Writing.
Essays on AI, product marketing, and building small things. Newest first.
2026
on the upsun blog
guest posts elsewhere
- nov 2026 A human in the loop can't be your whole AI security plan Argues that "human in the loop" doesn't settle the AI security question: agent permissions should be scoped to the consequences of failure, with low-risk, reversible work running freely inside tight boundaries and people stepping in deliberately, before the consequential action, rather than rubber-stamping endless approval prompts. cybersecurity-insiders.com
- nov 2026 Overcoming the 'AI is Magic' Mindset: Why Learning How AI Models Work Must be a Companywide Priority Argues that companies need to stop treating AI as a black box: because model outputs are probabilistic, not deterministic, safely deploying AI agents requires real infrastructure-based guardrails, transparent auditing, and controlled permissions, not just trusting the model to behave. unite.ai
- oct 2026 KI-Agenten: Warum technische Grenzen entscheidend sind Argues that prompt instructions alone aren't a real technical control: AI agents need hard technical limits on consequential actions, with autonomy earned incrementally. itdbusiness.at