Alexander Arvidsson

A Wharton study put 1,372 people in front of a deliberately wrong AI. 73% followed it anyway. Not from carelessness - from fluency, confidence, and the absence of a trained skeptical reflex. We built that gap ourselves: ten years of certifications that taught procedure and skipped judgment. Now the procedures belong to agents, and the engineers we trained have no independent model to cross-check against. This is about what we teach instead - and why broken things are the best place to start.

Most speakers spend weeks building a presentation and zero minutes thinking about how to end it. Then they hand the microphone to the audience and call it a close. That’s not a close - it’s an abdication. The serial position effect means your ending is the thing most likely to survive in memory. This post breaks down why Q&A endings undermine you, why stories work at a neurological level, and what the research actually recommends instead.

SQLBits is in Newport, Wales this year - a place that requires actual planning to reach, and a useful reminder that modern travel infrastructure is more remarkable than we give it credit for. I’m delivering two sessions: a workshop on data communication with Valerie Junk, covering why insights fail to drive action and what to do about it, and a Fabric FinOps session on what Microsoft Fabric workloads actually cost when you measure outcomes, not just Capacity Units.

You can delete a post. You can unpublish an article. You can invoke GDPR and demand erasure. What you cannot do is remove something from an LLM. Once data dissolves into billions of model weights, there’s no row to delete, no file to erase. And it gets worse: the models now training on AI-generated outputs are degrading with each generation, narrowing toward a statistical echo of themselves. The right to be forgotten has no technical implementation path. None.

Goodhart’s Law - when a measure becomes a target, it ceases to be a good measure - is quietly undermining data strategies everywhere. Support tickets get closed without being solved. Active users log in and leave. Velocity numbers climb while codebases rot. The problem isn’t bad metrics; it’s the assumption that better metrics fix the underlying issue. The real question to ask before deploying any KPI: if we optimize for this, what will break?

AI tools force knowledge workers into constant task-switching that neuroscience proves is cognitively expensive. Only 2.5% of people can effectively multitask, yet we’re asking everyone to perform like fighter pilots without selection or training. Berkeley researchers found workers experienced burnout as managers consumed efficiency gains with increased assignments—repeating the Jevons Paradox from the telegraph and typewriter eras. Organizations measure output volume while ignoring cognitive sustainability, creating hidden supervisory labor that extracts maximum throughput regardless of cost.

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.