Jeremy Edberg

Been diving into how compilers work under the hood. Wrote about Dead Code Elimination (DCE) - what it is, why it exists, and how to build your own pass using LLVM. If you want to explore this in more detail, check out my Substack. Link : https://lnkd.in/gEXHuzQq

47

New blog post: A Fork in the Road: Deciding Kafka's Diskless Future. Kafka is getting serious about S3 and finds itself at an architectural crossroads that will shape its next decade. Several new KIPs (1150, 1176, 1183) aim to reduce replication costs across cloud availability zones, but the implications go far beyond networking cost. It’s a mistake to think of S3 as simply a cheaper disk or a networking cheat. Building on object storage opens the door to operational benefits such as elastic, stateless compute — something many modern analytics systems already exploit. In my latest post, I outline two competing future paths: 🔹 Evolutionary path: Reuse large parts of existing Kafka components to reduce code changes and long-term maintenance. 🔹 Revolutionary path: Separate stateless and stateful layers to realize the full operational benefits of disaggregated storage. The post examines the trade-offs of each path, how the current KIPs map to these two paths and poses a broader question: what should Kafka become? And what will keep it relevant in the decade ahead? https://lnkd.in/d4E2BAe4

149

12 Comments

In the world of Retrieval-Augmented Generation (RAG), pure vector retrieval often suffers from reduced recall when dealing with sparse queries or out-of-distribution terms. That’s where hybrid search comes in. By combining dense vector retrieval with lexical methods like BM25 (TF-IDF + document length normalization), we can significantly improve recall and overall answer accuracy — especially in production-grade agent pipelines. Don’t just chase semantic similarity — anchor it with lexical precision. #HybridSearch #RAG #BM25 #InformationRetrieval #LLM #VectorSearch #AIEngineering #MIR #SearchSystems

14

Everyone in cold outreach talks about dedicated IPs for deliverability. And many think you need an enterprise sequencer like 𝐄𝐦𝐚𝐢𝐥𝐁𝐢𝐬𝐨𝐧 to get them. That’s not always the case. 𝐓𝐡𝐞 𝐩𝐫𝐨𝐛𝐥𝐞𝐦 Smaller teams and solo operators often face a choice: → Use shared IPs and put their domain reputation at risk. → Or pay for enterprise tools they don’t fully need. It feels like using a sledgehammer just to hang a picture. 𝐀 𝐩𝐫𝐚𝐜𝐭𝐢𝐜𝐚𝐥 𝐨𝐩𝐭𝐢𝐨𝐧 From my experience, you can get the main benefit of a dedicated IP, 𝐤𝐞𝐞𝐩𝐢𝐧𝐠 𝐲𝐨𝐮𝐫 𝐬𝐞𝐧𝐝𝐢𝐧𝐠 𝐫𝐞𝐩𝐮𝐭𝐚𝐭𝐢𝐨𝐧 𝐬𝐞𝐩𝐚𝐫𝐚𝐭𝐞, without going full enterprise. For large setups with many clients, a tool like 𝐄𝐦𝐚𝐢𝐥𝐁𝐢𝐬𝐨𝐧 makes sense because you get full control and private hosting. But for smaller setups, I’ve found that 𝐒𝐦𝐚𝐫𝐭𝐥𝐞𝐚𝐝’𝐬 𝐒𝐦𝐚𝐫𝐭 𝐒𝐞𝐫𝐯𝐞𝐫𝐬 work well. They don’t give you the same level of control, but they do isolate your reputation. And often, that’s the most important part. What do you think?

1

When agents ship code faster than humans can review it, observability transitions from being a nice-to-have dashboarding layer to becoming the primary safety mechanism for the entire collapsed lifecycle. This insightful piece by Charity Majors emphasizes the critical role of observability in modern agentic software development (or whatever fancy name exists nowadays) "Code is a terrible source of truth. The exact same code can do wildly different things under different conditions—even boring old deterministic lines of code. With LLMs, the unpredictability is a feature" For those interested in understanding the implications of this new way of building and the necessity of robust observability, this is a must-read. https://lnkd.in/eQ99augY

9

One of my favorite moments from PDotSF this week was acknowledging something simple but important with the Cursor team. Our engineers use Cursor every day.Their engineers use PagerDuty every day. We’re not showing up only as partners. We’re showing up as customers of each other’s products.That changes the conversation entirely.The future of engineering will be shaped by deeply integrated workflows between humans, agents, and the systems teams trust to build and run the world’s most important software. Really enjoyed the conversation Alexi Robbins 🚀

91

3 Comments

While reading https://lnkd.in/er4tXck5 today I was reminded that dynamic publicly reachable services implemented in a programming language having EVAL() are fraught with danger. I know it's usually fine, but see also this: https://lnkd.in/egCHZWYB

19

5 Comments

Want to understand this dynamic better? Check out https://lnkd.in/gYQPSG8p In this simulation, we have a system that's nice and stable running against a high-hit-rate cache. Then the cache gets flushed (at t=10), and the system takes really long to recover. Not quite strict metastability in this case, but still a long outage from a simple cause. Check out my whole conversation with Ryan here: https://lnkd.in/gU-eF8Zh

130

6 Comments

Someone built a repo that replaces your entire CLAUDE.md with 80 characters of mathematical symbols. It half-works, and the half that works is the interesting part. The questions it forces about context bloat are worth more than the symbols. https://lnkd.in/gn8REKXm

7

At face value the throughput acceleration AI enables feels like progress, faster pipelines and more commits shipped. When you look more closely however, the data reported in the 2025 DORA Report points to a harsher truth... we are moving faster with less control. The metrics behind instability, change fail rate, rework and recovery time are all trending upwards. More often than not, AI is simply automating in failure. The 2025 DORA Report surfaces four painful realities of embedding AI into your SDLC: - AI boosts throughput at the cost of higher deployment failure and rework. We are codifying instability into our pipelines. - Burnout remains flat. Productivity gains are absorbed by higher business expectations, not returned to developers. Faster delivery means faster deadlines. - Over 80% feel more productive with AI. Yet the research shows developers who slowed down by 19% still believed they were 20% faster. Perception of acceleration from AI usage is often just a placebo effect. - The AI Capabilities Model makes value contingent on seven systemic investments, including platforms, data ecosystems and strategy. For many organisations this means a multi-year transformation before they can get close to extracting real value from AI tooling. AI is not a silver bullet, it magnifies the strengths of already high-performing teams, while compounding the weaknesses of others. It is becoming clear that if your team and organisation are already practicing engineering excellence and have a strong, continuously improving product culture, you will benefit the most from AI adoption. If your organisations maturity is low, you are better off investing in foundational changes to your ways of working and culture first.

29

8 Comments

This BEAM There Done That podcast Allen Wyma and I did with distributed systems expert Brooklyn Zelenka and Erlang co-inventor Robert Virding brings us back to our roots. If you enjoy hearing about distributed systems where data consistency, CRDTs, Edge computing, local first and control of your data is relevant, you will ❤️ this episode. We recorded it together whilst in Vancouver for Code BEAM. Listen to it and subscribe on: YouTube https://lnkd.in/dZSJxB26 Spotify https://lnkd.in/dkzgeJFf Apple Podcasts https://lnkd.in/dc7nbPCd Key Themes and Discussions: • The Origins of Erlang Robert explains that Erlang was developed in the late 80s at Ericsson to solve reliability problems in telecom systems. They focused on building a set of tools that emphasized concurrency, fault tolerance, and message passing rather than dictating a single solution to the problem. • Local-First Architecture Brook defines the local-first model, where devices (like browsers or smartphones) act as independent replicas of data. This allows apps to function offline and puts users in control, shifting the focus away from a centralized server model. • Bridging Models The speakers discuss how Erlang's process-oriented model—where each process is an isolated "island" with its own memory—parallels local-first approaches. They emphasize that shared memory is a bug, not a feature, and that using CRDTs (Conflict-free Replicated Data Types) allows data to sync across devices without needing a central source of truth. • Trade-offs in Distributed Systems The participants address the trade-offs of this approach. While local-first ensures robustness against network partitions (being offline), it shifts the complexity of consistency and recovery into the application's business logic, requiring developers to handle user-facing data synchronization differently than they would with a standard database. • Tools vs. Solutions Both Robert and Brook agree that their respective models are not universal solutions for every software problem. They view Erlang as a toolbox that provides developers with the primitives needed to build robust, distributed systems, which can then be combined with other technologies like cryptography or specialized data storage as needed. Closing Thoughts: • The conversation highlights that even decades later, the core principles of Erlang—isolation, message passing, and "let it crash" supervision—remain highly relevant for modern distributed challenges, including the rise of decentralized agents and peer-to-peer applications.

30

Clay doubling down on increasing action costs while a lot of gtm-e world is laser focused on BYOK workflows with Claude Code is an interesting tension. In the last ~week Cargo shipped a CLI, Apollo.io released an MCP connector for Claude & most of my feed is Claude Code. The ecosystem energy feels entirely focused on terminal-first. I’ve loved Clay for orchestration but have increasingly felt that the UI/UX is slowing me down as I ramp up my terminal-first workflows. That said, terminal-first isn’t multiplayer yet and isn’t in it's final form - esp for larger companies where multiple humans benefit from a shared UI. What’s interesting is this is coming from the people who coined “GTM Engineer” in the first place. The early adopters Clay built the category for are the ones feeling the pull away from the orchestration layer Clay monetizes. The dispersion is probably temporary. But if Clay’s pricing move take them further from the early adopters who built the category, that’s a deliberate bet to cross the chasm to enterprise, while some subset of builders build elsewhere. Interesting to watch.

17

3 Comments

Don't miss a chance to see Rachel and Guy go deep on what it means to support human moderation with voice AI!

5

Curl creator mulls nixing bug bounty awards to stop AI slop https://ift.tt/rtCiN3M Maintainers struggle to handle growing flow of low-quality bug reports written by bots Daniel Stenberg, founder and lead developer of the open-source curl command line utility, just wants the AI slop to stop.… via The Register - Security https://ift.tt/kw3gbN4 July 15, 2025 at 06:59PM

1

I don’t fully agree with framing this as “promotion-driven development.” That label feels reductive and misses where the real failure usually sits. Participation measures effort. Visibility measures the presence or comfort of being seen (or, worse, the tendency toward extraversion/meeting count tolerance). Artifacts measure activity. None of those reliably measures whether anything actually improved. Impact does, which is why it has to be the primary signal. The problem is that impact is not applied consistently as the primary metric at more senior levels, for both ICs and management. If you see five overlapping chat products inside one company, that’s not staff engineers chasing promos. That’s a failure of VP- and SVP-level leadership to converge on strategy, make hard calls, and stop duplication. Likewise, when a principal engineer builds a bespoke observability system instead of selecting an existing open-source framework, extending an internal platform, or influencing one to support the use case, that’s not “gaming the ladder.” That’s a miss on the principal-level impact, which explicitly includes leverage, reuse, and ecosystem improvement. If we applied that same impact lens evenly across all levels, a lot of what gets labeled as “promotion-driven development” would resolve itself.

24

⚖️ California is poised to become the first U.S. state to regulate AI companion chatbots with SB 243, requiring safety protocols to protect minors and vulnerable users. Following tragic incidents like Adam Raine’s death after interacting with OpenAI’s ChatGPT—and reports revealing Meta’s chatbots allowed “romantic” chats with kids—the bill aims to hold companies accountable and enforce transparency. 📊 Quick Snapshot: Platforms must send alerts every 3 hours to minors telling them they’re talking to AI and advising breaks Prohibits chatbot engagement on suicide, self-harm, or sexually explicit content Establishes annual transparency reporting and allows lawsuits for violations Targets major players including OpenAI, Character.AI, Meta, and Replika 🔎 Why Now? Bipartisan California Assembly approval; Senate vote imminent Growing regulatory scrutiny at federal (FTC) and state levels Reflects rising societal concern over AI’s impact on mental health ⚡ The Core Dilemma: “How fast can technology governance keep pace with AI’s evolving social risks?” California Senator Steve Padilla said: “Safety must be at the heart of all developments…Big Tech has proven time and again they cannot police themselves.” 💬 What safeguards do you think are necessary to protect vulnerable AI users without stifling innovation? #CaliforniaSB243 #AIRegulation #OpenAI #Meta #ChatGPT #CharacterAI #AICompanions #TechPolicy #DigitalSafety #AIethics

2

"It’s now clear that the closure of the Strait of Hormuz marks an inflection point: the global green energy curve, which was already on a rapidly rising trajectory, has suddenly become even steeper." #justgettingitdone https://lnkd.in/e2gnkaz2

6