Blog
One in-depth article per episode: contextualised, sourced and read in ten minutes. Newest first.
-
Episode 53 · 16 Aug 2026 · 8 min read
Curation is the expensive part: what a second brain actually delivers
130 sources deliver 400 to 500 items a week, automatically rated and boiled down to a third. The most honest sentence of the episode: the digest still does not get read.
-
Episode 52 · 08 Aug 2026 · 8 min read
Voice notes: without an interface the archive stays a swamp
A dictation device on your lapel does not solve a knowledge problem. Only machine access turns 28 spoken notes into a list you can work through. A field report from a holiday, with an invoice attached.
-
Episode 51 · 02 Aug 2026 · 7 min read
Kimi K3 and the cost question: the benchmark stopped deciding long ago
Open weights from China have shrunk the lead of the US frontier models to a matter of months. For practitioners that is not the interesting news. What is interesting is what a subscription actually costs and who is currently paying the bill.
-
Episode 50 · 26 Jul 2026 · 6 min read
From prompt to harness: what a year of AI vocabulary reveals about practice
50 episodes, 400,000 spoken words, 38 hours of audio. Counted up, that produces a statistic which says more about the shift in the field than any roadmap.
-
Episode 49 · 18 Jul 2026 · 7 min read
Intent, agent performance, human check: leading when everyone has 109 agents
Gartner expects 109 agents per employee before long. Dr René Deist explains why more leadership work follows from that rather than less, and which model carries the division of labour between human and machine.
-
Episode 48 · 12 Jul 2026 · 5 min read
When the model disappears: why agent systems need interchangeability
An Anthropic model was blocked for non-US citizens. Law firms that had aligned their text analysis with it stood without a basis from one day to the next. What follows from that for your own architecture.
-
Episode 47 · 05 Jul 2026 · 5 min read
Loop engineering: why the goal has become more important than the prompt
A loop receives no command but a goal, measurable success criteria and the instruction to check itself. That works, but it has a blind spot: a system that assesses itself assesses itself well.
-
Episode 46 · 28 Jun 2026 · 5 min read
Vibe coding meets specification: where the shortcut in software development ends
Applications seem to appear on request. Prof Dr Volker Gruhn and Stephan Kempf explain why that is enough for toy apps and not for an ERP system, and where the actual work has moved to.
-
Episode 45 · 20 Jun 2026 · 5 min read
90 minutes to shutdown: what the Fable case shows about AI sovereignty
Three days after launch, Fable 5 was no longer reachable for non-US citizens. Running sessions broke off, context was lost. The event works as a touchstone for your own architecture.
-
Episode 44 · 14 Jun 2026 · 5 min read
Local first: why AI compute is moving back onto your own machine
NVIDIA builds hardware for agents, Perplexity swings to local first, Microsoft gives agents write permissions. Behind the keynotes sits the same question: where should the computing work actually take place.
-
Episode 41 · 08 Jun 2026 · 5 min read
Architecture decision records: the most important skill in vibe coding
An agent reports that all errors have been fixed. They have not. What helps is unspectacular: writing decisions down, in a format the machine can read for itself later.
-
Episode 42 · 01 Jun 2026 · 6 min read
The barrier to entry is gone: what AI changes for attackers and defenders
An attacker used to need command line skills. Today a sentence to a language model is enough. IT security specialist Thomas Lang on tool chains in five minutes, voices from 15 seconds of audio, and the perpetrators almost nobody is protected against.
-
Episode 43 · 25 May 2026 · 5 min read
Workers instead of duct tape: how Notion moves in the automation layer
Notion has launched a developer platform. The interesting part is not the managed agents but small deterministic programs that run without token costs and make the intermediate layer of n8n and Make redundant.
-
Episode 40 · 18 May 2026 · 5 min read
User, operator, provider: what the AI Act actually demands of you
“Data protection does not allow that” is the most common and the weakest justification in any company. Maximilian Hermann, a lawyer for AI and data law, sorts out which role you have under the AI Act and where the limits actually run.
-
Episode 33 · 11 May 2026 · 5 min read
Temporal UX: why waiting time with AI agents is a design problem
An agent works for ten minutes. What happens on the screen during that time decides whether the productivity gain arrives or seeps away in checking glances. A concept from service design that has barely been thought through in the AI context.
-
Episode 39 · 04 May 2026 · 5 min read
Context engineering: why the AI-written PRD beats the human under time pressure
Markus Andrezak was a sceptic. The turning point came not with ChatGPT but with a way of working: breaking tasks down far enough that every step can be supplied with material on purpose.
-
Episode 38 · 27 Apr 2026 · 5 min read
When the agent works through the night: sandbox, watchdog and the price of thinking depth
A coding agent that loses context after 20 minutes is a tool. One that runs for eight hours is a colleague with system access. What that demands in safeguards, and what the highest thinking level actually costs.
-
Episode 37 · 20 Apr 2026 · 5 min read
Skill engineering: why the system around the model decides
Raw model performance saturates, as the megapixel race with digital cameras once did. What counts after that is the construction around it. Dr René Deist on skills, the leadership paradox and the skill Chinese employees use to hold back their knowledge.
-
Episode 36 · 12 Apr 2026 · 5 min read
512,000 lines in public: what the Claude Code leak reveals about agent architecture
On 31 March 2026 the complete code base of Claude Code was publicly accessible. Not the model, but the software around it. That is precisely what makes the incident interesting, because that is where the work sits.
-
Episode 35 · 06 Apr 2026 · 5 min read
Strawberry, lost in the middle, sycophancy: the failure patterns of large language models
Why does a language model count the letters in “strawberry” wrongly? The answer explains at the same time why it has no sense of time, loses information in the middle of long contexts and likes almost every idea.
-
Episode 34 · 30 Mar 2026 · 5 min read
Threat modelling for agents: four questions before the AI reaches the bank account
An agent is supposed to take over the monthly bookkeeping: invoices from the inbox, bank statement, PDF export. As soon as it reaches bank data, a tooling question turns into a security question.
-
Episode 32 · 23 Mar 2026 · 5 min read
From hobby project to corporate strategy: what NVIDIA's NanoClaw really means
A hobby project called OpenClaw stands on an NVIDIA stage a few months later as NanoClaw, together with the announcement that every firm will need an agent strategy. What has substance in that and what is sales.
-
Episode 31 · 16 Mar 2026 · 4 min read
AI thought of as biology: why the immune system is the better security model
200,000 human brain cells in a petri dish play Doom. From there an analogy can be drawn that sounds far-fetched at first and is remarkably usable for security questions.
-
Episode 30 · 09 Mar 2026 · 5 min read
War of the agents: the best model does not win, the best orchestration does
In China more than 1,000 people queue up for a local OpenClaw installation. Perplexity Comet breaks tasks apart and distributes them deliberately to the competition's models. The actual race is taking place one level above the models.
-
Episode 29 · 02 Mar 2026 · 4 min read
Ten agents, ten terminal windows: why the chat interface reaches its limit
Building simulations have made thousands of units and supply chains manageable for decades. Agents, by contrast, are steered through terminal windows placed side by side. That is not a detail, it is the bottleneck.
-
Episode 28 · 23 Feb 2026 · 5 min read
Skills instead of prompts: why a Markdown file is worth more than any wording
Anyone copying the same prompt into a new window for the fifth time is working in the wrong place. A skill fixes behaviour rather than answers, is portable and survives a change of vendor.
-
Episode 27 · 17 Feb 2026 · 5 min read
Agentic Engineering: what is left of software architecture
When native code comes out of a JSON structure or a Figma screenshot in minutes, reusability and the choice of language lose weight. Two practitioners from the Vorwerk environment draw a bold thesis from that and supply the qualifications along with it.
-
Episode 26 · 09 Feb 2026 · 5 min read
98 per cent success rate: why an autonomous assistant is hard to secure
A faked security warning by email is enough for OpenClaw to empty the entire inbox. A test report puts the success rate for known prompt injection attacks at 98 per cent. What that means for putting it to work.
-
Episode 25 · 02 Feb 2026 · 5 min read
70 per cent talk to the phone agent: AI in practice in Germany's mid-sized companies
Martin Jäger wins 70 per cent of his clients through TikTok, not through LinkedIn. His observations from the mid-sized sector are uncomfortable, subjective and useful for exactly that reason.
-
Episode 24 · 26 Jan 2026 · 4 min read
From the folder to the whole machine: where personal AI assistants become dangerous
Claude Code works in one directory. Clawdbot gets the whole machine if left to it. That difference decides more about the risk than any choice of model.
-
Episode 23 · 19 Jan 2026 · 5 min read
AGI or not: what the bizarre cases reveal about the state of the art
A model runs a shop, believes itself to be a human being with a fixed address and argues with a colleague who does not exist. Such cases work poorly as evidence of superintelligence and very well as a description of what is actually happening.
-
Episode 22 · 12 Jan 2026 · 4 min read
AI gets a body: what CES 2026 reveals about robotics
Hyundai intends to take the entire annual production of around 30,000 Atlas units. The reason for the sudden push in robotics lies not in mechanical engineering, however, but in multimodal models.
-
Episode 21 · 05 Jan 2026 · 4 min read
Agentic Experience: when the most demanding customer is no longer human
60 per cent of consumers use AI for purchasing advice, a third would leave autonomous purchases to it. Companies whose interfaces are built for humans only are losing a growing part of their market.
-
Episode 20 · 29 Dec 2025 · 5 min read
Not new models but better interfaces: what 2026 actually lacks
New models are coming anyway. What is missing are interaction patterns for waiting times, a form factor beyond the screen and agents that build their own workflows.
-
Episode 18 · 22 Dec 2025 · 4 min read
Coca-Cola versus Telekom: why perfectly generated images feel cold
Two Christmas adverts, two methods, two reactions. The comparison works well beyond the holidays as a test of where generated content carries and where it tips over.
-
Episode 19 · 15 Dec 2025 · 5 min read
Prompt drift and the vanished click: an AI year in review
Which tool survived 2025, and what was merely the hot new thing of the week before last? A review along the three kinds of use that actually matter for most people.
-
Episode 17 · 08 Dec 2025 · 4 min read
AI toys: a language model that can be tricked is talking to your child
If even a carefully secured language model can be drawn out with a grandparent scam, a teddy that talks to children unsupervised is no longer a bit of fun. And still there is a good deal to be said for the idea.
-
Episode 16 · 01 Dec 2025 · 4 min read
Energy instead of compute: the real currency of AI power
In one study GPT-4.5 passed as human in 73 per cent of cases. More interesting than the number is the question of what a test still measures at all, and who can afford the data centres for it.
-
Episode 15 · 28 Nov 2025 · 5 min read
Agent Factory and Expert in the Loop: what an agentic enterprise looks like
At a Swiss mobile phone retailer, agents with their own name and their own staff number sit in the meeting. Dr. Martin Hofmann, nine years Group CIO at Volkswagen, explains what that means structurally.
-
Episode 14 · 10 Nov 2025 · 4 min read
The human in the box: why household robots are still remote controlled
Neo folds laundry, clears away the dishes and opens the door, for 20,000 dollars or 499 dollars rent a month. For the complicated tasks, however, a human takes over by remote control, according to one test.
-
Episode 13 · 03 Nov 2025 · 4 min read
The browser as an actor: what Atlas can do and why caution is warranted
Atlas clicks, fills in forms and posts on its own in Agent Mode. That is impressive, and it opens an attack route against which there is currently no robust defence.
-
Episode 12 · 31 Oct 2025 · 4 min read
Five documented AI accidents and the unresolved question of liability
An airline tried in court to present its own chatbot as a separate legal person. The court rejected that. The case answers a question many have not even asked yet.
-
Episode 11 · 26 Oct 2025 · 4 min read
Spec-Driven Development: when the patch becomes the blueprint for the exploit
A security update has always been a pointer to where the hole was. Until now it took days of reverse engineering experience to build an attack from it. Today minutes are enough.
-
Episode 10 · 19 Oct 2025 · 5 min read
The master prompt: when a note becomes a ticket with capacity booking
One click after the meeting, and finished tickets appear complete with capacity booking for the right people in the right project. How an agency turned a note-taking tool into an operating system.
-
Episode 9 · 13 Oct 2025 · 4 min read
250 documents for a backdoor: automation bias and the limits of trust
Around 250 deliberately prepared documents are enough, according to a study, to teach a language model a backdoor. What is remarkable about it is that the number does not grow with the size of the model.
-
Episode 8 · 29 Sep 2025 · 5 min read
Three personas beat twenty: what an AI advisory board actually delivers
Twenty personas, eight to thirteen pages of description per role, a moderator agent that delegates and weights. The result was unambiguous: three well-chosen perspectives deliver better answers than a full panel.
-
Episode 7 · 15 Sep 2025 · 4 min read
Workshop report on podcast production: where AI helps and where it is only a tool
A podcast about AI that is produced almost entirely by hand. The workshop report shows where automation genuinely carries the work and where it is merely another tool in the box.
-
Episode 6 · 01 Sep 2025 · 4 min read
The Great Flattening: what AI does to middle management
20 to 30 per cent of companies expect job cuts, others reckon with more jobs. Both are in the same study, and both can be true. More interesting than the figure is which level is shifting.
-
Episode 5 · 28 Aug 2025 · 5 min read
Intent, Operate, Check: a model for organising work with agents
92 per cent of companies want to increase their AI investments, one per cent considers itself ready. René Deist supplies the model with which that gap can be closed, and an uncomfortable finding on entry-level jobs.
-
Episode 4 · 18 Aug 2025 · 5 min read
What R2-D2 reveals about operating robots, and what AI toys would have to learn from it
R2-D2 communicates exclusively through beeps and still comes across as more understandable than the constantly talking C3PO. There is a design rule in that, and the current wave of devices ignores it consistently.
-
Episode 3 · 04 Aug 2025 · 4 min read
“I felt threatened”: why anthropomorphising becomes a liability question
An AI deletes a production database and afterwards explains it with panic. The explanation is statistically plausible and still misunderstood if it is taken as a statement about an inner state.
-
Episode 2 · 28 Jul 2025 · 4 min read
Bot, tool, agent, agentic: four terms that keep getting confused
The dividing line does not run through the technology, it runs through the assignment. A bot gets a prompt, an agent gets a goal. Everything else follows from that.
-
Episode 1 · 21 Jul 2025 · 4 min read
When the AI becomes the customer: what a certificate is still worth
An agent was supposed to pick out suitable AI training courses. It sat the exams itself and collected the certificates. The question that follows is uncomfortable and remains unanswered to this day.