Alexander Arvidsson

I’m in Redmond this week for the Microsoft MVP Summit - 4000 MVPs, 83 buildings across 500 acres, and not a single roadmap leak to share. The summit is NDA-covered, but that was never the point anyway. The most valuable thing you can do here isn’t sit in a preview session. It’s finding the engineer whose documentation you’ve been reading for years, buy them a terrible American coffee, and ask the question that doesn’t fit in a support ticket.

Non-determinism in LLMs isn’t a bug to fix - it’s a parameter to control. But most teams are tuning the wrong things: obsessing over temperature settings while ignoring the architectural choices that actually matter. Structured outputs, deterministic skills for calculations, plan-then-execute agent patterns, retrieval logging - these are what separate systems you can trust from ones that roll dice in production. The goal was never a deterministic LLM. It was knowing exactly where randomness is allowed to live.

Power BI User Days in Utrecht has something most events don’t - soul. But beyond the atmosphere, something more interesting is happening: attendees are no longer asking how to build things. They’re asking why, and when, and whether. A maturing community, combined with LLMs absorbing the basic how-to questions, is shifting the conversation toward applied thinking. It’s the shift I’ve been waiting for. Utrecht suggests the appetite is finally there.

For every hour on stage, a hundred hours of preparation. That ratio stopped feeling like a personal failing when the blog became part of the process — not as a publishing platform, but as a thinking tool. A post forces the argument into shape before the slides exist. If the logic holds on the page, it’ll hold on the stage. If it doesn’t, better to find out now than in front of a sold-out room in Utrecht.

Non-determinism in LLMs creates real operational problems: flaky tests, irreproducible bugs, compliance nightmares, and unreliable agents. Most people only know about token sampling, but randomness creeps in across six distinct layers—from floating-point variance to hidden system prompts. Temperature=0 and random seeds help less than you’d hope because they constrain token selection, not reasoning paths. The solution requires structural constraints, not parameter tuning. Until then, you’re rolling dice in production.

One of the most difficult events to get into for me has always been the Power BI Gebruikersdagen - a.k.a the Dutch Power BI User Days. I’ve always held it in very high regard, and I never stopped submitting in the hopes of one getting to speak. Last year I was invited to deliver a half-day session on presentation skills and a normal one-hour long session on data modeling. I had a blast.

Using Claude Code, I built a full-stack application to replace Microsoft Access - without being a developer or understanding React, TypeScript, or Node.js. LLM-assisted coding enables rapid prototyping and bridges the gap between business requirements and technical implementation, but like flight simulators, it doesn’t make non-developers into developers. The parallel matters: functional prototypes aren’t production-ready systems, and knowing the difference requires actual expertise.

Prompt injection is architecturally unfixable in current LLMs, but defense-in-depth works. Training-time defenses like Instruction Hierarchy, inference-time techniques like Spotlighting, and architectural isolation create practical systems. Microsoft’s LLMail-Inject showed thatadaptive attacks succeed at 32% against single defenses, 0% against layered approaches. Real failures like GitHub Actions compromise prove that securing obvious surfaces isn’t enough. Like SQL injection, it’s manageable with layering.