About
Activity
341 followers
-
Sebastian Macke shared thisAs the first professional group, we have the privilege of working with AI agents that truly deserve the name. In the last months, it’s become clear that the craft is shifting from implementation to direction: specifying intent crisply, verifying relentlessly, and orchestrating tight feedback loops. In this blog post (English) and on LinkedIn (German), I share what changed, what matters now, and how to adapt. https://lnkd.in/dNP9s9Zj https://blog.qaware.de/
-
Sebastian Macke shared thisLast (German) AI newsletter of this year. Have fun! https://lnkd.in/dstjAcMhGemini 3, GPT-5.2 & DeepSeek 3.2: Das neue Kräfteverhältnis der KIGemini 3, GPT-5.2 & DeepSeek 3.2: Das neue Kräfteverhältnis der KI
-
Sebastian Macke shared thisWhy do language models sometimes just make things up? We’ve all experienced it: you ask a question, get a confident-sounding answer—and it’s wrong, but it sounds convincing. In this blog post, I describe with examples what the reason for this is and possible solutions. https://lnkd.in/eEyk_7wf
-
Sebastian Macke shared thisBei der Internationalen Matheolympiade zeigt die KI, dass sie fast jedes mathematische Problem lösen kann. Außer denen, für die man tatsächlich eine eigene Idee haben muss. https://lnkd.in/eqxxTrt3
-
Sebastian Macke shared thisRAG (Retrieval-Augmented-Generation) is just as poorly defined as the word Agent. You can decide for yourself whether an implementation is RAG, is not RAG, or whether RAG is perhaps already dead. In the newsletter (in German), I go into this in detail and talk about the future of RAG. https://lnkd.in/entPDsUX
-
Sebastian Macke shared thisJanuary was just crazy in AI. There were so many great papers and so much news. In two German newsletters, I write about emergent behavior in chain-of-thought prompting and about the new generation of tiny models with reasoning. https://lnkd.in/gePkgdFj https://lnkd.in/gk7M-8q6
-
Sebastian Macke shared thisI've been trying my hand at a newsletter with focus AI for the last few months. So far it has been a fun experience. https://lnkd.in/eQ72RjbPpublic_profile__posts
NEWSLETTER
QAware AI Insights
News und Einblicke rund um KI von unserem Experten Dr. Sebastian Macke. Alle 2 Wochen. Sorgsam recherchiert.By QAware GmbHPublished biweekly
-
Sebastian Macke liked thisSebastian Macke liked thisWe barely review code at Giant Swarm anymore. At least not in some of our teams :) That's not a quality problem though. Our coding agents simply produce so much code that keeping up with pull request reviews stopped being realistic. The first thing that changed: user stories disappeared from our standup. They became too small an entity to be worth planning. Our engineers now pick up whole epics, and an epic is defined by a customer outcome. What can this person do afterwards that they couldn't do before, and is that worth building at all? All the product exploration happens right there. But you can't just delete the review step from your SDLC. So we moved it. We review the plan now. Review shifted left. Roughly how a plan gets made: It starts with /grill-me, a skill from Matt Pocock. We adapted it a bit for our context, but it's generally excellent. It interviews you relentlessly about your idea, walks every branch of the decision tree and drags out all the assumptions you didn't know you were making. That conversation becomes a PRD. Then /ground-truth takes every open question and actually researches it. The real code, the web, how other tools solved this, whether the thing is even feasible. Hallucinations (should) die here. Then the fun one: /contrarian. It tries to kill your plan. It hunts for existing capabilities that already cover the problem and demands a written justification for every new piece you want to add. Verdict is BUILD, SIMPLIFY or KILL, and anything other than BUILD sends you straight back to grilling. And you know me...Almost every plan I have run through it came back with SIMPLIFY. And every single time it was right :D The output is a little website. Reviewers read the summary and drill into the PRD, the research and the contrarian verdict when they want details. That's what goes into a pull request, and that's what humans actually argue about. Once the plan is agreed, the agents implement it. Tests and test gates are part of the plan, so they know what "done" means. An agent does the code review. It merges itself. Result: we ship an epic in two to three days. I think this is far from the final shape of it, and not every team here works this way yet. But reviewing intent instead of output feels like a much better place to spend human attention. So how are you handling this? Because if your agents got fast too, that review bottleneck is coming for you :)
-
Sebastian Macke liked thisSebastian Macke liked thisQwen-3.8-Flash-Next is a big deal and a herald of things to come. Flash models are models distilled down from full size models, in this case from Qwen-3.8-Max, a 2.4T model requiring some 8 datacenters GPUs to run. The Flash model instead features 125B parameters, roughly 5% of the original model of which again only 6B are active per token (4%). Compare that to 27B active parameters of Qwen-3.8-27B dense, the current favorite for local AI. Supplementing the performance are an additional 51B ngram embeddings supplementing for lost “knowledge”in the distillation process that can be offloaded to system memory(!) The result is a model that can run in an optimized NV4 quant with 256k context, generous KV cache and multimodal image and video capabilities (!) on 94GB of VRAM - enough for a single Blackwell RTX 6000 at about 80 t/s or 128k context supporting multiple concurrent lanes with 130 t/s - Faster than it’s 27B brother and significantly better. Lower quants could conceivably run on something like a 3090 while the 4bit quant should still be able to push 50t/s on a 128GB+ mac studio None of this would matter if the performance didn’t match up and, all benchmarks aside, it’s phenomenal in my real world tests: I regularly port old dos games to the web using AI and no local model has been able to handle the tasks involved, in particular reverse engineering, driving a debugger and writing the required complex systems code until now. Qwen-38-27B, with a lot of handholding, was able to somewhat brute force itself through one of these tasks for me earlier this week, but it wasn’t economic, too slow and not competitive with cheap models like gpt-5.6-luna. Qwen-38-Flash-Next is something else. It’s the first local model that reaches Opus performance for me, maybe not 4.8 but better than 5 (lol Opus 5) and better than 4.6 which for many people was the “good enough” crossover point after which improvements become mostly optional. “Good enough for challenging engineering work” has arrived on local machines and judging from the buzz around GLMs flash model, there’s still a lot of room to go here All of which is very bad news for US labs, because whatever HBM shortages are stopping companies and consumers from having local capabilities are guaranteed to be temporary and these models scream commoditisation. Another takeaway: I was able to squeeze this model onto a 96GB card with a handful of prompts less than 24g after release. Gpt-5.6-sol was able to deploy it, benchmark configurations, work through bugs and make custom patches to the inference runtime (SGlang) autonomously. This matters A LOT because it shows you that “what’s in the weights” doesn’t constrain the technology much anymore. 5.6 does not know a thing about Qwen3.8, its architecture and trickery and it could still autonomously optimize it from search alone. Image: Test Drive III (1990) port to mobile web in progress using Qwen-3.8-flash-next on a local RTX-6000 MaxQ
-
Sebastian Macke liked thisSebastian Macke liked thisFable 5.1 is a significant improvement, more than most benchmark scores give credit. It autonomously reverse engineered Dragon Strike (1990) to playable, almost complete state and ported it to the web in 8 hours from a single prompt [0] The way I evaluate coding models, rather than drawing SVG Pelicans, threeJS scenes or similar tasks is to give them a real world job: "Autonomously reverse engineer this game and port it to web". Reverse engineering and porting Re complex, multi modal tasks and the process usually takes days to complete (or a week running on a single lane local AI system as seen here [1]) No model so far has fully autonomously finished the job from the initial prompt, usually they get stuck at about 50-60% completion and require handholding for the rest but Fable 5.1 is the first to get really close (85%), on about 1/2th of a pro max weekly fable budget no less (in part because the model uses subagents quite efficiently when instructed). The video below is the point at which Fable hit the first wall (the breath weapons are not working properly) I chose a game not popular enough to have any open source implementation or extensive format documentation - Unlike Syndicate [3] or Populous [4] which have ports and a huge fan community, for this game the model had to heavily rely on performing research evidence gathering and reverse engineering by running the game in a debugger and disassembling it's loaders and file formats. If you're on the fence about the impact of these models on the knowledge economy, consider that reverse engineering and porting is a high skill/experience, complex job involving multiple distinct disciplines - the kind of complex work only a small fraction of people out there have the experienced to perform. While there's a lot of hype and nonsense around the technology, the fundamental primitive underneath - we made all published human knowledge accessible and retrievable and we built a machine that can retrieve and transform/interpolate between representations - points to "we've taken scarcity of knowledge out of the knowledge economy", which, like taking oil out of the energy economy and replacing it with extremely cheap renewables, has dramatic consequences.. The question in these kind of economic disruptions is not however whether the technology is going to be valuable. It's who will profit, who gets to extract that value. And that's the real global battle unfolding right now. Anthropic would like to be that party, of course, and it raised unfathomable money, like OpenAI, to secure the riches of the AI frontier, or, as we put it, re-distributed value from white collar knowledge workers. At the same time, we're now running Opus 4.7 at home [1], indicating that much of the technology is math that has no moat and, like the calculator, won't make the manufacturers very rich. Interesting times
-
Sebastian Macke liked thisSebastian Macke liked thisTen years ago, we started the Cloud Native Night Meetup in Munich and Mainz with a simple idea: to explore, discuss and better understand emerging technologies together with the community. At the time, Cloud Native was still a pretty new idea. These days, it's basically the foundation of modern software engineering, and it's the platform that makes all the systems we're building today possible. The technology has changed a lot in the last ten years, but one thing hasn't: the value of an open community where practitioners share experiences, challenge ideas, and learn from each other. Our 10th anniversary is a chance to think about how we've got to where we are today. We've come a long way from containers and Kubernetes to platform engineering, cloud-native architectures, and now AI-native applications. So let's have a chat about what the next 10 years of software engineering might look like. I'm really looking forward to the evening's lineup. We'll kick things off with a panel discussion featuring Nico Meisenzahl, Max Körbächer, Matthias Haeussler, Sebastian Kister, and myself. We'll be chatting about the early days of Cloud Native, our individual journeys, the role of Cloud Native in today's AI landscape, and where we see the industry heading next. Then we'll hear from Robert Hoffmann, who'll explore how Amazon and other industry leaders made AI coding work, and finally Sebastian Kister will be delivering the closing keynote. "The Agentic Enterprise: Architecting People, AI and Sovereignty." We're looking forward to celebrating this milestone with you. Details & registration: München: https://lnkd.in/eDt7KYwB Mainz: https://lnkd.in/ej8zwzw8
-
Sebastian Macke liked thisSebastian Macke liked thisDie AI ist nicht mehr der Engpass: Warum jetzt wir das Limit sind Vor drei Jahren war die ehrliche Antwort auf viele AI-Anfragen noch: Das geht nicht. Coding-Agenten taugten bestenfalls für Unit-Tests, komplexe Excel-Modelle waren tabu, und „Chat with your data" scheiterte an den eigenen Versprechen. Das Limit war die Technik. Dieses Jahr hat sich das gedreht: Agenten arbeiten stundenlang selbstständig, gehen durch dutzende Dokumente und schreiben Software. Was heute über Erfolg oder Misserfolg entscheidet, ist das Gerüst drumherum. Die neue Ausgabe der QAware AI Insights zeigt, wo dieser Wendepunkt herkommt und was er für Unternehmen bedeutet: ➡️ Warum dieselbe AI mal bei 8 und mal bei 95 Prozent landet. ➡️ Was „Harness Engineering" wirklich bedeutet. ➡️ Wie eine AI-native Firma aussieht. ➡️ Woher die eigentliche Gefahr für etablierte Unternehmen kommt. Viel Vergnügen beim Lesen!
-
Sebastian Macke liked thisSebastian Macke liked thisFrom Architecture to Implementation of Agentic Systems In my previous article, I discussed the reasons why most AI initiatives never make it into production. However, Agentic AI presents another structural challenge: even validated, high-value use cases often fail to reach production. This is not because the models fail, but because today's IT architectures were not designed to orchestrate autonomous, non-deterministic AI agents securely and in a controlled manner. In my latest article, I introduce the Agentic Control Plane: an architectural approach that enables the operation of Agentic AI in production and helps organisations to turn promising use cases into sustainable business value. I also demonstrate how these principles are implemented in our Kubernetes-native, open-source project: the Agentic Layer.Agentic Control Plane and Agentic Layer: From Architecture to Implementation of Agentic SystemsAgentic Control Plane and Agentic Layer: From Architecture to Implementation of Agentic Systems
-
Sebastian Macke liked thisSebastian Macke liked thisAgenten auf dem Firmenrechner: Wer kontrolliert hier eigentlich wen? Wer morgens den Firmenrechner aufklappt, betritt eine ziemlich gut gesicherte Festung. Jahrzehntelang wurde sie auf menschliche Schwächen ausgelegt. Doch jetzt sitzt dort eine Software, die im Prinzip alles tun kann, was man selbst über Tastatur und Maus tun könnte. Die neue Ausgabe der QAware AI Insights geht der Frage nach, wie sich Coding-Agenten sicher kontrollieren lassen, ohne sie bis zur Unbenutzbarkeit einzuschränken: ➡️ Was der Coding-Agent alles kann: Wie ein Agent für eine scheinbar kleine Aufgabe selbstständig die Entwicklungsumgebung startet, echte Browserfenster öffnet, JavaScript einschleust und dabei wie selbstverständlich Screenshots in die USA überträgt. ➡️ Warum Abnicken keine Kontrolle ist: Warum Freigaben im Drei-Minuten-Takt zu Decision Fatigue führen und viele Entwickler direkt in den gefährlichen „alles erlauben"-Modus wechseln. ➡️ Das Chaos beim Sandboxing: Warum sich der Ausschluss sensibler Dateien wie secrets.env mit einem kurzen Skript umgehen lässt und warum sprachliche Anweisungen keine harte technische Grenze sind. Reicht menschliches Risiko als Maßstab? Warum das Ziel vielleicht nicht lauten sollte: „Der Agent kann nichts Gefährliches tun", sondern: „Der Agent verursacht seltener Schaden als ein Mensch mit denselben Rechten." Viel Vergnügen beim Lesen!Agenten im Firmenrechner: Wer kontrolliert hier wen?Agenten im Firmenrechner: Wer kontrolliert hier wen?
-
Sebastian Macke liked thisSebastian Macke liked thisJava-Legacy-Systeme: Zwischen technologischem Goldstandard und akutem Wartungsstau. Java ist das Rückgrat moderner Backend-Infrastrukturen. Doch viele langlebige Anwendungen stehen vor einer kritischen Schwelle: Über Jahre eskalierte Qualitätsprobleme, veraltete Bibliotheken und komplexe Monolithen bedrohen die Stabilität und Agilität im Business. Aktuelle Marktstudien (Konveyor & Lünendonk & Hossenfelder) zeigen ein klares Bild: Getrieben durch die wirtschaftliche Lage und den demografischen Wandel priorisieren Unternehmen in Deutschland die Modernisierung bestehender Systeme aktuell massiv gegenüber dem Neubau. Doch wie lässt sich diese Transformation planbar und kosteneffizient umsetzen? In seinem aktuellen Fachartikel „Legacy Reloaded – Strategien zur skalierbaren Java-Modernisierung“ beleuchtet QAware Lead Software Architect Andreas Zitzelsberger genau diese Praxis-Herausforderung. Er zeigt, wie der Spagat zwischen laufender Weiterentwicklung und risikofreier Modernisierung gelingt – von der datenbasierten Bestandsaufnahme bis hin zur KI-gestützten Beschleunigung. Schaffen Sie eine verlässliche Arbeitsgrundlage für Ihre Systemlandschaft. Der Fachartikel steht ab sofort kostenlos zum Download bereit. https://lnkd.in/daEfBXiQ zitzelsberger #Java #SoftwareArchitektur #LegacyModernisierung #SoftwareEngineering #QAware
-
Sebastian Macke liked thisSebastian Macke liked this📢 Reminder - Cloud Native Night: Ship It Safely: Mastering AI Agent Orchestration 🗓️ 09. Juni | 🕐 18:30 Uhr | 📍 Mainz & Remote | 🎤 Deutsch Reminder, CloudNativeNerds! Die nächste Cloud Native Night in Mainz rückt näher und diesmal sprechen wir über den Agentic Stack, autonome Systeme und die Zukunft moderner Softwareentwicklung. Freut euch auf zwei spannende Talks rund um AI Agents, agentische IDEs und die Frage, wie sichere und zuverlässige KI-Systeme im Enterprise-Umfeld entstehen. 🙌 🗣️ Talks: 💡 The Software Development Life Cycle (SDLC) in the age of AI Agents mit Harald Seipp, Principal Client Engineering EMEA Solution Architect & STSM und Thomas Suedbroecker, Senior Engineer and Solution Architect bei IBM Harald und Thomas zeigen, wie agentische KI den gesamten Software Development Life Cycle verändert - von der Anforderungsanalyse bis hin zu Deployment und Betrieb. Im Fokus stehen autonome KI-gestützte Entwicklungswerkzeuge, die Produktivität steigern, Entwicklerrollen neu definieren und eine verantwortungsvolle Einführung von AI im SDLC ermöglichen. 💡 Mastering Multi-Agent Systems with the Agentic Layer mit Florian Mallmann, Senior Software Engineer bei QAware GmbH Florian zeigt, wie das Open-Source-Projekt „Agentic Layer“ von QAware sichere, testbare und enterprise-taugliche KI-Agenten ermöglicht. Der Fokus liegt auf Agenten Orchestrierung, automatisierter Qualitätssicherung und Compliance-Guardrails. Der Talk liefert praxisnahe Einblicke und Live-Demos für den Aufbau skalierbarer und zuverlässiger Multi-Agent-Systeme. Lasst uns gemeinsam neue Perspektiven entdecken und Erfahrungen teilen - wir sind gespannt auf euch! 🫵 👉 Details und Anmeldung: https://bit.ly/cnnmz0626
Experience & Education
Volunteer Experience
-
Keyboard player in the 1st orchestra
Handharmonika Club Ditzingen
- Present 31 years
Arts and Culture
Active membership in the registered association of Handharmonika Club
Ditzingen
- 15 years playing the keyboard
- 2 years as youth supervisor
- 3 years as IT and homepage administrator -
Mentor
Google Summer of Code
- Present 11 years
Science and Technology
Implementing RISC-V emulation to the jor1k emulator
-
Faculty Lecturer
Rosenheim Technical University of Applied Sciences
- Present 5 years 1 month
Education
Lecturer for the lecture "Concepts of Programming Languages"
Projects
-
Concepts of Programming Languages
-
Lecture materials
Languages
-
English
Professional working proficiency
-
German
Native or bilingual proficiency
Organizations
-
Deutsche Physikalische Gesellschaft
-
- Present
Other similar profiles
Explore more posts
-
Quantum Science | University of Vienna
🐈⬛🤖 How big can Schrödinger’s Cat be, and what does that have to do with modern technology? Group leader Markus Arndt and his team investigate fundamental questions of quantum physics in the laboratory, using molecules and nanoparticles. In the video, Markus Arndt talks about what drives him — and demonstrates the potential of research that will shape the world of tomorrow. ⤵️
1 Comment -
ScoutinScience
🔬 𝐀 𝐧𝐞𝐰 𝐰𝐚𝐲 𝐭𝐨 𝐬𝐞𝐩𝐚𝐫𝐚𝐭𝐞 𝐚𝐧𝐝 𝐮𝐧𝐝𝐞𝐫𝐬𝐭𝐚𝐧𝐝 𝐧𝐚𝐧𝐨𝐩𝐚𝐫𝐭𝐢𝐜𝐥𝐞𝐬 A new study published by researchers from FAU Erlangen-Nürnberg, TU Bergakademie Freiberg, and University of Kassel explores a new way to separate nanoparticles using specially engineered supraparticle powders. 📄 Paper: 𝘚𝘶𝘱𝘳𝘢𝘱𝘢𝘳𝘵𝘪𝘤𝘭𝘦 𝘱𝘰𝘸𝘥𝘦𝘳𝘴 𝘢𝘴 𝘴𝘵𝘢𝘵𝘪𝘰𝘯𝘢𝘳𝘺 𝘱𝘩𝘢𝘴𝘦 𝘮𝘢𝘵𝘦𝘳𝘪𝘢𝘭𝘴 𝘧𝘰𝘳 𝘴𝘪𝘻𝘦-𝘦𝘹𝘤𝘭𝘶𝘴𝘪𝘰𝘯 𝘤𝘩𝘳𝘰𝘮𝘢𝘵𝘰𝘨𝘳𝘢𝘱𝘩𝘺 𝘰𝘧 𝘯𝘢𝘯𝘰𝘱𝘢𝘳𝘵𝘪𝘤𝘭𝘦𝘴 Authors: Umair Sultan | Lukas Hartmann | Céline Kohl | Allison Götz | Anna Krapf | Ralf Ditscherlein | Erik Löwer | Benoit Merle | Urs Alexander Peuker | Erdmann Spiecker | Martin Hartmann | Wolfgang Peukert | Benjamin Apeleo Zubiri | Malte Kaspereit | Nicolas Vogel 𝐖𝐡𝐚𝐭 𝐭𝐡𝐞 𝐫𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐢𝐬 𝐚𝐛𝐨𝐮𝐭 Nanoparticles are incredibly small materials used in areas such as medicine, electronics, and advanced materials. But because they are so tiny, it can be difficult for scientists to separate particles of different sizes and study them properly. In this study, researchers tested a new type of material called supraparticle powders. These are larger particles built from many smaller nanoparticles that cluster together, creating a structure with tiny pores and channels. The team used these supraparticles as the material inside a chromatography column, a tool scientists use to separate particles based on their size. 𝐖𝐡𝐲 𝐭𝐡𝐢𝐬 𝐦𝐚𝐭𝐭𝐞𝐫𝐬 Accurately separating nanoparticles is essential for research and industrial applications. Many technologies rely on nanoparticles with very precise sizes and properties, but existing separation methods can struggle to distinguish them clearly. 𝐖𝐡𝐚𝐭 𝐭𝐡𝐞𝐲 𝐟𝐨𝐮𝐧𝐝 Using gold nanoparticles as a test, the researchers showed that supraparticle powders can help separate extremely small particles by size: - The pore size of the supraparticles strongly influences how nanoparticles move through the column. - Different nanoparticle sizes exit the column at different times, enabling separation. - By tuning the structure of the supraparticles, scientists could potentially control how particles are sorted and analyzed. 𝐖𝐡𝐚𝐭 𝐢𝐭 𝐜𝐨𝐮𝐥𝐝 𝐦𝐞𝐚𝐧 𝐟𝐨𝐫 𝐭𝐡𝐞 𝐟𝐮𝐭𝐮𝐫𝐞 This approach could lead to more precise tools for analyzing and sorting nanoparticles. That could benefit research fields ranging from nanomedicine and catalysis to advanced materials and electronics. More reliable separation methods also help scientists better understand how nanoparticles behave, which is key to developing new technologies. 𝐖𝐡𝐚𝐭’𝐬 𝐧𝐞𝐱𝐭 The next step is testing whether this method works equally well for other nanomaterials. If successful, this could become a new platform for high-precision nanoparticle characterization. 🔗 Full study: https://lnkd.in/eEnWjJtX #Nanotechnology #Nanoparticles #MaterialsScience #ParticleScience #ResearchImpact #Innovation #ScoutinScience
1 Comment -
DIGZON
[PDF] Parallel Algorithms and Cluster Computing: Implementations, Algorithms and Applications Karl Heinz Hoffmann, Arnd Meyer https://lnkd.in/eYX2b9Ub This book presents major advances in high performance computing as well as major advances due to high performance computing. It contains a collection of papers in which results achieved in the collaboration of scientists from computer science, mathematics, physics, and mechanical engineering are presented. From the science problems to the mathematical algorithms and on to the effective implementation of these algorithms on massively parallel and cluster computers we present state-of-the-art methods and technology as well as exemplary results in these fields. This book shows that problems which seem superficially distinct become intimately connected on a computational level. digzon #simple #Engineering #ArndMeyer #KarlHeinzHoffmann https://lnkd.in/en5rbe6z
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content