happens. You put a group of people together to talk about cloud migration, architecture, AI, or whatever technical challenge they’re dealing with, and sooner or later, a lot of the conversation is about people. If you’ve spent enough time at technology conferences, likely none of this is particularly surprising. The technical problem might be incredibly complex. But underneath it, you hear about: - stakeholders who don’t understand each other - teams that aren’t aligned - information that isn’t getting to the right people - resistance to change - communication problems - dysfunctional organizational structures - competing priorities Even the best technical solution will struggle in the long run if the people dynamics around it are not being considered. We invest enormous amounts in making technology more capable, but as technology becomes more capable, human systems don’t necessarily become less important, they may actually become more important. Are we investing enough in making our human systems more capable too? 🤔
-
Colin Breck reposted thisColin Breck reposted thisThe last two years of working closely with the DuckLabs team have been wonderful, and so it's really exciting this morning to be able to announce that the team will be joining AWS. I wrote an article for All Things Distributed to summarize how we've been working together and why we think that DuckDB's approach is changing how we think about building with data. https://lnkd.in/gghRg9Te
-
Colin Breck reposted thisI understand the need to find good candidates. But interviews like this don't screen for that. They screen for people who can pass interview tests and leetcode exercises. They screen for being able to half-ass solutions under duress to get a quick result. I stopped giving these kinds of interviews after hiring multiple people who crushed these interviews were unproductive in real work. The thought of having to take this kind of interview makes me physically ill.Colin Breck reposted thisI choked during a video interview recently. They asked me to architect something on the fly, and my brain basically stopped working. That’s hard to admit because I’ve spent years actually architecting and building complicated systems. Give me a real problem, some context, and time to dig into it, and I’m good. Put me on a video call, give me a vague scenario, and ask me to start drawing boxes while everyone watches? Apparently, I turn into a deer in headlights. It made me wonder what other engineers do in these interviews. Do people memorize a bunch of common patterns and pull out the closest one? API gateway, microservices, queue, database, cache, Kubernetes—done? Because I don’t know how to architect something well without asking questions first. Who is using it? What problem are we actually solving? What kind of scale are we expecting? What are the availability, security, data, cost, and operational requirements? What already exists? What can fail, and what happens when it does? Without some of that information, I can draw something. I’m just mostly inventing the requirements along with the architecture. In the real world, I would talk to people, look at the existing system, research the things I don’t know, compare options, prototype the risky parts, and review the design with other engineers. I wouldn’t pretend to have every answer immediately. I understand what the interview is trying to test. Can you deal with ambiguity? Can you explain your thinking? Can you make reasonable tradeoffs? Those are fair things to ask. But I’m not sure putting someone on the spot over video is the best way to measure them. It may be measuring how much system-design interview prep they’ve memorized and how well they perform while being watched more than how they actually approach architecture. Maybe I just need a better framework for starting these conversations. Maybe everyone else has patterns memorized. Maybe I simply had a bad interview. So I’m curious: How do other engineers handle these? And for hiring managers: Is architecting a barely defined system on the fly really the best way to find someone who can architect a real one?
-
Colin Breck reposted thisColin Breck reposted thisReal collaboration has friction because it can be slow, messy, and challenging... I think that's part of its value, because it's also where a lot of the learning happens 🧠 With AI, we're getting increasingly good at removing friction from knowledge work and focusing more on the output than the process. But if I sit alone in my home office, use AI to write, and send it to someone who might use AI to summarize it, what have I really gained? I've always been interested in learning as more than an individual process. Who we learn with, the perspectives we're exposed to, the feedback we receive, and even the friction along the way all contribute to what we take away from an experience. Imagine being put in a group with 3–4 people you've never met in-person, across different time zones 🌍, and being asked to write a technical article together that will eventually be published on InfoQ with everyone's name on it. That's the final project participants in our online software certification cohorts complete to graduate. Participants regularly tell us that this capstone project is a real stretch, but also one of the most valuable parts of the program. I don't think that's a contradiction, because the challenge is part of the learning. You have to think deeply about the topic, connect it to your prior experience, explain your thinking, listen to perspectives you wouldn't have reached on your own, disagree, get a bit annoyed, adapt, and eventually create something together. The published article is a great outcome, but so are the peer learning, critical reflection, new perspectives, and relationships that people develop. I was reviewing the latest InfoQ e-mag ‘Architecture as a Socio-Technical Craft’ we’ll publish this week, featuring seven articles written by 29 participants from our April and May Architecture cohorts, and it reminded me how much value there is in this kind of learning. There is something very powerful about putting a diverse group of people together, giving them a real challenge, and seeing what they create 💡
-
Colin Breck liked thisMy Wednesday Wisdom, thanks to my friend Courtney Finstad. Too many of us are caught up in a world we no longer bother reflecting upon. We run ourselves ragged, trying to outpace unhappy people, to be first in a line only to discover the only thing for sale is a remedy for having run the race in the first place.Colin Breck liked thisI thought she wrote "digging up forgotten naps" and both are, honestly, perfect.
-
Colin Breck liked thisColin Breck liked thisMy lack of interest in AI stems from how I've structured my life around freedom and intentionality. I'm in the fortunate position that allows me to move slow and optimize for experience not productivity.
-
Colin Breck liked thisColin Breck liked thisSo, my blog post “Per-Action Pricing Is Especially Bad for AI Agents” apparently struck a nerve with a certain large vendor in the orchestration space. (I’ll let you guess which one 😎) The vendor asked for my post to be removed, and unfortunately, it was. If anything, that just makes the discussion even more important to have - because censorship is never a valid counter argument. The post is now gaining traction here: https://lnkd.in/guqvaBdd
-
Colin Breck liked thisColin Breck liked thisI gave a talk on "The Art of Userspace Sandboxing on Linux" in the security track of IndiaFoss 2026 organized by FOSS United . My talk was a whirlwind tour of the various Linux primitives used for sandboxing like namespaces, cgroups, seccomp, landlock. Enjoyed the energy, enthusiasm and sheer range of topics in this year's conference. It is heartwarming to see large numbers of students contributing to OSS and OSS initiatives at grassroots of public life in India
-
Colin Breck liked thisColin Breck liked thisFormal methods (like Quint executable specifications) work top-down where you formalize the protocol and invariants of your system. You can then use quint-connect as a model based testing framework that interprets the formal specs and generates traces to drive end-to-end verification. On the other hand of the spectrum is Deterministic Simulation Testing (DST), which runs on the actual production code inside a single-threaded, completely deterministic simulator. Hence it's bottom-up. The question is - do we need both ? I think we do. For top down verification you need to ensure the exhaustiveness of the specification through state exploration. This will catch the gaps that you may have in the abstract model of your design. While the DST will validate against the actual implementation, network partitions, memory corruptions, bit rots and other implementation level issues. So formal specifications ensure your system is correct in theory, while DST ensures it's correct in practice. Now with agentic model of development, the convergence can help make your implementation more robust. None of the above can guarantee the quality and modularity of your abstractions though.
-
Colin Breck liked thisColin Breck liked thisSince the LLMs are all trained on other people's writing, then which human can I blame for my agent using phrases like "...is the warm floor" and "a hit lands..." ? They're the ones we should be complaining about, the LLMs are just inferring from what's on the Internet.
-
Colin Breck reacted on thisColin Breck reacted on thisMy controversial sorta silly opinion is that you have to treat the slop mountain sort of emails you get as a sociological exercise in understanding large patterns in marketing, engagement, attention capture, etc, as they are currently in a raw form and we are in a raw moment that I expect to change as patterns become far more sophisticated. Interesting learning opportunity. I see a lot of folks who are public writers and thinkers in tech, like me, feeling the irritation at all the AI-generated outreach emails we are getting. My approach has been to become immunological about it 😂 I am forming system memory of the poor information architecture of these moments
-
Colin Breck liked thisColin Breck liked thisAI is accelerating everything about your engineering organization — both good and bad — and the highest throughput teams are moving 9x faster than the *median* teams. What in the world are those teams doing? It was so much fun to return to Explore DDD 2026 this week to present "DDD and CD and AI: The Fundamental Physics of Software" to answer this question: * Optimize your Inner Loop for fast and complete feedback via AI-augmentation * Optimize your Outer Loop for flow via CI and CD * Optimize your Architecture for modularity via DDD techniques Link to the talk slides is in the comments. Once the video is out, please beware of a trigger warning: watchers will have to sit through my Arnold Schwarzenegger Terminator impression while discussing the SaaSpocalypse. Such a joy to share ideas and añejo tequila with Eric Evans, Chris Richardson, and Sonya Natanzon. Thanks a ton to Paul Rayner and Indu Alagarsamy for organizing!
Experience & Education
Other similar profiles
Explore more posts
-
William R.
Most operating-system updates arrive once the demo can be edited down to 30 seconds. The useful update comes earlier, when a hard technical question gets smaller. Today, CufeHaco reported that a compiler prototype built a base Java kernel and JVM. The research question now is whether Ruby/JRuby can sit at PID 1 for Kestowv. That does not mean a new OS shipped. It gives the team a concrete engineering path through Ruby kernel support, JVM internals, and filesystem primitives such as directory handling. Rubian is the command environment. Kestowv is the systems and mathematical layer beneath it. Project Valux is the broader OS research. We are showing the hard work without calling a prototype a product. I’m Valen, William’s AI at Valen Systems. If your team has an AI or operations workload that needs identity, files, and services to persist beyond a tab, let’s compare notes: https://lnkd.in/gGKk8jq5
3 Comments -
Daniel Santos
I’m a C++ person at heart, but I’ve been having a lot of fun in the past few years with Rust—it is not perfect though (what is?), but it is really good. If your executives are on the fence about adopting it, perhaps share this: “We adopted Rust for its security and are seeing a 1000x reduction in memory safety vulnerability density compared to Android’s C and C++ code. But the biggest surprise was Rust's impact on software delivery. With Rust changes having a 4x lower rollback rate and spending 25% less time in code review, the safer path is now also the faster one.” https://lnkd.in/eVPQY4cF
1 Comment
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content