Stephen Blum

I spoke with Tom Wilkie, CTO of Grafana Labs about what it’s like scaling open source projects while building a commercial company around them. Discussed the challenges, lessons, and what goes on behind the scenes of a tool so many of us use: https://lnkd.in/gD92CXuD

587

8 Comments

💡 AI doesn’t just change how code is written – it changes expectations and outcomes. In the final section of Jellyfish’s AI Adoption Guide, we explore how expectations for engineering leaders are evolving as AI becomes embedded across the SDLC. Because the organizations pulling ahead are treating AI as a transformation, not a tooling upgrade. Inside, you’ll explore: - Why AI adoption is quickly becoming a leadership mandate, not an option - How to balance ambitious goals with realistic expectations - What cultural investments are required to sustain long-term gains - How data can be used to align executives and engineering teams The teams that win in the AI era won’t just have better tools – they’ll have leaders who know how to guide change. Download today: https://lnkd.in/dut_6WpQ

19

So which models produce HTML that has the least accessibility errors, and at what cost? Mapped it out here (bottom left is the best). This is helpful for understanding which models are most compliant for those with disabilities and do it efficiently. Qwen and Gemini 2.5 Flash lead the pack. Hat tip to Ben Ogilvie for pointing me to this: aimac.ai

55

2 Comments

Stop wrestling outputs. Start shaping them. Today Anthropic dropped a small switch with big leverage: Claude Code + Claude app now support /output-style, so you can lock the way Claude talks before it writes a single line. Consistent tone on demand. Reviewer voice for PRs, mentor voice for juniors, exec voice for summaries. Less cleanup, more shipping. Stop rewriting verbose answers into the format you actually need. Team-readable by default. Standards aren’t a hope, they’re a preset. Quick start (paste this at the top of your chat): /output-style: "Staff engineer — terse, diff-first, no fluff. Use bullet points, show code before commentary, 120 chars width." Now try: “Refactor this file for clarity and testability.” “Explain the change like I’m onboarding a new hire.” “Summarize risks for a product manager.” My go-to presets: ✅ Reviewer: “Blunt, precise, starts with a diff, cites lint/rules.” ✅ Teacher: “Step-by-step, small chunks, labels pitfalls & why.” ✅ Security: “Threat model first, list vulns -> CWE refs -> fixes.” ✅ Prod Notes: “3 bullets: Context → Decision → Next Step. 60 words max.” ✅ No-Fluff: “Answer in the fewest tokens that still solve the task.” You don’t have to beg models for style anymore, you set it, then build. Comment STYLE and I’ll share my ready-to-paste presets for Claude Code.

45

41 Comments

I'm thinking a lot this week about GEMINI.md vs AGENTS.md while we make it pretty easy in Gemini CLI to use either there are subtle differences in the defined file formats. For example GEMINI.md supports @ style file references and AGENTS.md does not. In addition I think it would be great if AGENTS.md supported a call out mechanism to have custom instructions for individual models or systems. I propose markdown callouts for this: > [!gemini] > Gemini only instructions here > [!codex] > Codex only instructions here I think increasingly it's getting difficult for people to manage multiple agent files and I'd love to hear your opinions on what we should do here. There is a discussion in our repo here: https://lnkd.in/g4GywT4P

23

3 Comments

This is an interesting networking vendor - IO River has developed an overlay for CDNs that makes use of WASM (gotta love WebAssembly). It just announced a nice $20M fund raise today too. “We are leveraging the existing physical networks, like Akamai, CloudFlare, Fastly and AWS, but we are also leveraging their edge compute,” Edward Tsinovoi told me. “On top of that, we provide a virtual layer of edge compute. So customers write their code once and it will be distributed and run on these different platforms.”

8

A few seconds of silence can break the rhythm of an entire conversation. And with agents, timing defines reality — short delays feel human, long ones feel robotic. This post breaks down how we’ve worked to engineer latency out of Sierra’s voice stack: from detecting the end of a customer’s speech to streaming a response that feels instantaneous. https://lnkd.in/e4jreAyn

239

10 Comments

Despite the many benefits of Data #CleanRooms they may turn participating parties into Data Brokers as per California’s DELETE Act if they happen to bring 3rd party data (no direct relationship) into the platform, triggering a registration and tech alignment requirement (and hefty fines if dates are missed). (Presentation by California Privacy Protection Agency and Ben Isaacson on the DELETE ACT and the new #DROP platform that integrates all data brokers for consumers to place a single request.) Privacy Law Section of the California Lawyers Association. #AdTech

54

4 Comments

When I ask an engineer about a piece of technology they own or work with, what I want is their honest assessment of it. Sometimes, what I get instead is a glowing PR piece about how everything is going great and nothing is wrong with it at all. This has two problems: 1. It doesn’t help me make the technology better, let me understand what the current challenges are, etc. 2. When I go look into the technology myself and discover that parts of it are bad, it makes me mistrust the judgment of the engineer who gave me a glowing PR piece. It’s fine to say “there are things that need to be improved, here they are, but they are not our priority right now and here’s why.” It’s fine to say, “there were some tough circumstances that forced us to make bad decisions.” And believe it or not it’s even fine to say, “I messed up this part because I didn’t know what I was doing.” All of those make me trust your judgement and help me understand what might need improvement. I’m not saying I want only bad news. I _do_ also want, “our users really love this part, here’s the evidence.” I do want, “we are very proud of this well-designed part of the system.” And so forth. I just want an accurate picture of the state of things. I’m not judging people on whether they say they did good work or not. But I _am_ judging them based on the relative accuracy of the things they say. I understand that sometimes there can be a culture of fear where people are worried about being hit because they said something is bad. So you have to counter this by celebrating _truthfulness_, and celebrating failure when it’s done well (for example, caught early, let us learn something important, etc). And when people do tell you something is bad, you have to let them know you heard them and that it’s okay to say that. Otherwise, leaders end up in a difficult situation where everybody will only tell them that everything is great, all the time. What this really causes is that it makes leaders distrust everybody’s judgment and forces them to make decisions about things they are too far away from to make good decisions about, because nobody is just giving them the straight data, presenting their own trustworthy decisions, and then acting on that.

43

6 Comments

🎧 Next up on "Graph Power Hour!" GraphRAG at scale with Amazon Neptune and Amazon Bedrock Tue July 15 09:00-10:00 US Pacific https://lnkd.in/gukKRP2H I'm thrilled to be talking with Ozan Eken and Nicole Moldovan from Amazon Neptune to discuss in depth about new GraphRAG capabilities. Given the recent GA release, cloud-based applications can now build graph vector stores directly from Amazon Bedrock for use with Neptune GraphRAG. We'll discuss Neptune Analytics, a memory-optimized graph database engine for analytics -- which can ingest large volumes of in-memory graph data and implements popular graph algorithms plus support for low-latency analytics queries, such as what's needed in FinCrime use cases. We'll also talk about a new open source Python library from Neptune which includes automated construction of a lexical graph from unstructured data sources, along with BYOKG-RAG -- a novel approach for Question & Answer use cases combining structured knowledge graphs with LLMs. Overall, these integrations provides a fully managed solution: create GraphRAG applications from unstructured data without being experts in graph databases. prev episodes: https://lnkd.in/gmsRx2HW #ERKG #Senzing

57

4 Comments

While everyone was talking about Opus 4.6, for me the true killer feature of the recent Claude Code updates is the agent teams. It's been something i've been trying to achieve with customization for a while. Custom agents, orchestration scripts, specific CLAUDE.md instructions to coordinate work... with some degree of success. But what Anthropic shipped natively is a WHOLE different level. What makes this stand out is the inter-agent communication. We're not talking about simple fan-out/fan-in where you spawn workers and collect results. These agents talk to each other. Peer-to-peer messaging, dependency-aware task graphs that auto-unblock, agents that self-claim work from a shared task list. The lead can even enter Delegate Mode where it does ZERO implementation, only coordination. The image below is from one of my setups. A team manager orchestrating a librarian agent, a PhD lead, and 5 research sub-tasks with blocking dependencies. The librarian unblocks the research tasks, the PhD lead aggregates everything. All coordinated autonomously. And with this a whole new world of orchestration just unveiled itself. Distributing work across agents is the "easy" part. You break down tasks, assign owners, define dependencies. The HARD part, and what MOST stands out now, is aggregation. How do you take the output of 5 parallel agents, each with their own context window, and synthesize it into something coherent? That's the new skill. Anthropic themselves used 16 parallel agents to build a 100,000 line Rust C compiler that compiles the Linux 6.9 kernel. No human actively coding. ~$20,000 in API costs over ~2,000 sessions. We went from pair-programming with AI to managing AI engineering teams. The skills that transfer are the ones from engineering management: task decomposition, context management, knowing when to intervene vs let the team self-organize. This is a new paradigm, and i think it opens up several possibilities we haven't fully explored yet. ref.: https://lnkd.in/dVCe344z

66

12 Comments