Tejas Kumar

Experienced developers aren’t replaced by AI—they’re supercharged by it. In this thoughtful piece, Manuel Kiessling explores how senior engineers can use AI coding agents to improve productivity, quality, and joy in software development. 🔗 https://lnkd.in/dthdS-iZ #AI #SoftwareDevelopment #DevEx #Productivity #SeniorEngineers

13

1 Kommentar

Software development has fundamentally changed for me. I’ve been using AI as part of my daily development workflow for years now—it’s become a standard tool in my toolbox. But recently, I tried something new: ChatGPT Codex. Codex is essentially ChatGPT connected directly to your repositories. The difference? It doesn’t just read your code. It runs your tests, spins up supporting services like Redis, analyzes compiler output, reasons about test logs—and keeps iterating until it gets things right. This is a completely different experience from what I’d call “vibe coding”—where an LLM has some context about your codebase and maybe the file you’re working on, but no real awareness of the larger system. Vibe coding often leads to hallucinations: fake APIs, patterns that don’t exist, or large, irrelevant rewrites in unrelated parts of the system. It feels like one step forward, five steps back, followed by a hard reset to a clean git commit. Codex, on the other hand, is grounded. It makes small, precise changes based on real feedback from the environment. No wild guesses—just direct, incremental progress on the task at hand. Right now, I’m using it for practically everything: Updating Proto.Actor documentation and diagrams Porting features from C# to Go Simplifying overly complex code Generating examples and test cases Filling in gaps in test coverage Creating GitHub workflows It’s not going to build you the next Unreal Engine. But for all the everyday maintenance—fixing things, porting code, writing tests—it’s a game changer. And the best part? It almost never produces unmergeable code—as long as you have tests for it to validate against.

53

19 Kommentare

A week ago on #TheContext, Toby Padilla from GitHub talked about his view that opening up MCP publication on the GitHub registry will spur adoption. Very good episode.

9

Just finished watching this video by Maximilian Schwarzmüller A very subjective comparison of Claude Code, OpenCode, Cursor, and GitHub Copilot. What stood out to me wasn’t just the features, but the hands-on, real-world experience—how these tools actually fit into daily development. 📌 Key takeaway → There’s no single “best” AI coding tool. → It depends on your workflow, preferences, and how you like to think while coding. If you’re using AI tools seriously in your dev work, this is a must-watch. 📌 → Which one are you using right now, and what do you like (or dislike) about it? 📌 → Check out the video here 👇 https://lnkd.in/g2wiSS5N #AI #AICoding #Developers #SoftwareEngineering #DevTools #Claude #Cursor #GitHubCopilot #OpenSource

5

Did you know you can use mittwald's hosted AI models (currently in beta) for coding assistance in both VS Code and JetBrains IDEs? Just pushed a bit of documentation (read more at https://lnkd.in/gPKSUHVV) that explains how to configure both JetBrains AI Assistant as well as Continue (for VS Code, or JetBrains IDEs) and Cline (also VS Code) to use open-source models like Qwen3-Coder-30B-Instruct via the mittwald API. All models are hosted on german servers, and no user data is stored (let alone used for training purposes).

29

Did journalists get replaced by LLMs - even though they were already great at writing? I don’t think so. And I believe the same applies to software engineering. The craft has never been just about producing text or code. It’s about ‘the what, the why, and the how’ behind what gets built. Tools can accelerate output, but they don’t replace judgment, context, or intent. Boris Cherny, creator of Claude Code, recently predicted the emergence of a new wave of builders using tools like Claude Code - especially people for whom software development previously felt inaccessible. I think he’s right. We’re lowering the barrier to entry, and that’s powerful. But the real difference isn’t simply that more people can build. It’s how well things are built, and how quickly ideas move from concept to reality. In many ways, it reminds me of football (soccer): street football and the Premier League both involve the same game, but the level of execution, structure, and mastery changes everything. AI doesn’t remove the craft - it changes the playing field. Curious to hear how others see this shift. Are we witnessing an evolution?

38

3 Kommentare

The room went quiet at the Lausanne Craftsmanship meet-up today. Not when we showed the framework. Not during the live demo. During the research slide. DORA 2024: Delivery stability down 7.2% with AI adoption. Stanford: Experienced developers in complex codebases saw negative impact. METR: 19% time increase when developers felt 24% faster. Someone in the third row stopped typing. A few phones went down on tables. When we opened Mentimeter, the responses landed with a weight I wasn't expecting. One phrase kept surfacing: "We've been measuring the wrong things." That's not a criticism of anyone's intelligence. The metrics we inherited weren't built for this moment. Lines of code. Velocity points. PR throughput. They measure motion, not progress. What got me was the conversation after. Not "how do we adopt AI faster" but "how do we know if it's actually working?" and "how do we bring junior developers up to speed when the tools do the thinking for them?" Those are the right questions. And honestly, we don't have all the answers. Nobody does yet. But we have seven years of failing, measuring, adjusting. And two things that might help: A free AI-Readiness assessment. And nWave, the agentic AI framework we built for crafters after the obvious approaches stopped working. If you're asking similar questions, happy to share what we've learned. No pitch. Just patterns. To everyone in that room today, especially Andrea Laforgia Peti Koch Maarten Engels Pedro Santos Claudio Perrone... thank you. Your questions and reactions were better than our slides. What metrics is your team actually trusting right now? #nWave #AIReadiness #SoftwareCraftsmanship #ceo #cto

31

10 Kommentare

I'm on record as being quite sceptical about Large Language Models. There are a lot of people in whose interest it is to hype these things up, I'm not one of them. LLMs are certainly useful. Especially for linguistically driven tasks. Computer programming is a kind of linguistically driven task even if the language is rather specialised. So, what is "vibe coding" going to do to programming? As someone whose been writing computer programs professionally since 1990, I have something of a vested interest. I suspect they are useful in different ways to different kinds of programmer. For non/junior programmers they are a way in, they make things possible that would not be possible. For mid/senior programmers they are an assistant. Where I find them very useful and fairly reliable: Architectural spike of a well known but new to me area. Where I find them useful and somewhat reliable: Non-trivial but non-complex coding tasks that I would find boring to implement. Where I find them rather unreliable: Everything else. We've been here before. Off-shoring was supposedly the answer to needing expensive developers. We can just specify the application, hand it over, and get back working code. Well… it works but it wasn't that simple. It brings us back to age old questions that we were wrestling with in the '90s: * Buy vs build * Best point solution vs best integrated solution Should we use this existing SaaS platform or vibe code our own? It's so easy to vibe code! It feels great! However, you get to own that code and all it's problems and long term maintenance. And the data. And the fallout when everyone wants to vibe code their own hacky solution to every problem or challenge. How I think this plays out: Let's assume that most of the problems of vibe coding go away and it becomes reasonably reliable and accessible. What does it do to development? It commodifies the simpler end and creates opportunities to layer higher order services on top (c.f. Wardley Mapping). Those will not, by definition, be simple to execute properly. When I started programming a spreadsheet was a business and a major undertaking for good programmers. These days any old fool can put a spreadsheet in their app with git and a few lines of code to integrate a library. Now it's what you do with the spreadsheet module that matters, what leverage it gives you to do more interesting and valuable things. And so, assuming my more sceptical predictions don't play out, will it be here. You'll vibe code the simple stuff and turn to developers for the more complicated stuff. Unless and until some idiot cracks AGI. Then all bets are off.

8

2 Kommentare

Today's the day! The Distributed AI Hackathon in Berlin kicks off in just a few hours 🚀 We're bringing together students, researchers, and AI builders for Europe's first hackathon focused on federated and distributed AI. Over the next two days, teams will compete to train the best AI models using distributed workspaces across various datasets, proving that data-sovereign AI is achievable. What's at stake: 🏆 €5,000 cash prize pool 💻 Free GPU credits for all participants 🤝 Direct connections to Germany's AI industry and ecosystem 🎯 Real experience building federated AI solutions Flower Labs, as the world's leading opensource federated and decentralized platform is proud to organize this with key partners that include: - exalsius / logsight.ai (SPRIND-backed) – cloud GPU infrastructure for modern AI teams - Technische Universität Berlin's top technical university driving next-gen research - Einstein Center Digital Future – Berlin's inter-university digitalization research hub Doors open at 13:00 today. Let's build the alternative to centralized AI—together!

31

1 Kommentar

I really appreciated this post by Adam Jafer, describing an experience at Voi, which reflects an experience I think we've all had in tech, and I've had it multiple times in my experience. The surprise: Applying 'more' of something that worked in the past, doesn't work now because it's not exactly what's needed for the current set of problems. So in practice it looks like this: 1) Using AI gives you a quantum leap in speed of initial development - giving you a whole lot of functionality that you previously did not have, which lets you get that first customer, or that first investor but! Adding "more AI" doesn't keep on giving you more value, because your own time is limited, prompting gets harder and harder, and the problems require more and more skill to solve. 2) Hiring one programmer gives you a quantum leap in value, but! the second one gives you a little less, and eventually adding more people doesn't give you much more value at all. Because the programmers need managers, and they speak to each other, and they form teams, and they need to coordinate, and they need to discuss the development processes, and they need to re-evaluate what the project is actually about, etc. 3) Adding a server gives you a quantum leap in value, but! adding more and more servers will not help you if the actual resource problem is not actually CPU or IO bound ... 4) One can extend this idea "1 is great, but adding more and more gives less and less improvement" into all kinds of things. I would leave that as an exercise for the reader.. Who is not surprised by this: generally the people who already know this are experienced software developers, because, that's the nature of software development. To solve any problem in software, the very first thing to do is to figure out the exact nature of the problem, and solve that directly. And sadly, human nature, or possibly, the nature of business, is not to properly diagnose and solve problems! But, there are always some people in the team, or company, who do know! The hard part is to listen to their processes, their ideas, when much louder and more confident voices proclaim they already know what to do!

2

650 signups. 150 spots. Full house at the Claude Code Meetup Munich yesterday. I gave a lightning talk called "More Code, More Problems." The premise: AI makes individual developers faster. The evidence is clear on that. But teams might actually get slower. Why? Because the bottleneck moves from writing code to coordinating it. Review, integration, alignment. More code means more load on communication and coordination. More code, more problems. We also had great talks by Sebastian Weyrauch (nion digital) and Dominik Grusemann (Marbles AI). Big thanks to the orga team around Dr. Florian Steiner, M.B.R., Ramin Assadollahi, Jens Franke, Steffi Kieffer and Jens Rusitschka. You are doing a really great job organizing theses events, I really enjoyed them as a guest and now also as a speaker!

32

8 Kommentare

We just shipped something small, but important, for ZurichJS Conf. You can now report issues directly on the website. Typos. Confusing UX. Performance hiccups. Unclear information, for example around CFP deadlines. Anything that feels “off”. We are building this conference in public. That means the experience should not be dictated only by us. It should be shaped by you. The goal is simple: Make the website and the event experience as clear, smooth, and intentional as possible. If something is confusing, that is on us. If something breaks, we want to know. If something can be better, we are listening. There is now a dedicated “Report an issue” entry in the footer. Submit feedback, and we will review it personally. We are making incremental improvements every week. Your feedback directly influences what gets fixed next 💛 PS: We also shipped a few hidden Easter eggs across the site 😉 If you find them, you might unlock a secret ticket discount. No guarantees. This conference is community built. Help us refine it. #ZurichJS #ZurichJSConf #JavaScript #WebDevelopment #Frontend #React #TechCommunity #CallForPapers #ConferenceLife

16

3 Kommentare

This post made me think about AI & podcasting - why do we use the word AI as a catch-all term? Who's it easier for? And what are we sacrificing without the specificity? Instead of being a podcaster who uses: - a local shaping algorithm for audio improvement - an LLM-based transcription tool - an LLM for processing and analysis of transcript, and; - a machine learning algorithm for social media discovery ...we say, "I use AI in my podcast." AI is a marketing term to entice people to use products. AI is a catch-all term to ensure a program or process is understood. Our post-production process doesn't need to be simple to understand. And in response to the question, "Do you use AI in your podcast?", why should the answer be Yes or No? I give you permission to be specific. Specificity is hard when you're doing text-based editing in Descript, but if you use a variety of tools with varying functions & algorithmic capacities, be specific. And to Stephanie Fuccio's point, lets start categorising and explaining what we use to assess its impact. If I make your Zoom call sound better, its not the same as a 3,000 episode per week production schedule with no human oversight. And if your apps are starting to stuff their packages with AI-based tools, explore what, why and how to figure out what's happening in there.

3

1 Kommentar

Vibe coding tip of the week... Don’t build a “proper” software project… …in a streamed Google Drive folder. 🤦‍♂️ Everything was going just fine— until I decided to do some “folder housekeeping”. Turns out, moving 10,000+ tiny files takes forever… The sync broke. My computer ground to a halt. I was left staring at a bunch of empty folders. Gulp. Thankfully, 16 hours later, I’ve patched things back together. Six months of work is safe (thank you GitHub 🙌). Sharing this as part of the warts-and-all life of a vibe-coder. Mostly so you don’t make the same mistake! The new approach: - Local dev folder (with GitHub and Time Machine) - Drive just for PDFs, spreadsheets, and docs Lesson learnt. (Very understanding) client’s patience tested. Onwards! 😊 #happyfriday #vibecoding #ai #aiagents

12

Stigg is an abstraction and orchestration layer that sits between your application and your billing provider (like Stripe), designed to give you flexibility and control over your monetization strategy. Learn how to decouple pricing from billing and move faster. 👇 https://lnkd.in/g3j8AYZ4

7

1 Kommentar

Check out my connection Tobias Uhlig latest article on Neo.mjs and the future of frontend development. He makes a compelling case for a new "AI-native" architecture that goes way beyond the current "AI-assisted" tools we've all been using. If you're a JavaScript dev tired of AI hallucinations and outdated docs, you'll want to read this. The idea of a self-improving codebase where AI acts as a foundational partner is a game-changer. Read the full article here: https://lnkd.in/e3MQRdmE #JavaScript #FrontendDev #AINative #NeoMJS #AI #Development

5

3 Kommentare