Mathias Bæk Tejs Langhede

As part of an educational company, learning is central to us. The Nordic Region’s AI Champions Network gives our colleagues the space to explore AI hands‑on, learn from experts, and be inspired by best practices. It reinforces Sanoma’s commitment, as a team, to continuous upskilling to deliver the best learning solutions 😊 #SanomaLearning #Sanoma #AI #AIEducation #FutureOfEducation

16

1 Kommentar

I've been seeing debates about LlamaIndex vs LangGraph in some sites.. Based on 18+ production deployments tracked by Alice Labs, LangGraph ranks #1 for anything requiring explicit control: graph state, retries, human-in-the-loop flows, and time-travel debugging. It models agents as finite state machines, which sounds nerdy but is incredibly useful when your workflow has conditional branches and failure modes. LlamaIndex still wins when retrieval is the core of what your agent does. If your agent needs to reason over documents, APIs, or proprietary data, LlamaIndex's depth in that space is hard to beat. It significantly reduces hallucinations in document-heavy use cases. So, there's no universal winner. LangGraph for flexibility, LlamaIndex for RAG, The article covers many more insights about other frameworks, so check it out! https://lnkd.in/dYXzkCzV

3

1 Kommentar

❓Attending the Danish Digitalization, Data Science, and AI - D3A next week? On Tuesday (August 26th), Professor Jesper Ryberg will present on the ethical challenges of using AI in criminal sentencing. Frej Thomsen, who will be joining the Centre in September, will also give a talk in the same session on the ethics of AI-generated Child Sexual Abuse Material. 💡 Abstract for Jesper’s talk Artificial intelligence is becoming widely embedded in social institutions, including the criminal justice system. However, it is generally agreed that using AI to assist judges in meting out sentences gives rise to ethical challenges. This presentation provides an overview of the main ethical concerns, and argues that an overlooked complication is that, according to our best theories of punishment, many offenders are currently being punished in ways that are not ethically justified. This complication fundamentally affects how we ought to address the standard challenges associated with the use of AI in sentencing. 💡 Abstract for Frej’s talk The advent of widely available high-performing AI for image-generation has also made it easy to create morally challenging material. A central example is AI-generated child sexual abuse material (AI-CSAM). AI-CSAM differs from conventional CSAM in that producing AI-CSAM need not harm victims. This presentation explores alternative explanations for why producing and distributing AI-CSAM is morally bad, and discusses the policy implications including a controversial use-case: Police use of AI-CSAM for the purposes of pursuing and apprehending offenders. https://lnkd.in/dwUqCXuu #AI #Ethics #AIEthics #Law #technology #criminaljustice #research

15

An absolute pleasure to give the first recorded Life In Quantum Lecture at our NNF Quantum Computing Programme (NQCP). The talk is on recent work from our team on ultra-fast estimation of T1 (https://lnkd.in/e2sMJweJ, Fabrizio Berritta), where we used an on-FPGA Bayesian agent to observe 100x faster T1 dynamics in superconducting transmon qubits than reported before, and our related work on calibration and benchmarking of single-qubit properties on these subsecond timescales, also using on-FPGA control, optimization and analysis techniques (https://lnkd.in/erRrx3cn, Malthe Asmus Marciniak Nielsen). I also sneak preview a couple of upcoming manuscripts from our team on extensible Lindblad dynamics learning, hyper-fast parallelised robust phase estimation, and fun entanglement polytope implementation for efficient entanglement certification :-) Shout out to Stuart Woods and Clara Sika Helbo for starting the new tradition of filming these talks and Kim Splittorff for organizing these events!

58

How do you train AI for mental health when patient privacy makes data hard to get? 🤔 Our new paper co-authored with the amazing Fredrik Sand Aronsson in Frontiers in Digital Health explores a "synthetic-only" approach to clinical NLP. We used Large Language Models to generate thousands of synthetic patient narratives paired with clinical scores (like PHQ-9 for depression and PCL-5 for PTSD). We then trained standard ML models on this data to see if they could "learn" the clinical signal. The result? The models successfully predicted mental health scores with high accuracy solely from synthetic text. This isn't about replacing real clinical data, but about creating a privacy-preserving fast lane for research. It allows us to prototype, test, and debug complex models without needing immediate access to sensitive medical records. Check out the full study to see how synthetic data might accelerate the future of digital health! 👇 🔗 Link in comment #AI #MentalHealth #DataPrivacy #SyntheticData #NewResearch #FrontiersInDigitalHealth

39

9 Kommentarer

The codebase AI couldn't read. Lemonsoft Oyj engineering team adopted Copilot, Cursor, and ChatGPT back in 2025. Routine work decreased and team sentiment with AI-coding was positive. On paper, everything looked good, but something was quietly wrong, thought Janne Tammi, Lemonsoft's CTO. Each AI coding session started from zero - no memory of architectural decisions made the week before, no recollection that a certain approach had already been tried and abandoned. With ten developers starting fresh ten times a day, that added up to a hundred zero-start sessions every day. The AI produced code that was syntactically correct. However, it didn't fit the patterns, respect the conventions or account for decisions already made. The answer wasn't better prompts or different tools. The codebase itself needed to become a structured knowledge layer, something AI agents could query accurately and build on reliably. This is when Lemonsoft CTO called ModernPath. "I wish I had shared his confidence. Honestly, I didn't believe the project would succeed." Janne Tammi, CTO, Lemonsoft Read the full story: link in comments #AIcoding #LegacyModernization #EnterpriseAI #SoftwareDevelopment

35

1 Kommentar

Magic can happen if you embrace GitHub Copilot assisting your code! Listen to these great speakers, bring your code if you'd like and explore your future way of working.

8

In the ASCHAS pre-study, we (Efi Papatheocharous Anders Wallberg Sara Janhäll Rolf Blom and I) explored how to connect the STPA risk analysis method to attack trees in the cyber security domain. We also started work on a tool that supports the user during the modeling process. Read more in this blog post https://lnkd.in/dRpjwhcR RISE Research Institutes of Sweden Cybersäkerhet på RISE

18

1 Kommentar

"With the release of coding agents, working with both frontend and backend code at the same time has become simpler." Eivind Kjosbakken expands on some best practices in his latest article.

1

Vesiro, a spinout from Chalmers University, has secured €1.6 million in seed funding to tackle one of the biggest challenges in modern data infrastructure: search efficiency at scale. The round was backed by Chalmers Ventures, Industrifonden, Länsförsäkringar Göteborg & Bohuslän, Yuncture, First Gate Invest, E14 Invest, Mach One, and a group of angel investors including Stefan Mahlstein, Staffan Truvé, Jan Sparud, Petter Eriksson, and Magnus Lundgren. As global data generation accelerates, data centres—already responsible for ~1% of global energy consumption—are becoming increasingly costly and energy-intensive. Vesiro addresses this with a performance-focused plug-in for Elasticsearch, enabling organisations to process large datasets faster and with fewer servers. According to internal benchmarks and pilot deployments, Vesiro’s technology allows customers to maintain the same workloads while significantly reducing infrastructure needs—lowering both costs and energy consumption. “Data volumes are growing faster than today’s infrastructure can handle, and our technology makes it possible to analyse large datasets using far fewer servers,” — Oskar Hagman, Co-founder & CEO, Vesiro The funding will be used to expand the technical team, accelerate product development, and support market rollout, targeting Elasticsearch-heavy use cases across business intelligence, e-commerce, and AI. 🔍💡 Smarter search. Fewer servers. Lower energy use. #Funding #SeedRound #Elasticsearch #DataInfrastructure #GreenTech #ClimateTech #Sweden #Startups #EnergyEfficiency

5

Embedl and #HuggingFace in the same box! Embedl is building a hub solution for inference on the edge: https://lnkd.in/deTFut75 AI models running locally on chips other than Nvidia GPUs present an adoption opportunity that can democratize and embed AI in a wide variety of products and solutions. But AI on the Edge comes with its own set of problems - models don't compile (operators aren't supported, quantization doesn't work), or are too slow and memory intensive. The dev tools needed to deal with such problems stitching together Python code with compiled binaries, .so files and runtimes from vendors such as Qualcomm, TI, Hailo, Samsung etc are the real MVPs that will determine the value AI can generate. In the past two years of working at Embedl in this area, I have seen some incredibly cool use of AI at the edge. Excited to see what the future holds and how we can contribute to it. Check out our hub or reach out if you are interested in running AI models locally on the edge. We will also be at #EurIPS in Copenhagen to talk about our research and development in enabling AI at the edge.

8

Partial deadlocks have been a major problem in Go programs for many years. We discovered that by piggybacking on the garbage collector, it is often possible to efficiently detect and recover from such errors. This was a masterpiece of Georgian-Vlad Saioc's PhD work in collaboration with Milind Chabbi and Angelina Lee at Uber. Great to see that the implementation will now be broadly available in Go 1.26! 🥳 📜 https://lnkd.in/eR9sWxtN

78

1 Kommentar

Agreed! "Switching to Linux is not just the wise choice, it's a great choice. Europe should be all-in on Linux for public works, public computing, and all new system development."

1

Today starts the LUMI supercomputer summer school in High-Performance Computing for young researchers and students at Nuuksio, Espoo (https://lnkd.in/dW86Jqdn) . I am happy that DeiC is supporting 3 danish students and/or researchers to join the summer school this year. We hope that maybe next year we can support one from each university in Denmark. Why is this important? It is important to support young talents in programming HPC applications using MPI, OpenMP and other HPC frameworks and programming paradigms and learn the pitfalls of large scale computing (if not to avoid them, then know when you hit them hard and crash and burn). Read more here: https://lnkd.in/d5f2vxp9 Without these talents it is hard to develop the community in Denmark. #supercomputing #hpc

29

I judged an AI Hackathon Zero2Demo for DTU x CBS students over the weekend - and it completely flipped one of my “AI-Assumptions”. I used to think AI would mostly replace junior roles. Now I’m not so sure anymore. Students using AI today can reach “expert-level” execution absurdly fast. They build things in hours that used to take teams months. And sure – not everything can be solved by predicting the next token in a model. Some experience comes from years of scars, bad decisions, and awkward meetings you wish you could undo. But let’s be honest… a surprising amount of “experience” is really just pattern recognition. And AI is very good at patterns. (Sometimes too good… it spots patterns that don’t even exist.) So who’s at risk? Not juniors, not seniors either. It’s anyone, at any level, who stops adapting. AI won’t reward seniority nor experience. It will reward curiosity and speed of learning - which is great news for students.  Massive shoutout to the partners who made this possible: Antler, byFounders, node.vc , People Ventures, Creator Fund, Wave Ventures, United Founders, Lovable, Cursor, TechBBQ, and all other incredible sponsors. Huge kudos to the organizers Matyas Adam Michel, Milan Aha, Adam Bočev, and Janik Pöttinger!

80

7 Kommentarer

Exciting! Three speakers at one PyLadies Stockholm event! 💃💃💃 This doesn’t happen very often. The last time it happened at our 10-year anniversary 🥳✨ And… I guess it’s time for me to practice my public speaking again 💬😅 For my talk, I’m diving into something we all rely on but rarely question: 🚨 Alarms. Sensor sends data → alarm triggers → notification hits your phone. Simple, right? Yes and no. When I started working with alarms, I discovered some surprisingly deep and interesting concepts behind what seems simple on the surface: 🌳 Abstract Syntax Trees — used to define, validate, and evaluate complex rule expressions 🔄 Finite-State Machines — when alarms have multiple states and transitions in between them ⚡ Event-driven architecture — because when something big happens, we want to know immediately, right? And once you introduce events… you also introduce some pitfalls. In this talk, I’ll walk through a tiny Python-based alarm system and we explore: How rules are evaluated How events flow through a system Pros and cons of event-driven design So, dear ladies, if you’re curious to know more about how it all react behind the scenes — and what actually happens when 'something happens' — we will be happy to see you 💜 P.S. the event is full, but waiting list is available and there is always happen that many people change their RSVP and other take their spot 😉

3

Anders Hejlsberg - creator of Typescript (and more) says AI started out being the assistant. But now it’s switching around. The AI is doing the work and it doesn’t need an IDE in the same way that we need IDEs. https://lnkd.in/etFxzzde

3

Anders Hejlsberg - creator of Typescript (and more) says AI started out being the assistant. But now it’s switching around. The AI is doing the work and it doesn’t need an IDE in the same way that we need IDEs. https://lnkd.in/etFxzzde

7

What is the difference between a Petri dish and a Petri net? One is for growing tiny living things and the other is a model language for distributed systems. We combine both in our 2 June Open Game Data office hour, when Jeremy Roberts will be demonstrating a research platform for modeling learning game interactions as Colored Petri Nets, generating synthetic player telemetry with Bayesian Knowledge Tracing, and analyzing the results with process mining techniques. Jer is the Managing Director of The Incredible Pear. He serves on the Open Game Data steering group, and was formerly the Senior Director of Learning Technology for PDS KIDS Digital. Register today! https://lnkd.in/euQBFa-V #OGDOOH #LearningGames #gametelemetry #Petrinet #EdTech #publicmedia #PBSKids

8