The Lab
Applied AI
10 articles on AI in UX testing, GEO, prompting and voice AI. Examples show how models are tested and used in workflows.
Transcription Models Compared: Deepgram vs. Whisper
Deepgram vs. Whisper on 898 hours of the German podcast ungeskriptet by Ben Berndt. With audio clips: where the models mishear and why the speaker matters.
Read articleAleph Alpha's Kolibri-1 Tested: AI “Made in Germany”?
Aleph Alpha's Kolibri-1 tested: in a rewrite it dropped a selling point, and in a jailbreak test 12 of 20 answers held explosives instructions.
Read articleGerman Health-Insurance Websites: What to Review First
56,198 pages from 84 German statutory health-insurance sites, AI-assisted and triaged. Which review patterns recur and how teams turn them into a work order.
Read articleWeekly AI Briefing Agent: What My Filter Really Needs
A working AI news filter, its prompt failures and three rules for relevance, sources and real tests instead of daily FOMO.
Read articleGemini Live in the Browser: A Voice AI Prototype
A browser demo with native audio and function calling. What it shows—and what remains unresolved for costs, privacy and real-world use.
Read articleSchema, llms.txt and Structured Outputs Explained
Schema.org helps search engines, llms.txt can guide agents and Structured Outputs enforce a format. Four layers that are often confused.
Read articleHanzi Movie Method: How It Works (+ Free Script App)
Actor, Set, Room, Props, Script: how the Hanzi Movie Method turns Chinese characters into movie scenes. A free browser app writes the scripts — no sign-up.
Read articleGEO: What AI Overviews Really Change About SEO
GEO remains SEO, but adds a measurement layer: whether content is found, cited and represented accurately in AI Overviews and ChatGPT.
Read articleHumanizer (German): A Claude Skill for More Human AI Text
Humanizer (German) for Claude Code and Codex: edit 72 German AI-writing patterns without Python. Open source under MIT and CC BY-SA.
Read articleVerbalized Sampling: More Varied AI Responses
A prompt for more varied responses, with rough model probabilities, a cost trade-off, and a clear selection rule.
Read article