The Dispatcher Beats the Model Fan
When a weaker model delivers useful preparation work for one fifth of the price, model choice becomes an operating decision.
Margin Notes
Short observations, finds, thoughts-in-progress. No finished argument. No prompt kit. Thinking in progress.
When Flat-Rate AI Turns Into a Meter
What changes for companies when AI vendors rework the billing model in the middle of active use.
Open Standard, Open Entry Point?
MCP is being sold as a common standard for AI tooling. That is exactly why the question matters who is actually left carrying the security burden.
The Fault Is One Layer Up
When AI tools suddenly feel worse, the problem often does not sit in the model but in the work system around it.
Not Siri-ready. System-ready.
Preparing an app only for Siri is too narrow. The real task is systemic addressability.
The App with the Factory Badge
Many companies treat AI tools like harmless productivity software. Organizationally, they often behave more like external firms with factory badges.
The Boundary of Good Intentions
The real test of an ethical AI company is not its principles. It is its defaults.
Are Managed Agents Becoming a New AI Infrastructure?
Agentic workflows turn model buying into an infrastructure problem surprisingly fast. A managed layer may be forming exactly there, above inference as the underlying commodity.
From the Bottleneck to the Runtime: The Real Apple Play
Apple is not suddenly assembling one big Siri moment. Across the last two WWDCs, Apple has been laying the runtime layer that makes apps actionable for AI.
Ethical Until the Website Edge
Cloudflare's new metric pushes the ethics question in AI a step further. Not only: what does the model do. But also: what does the product give back to the sources?
The Model Company Is Not Yet a Model Organization
Big AI numbers sound like deep transformation. Often they only show that one part of the company already runs on a new rhythm while the rest still operates on the old one.
The Trust Tax
AI platforms will not just sell model capability. They will sell predictability. The real surcharge sits in billing, limits, appeals, and policy enforcement.
The Quiet Unbundling of Teamwork
AI often makes solo work faster. But what disappears when teamwork gets translated into one person plus an agent stack shows up on no productivity slide.
From Risk to Product
The same capability was a reason for restraint yesterday and is suddenly a product today. Not because the model changed, but because the narrative and access channel did.
Touched a thousand times. A thousand times, nothing happened...
Five million automated test runs hit a single line of FFmpeg code without finding the bug. A new model spotted it on the first pass. What does "tested" even mean anymore?
The Brand in the Prompt
The agentic commerce literature says: when agents do the buying, only data matters. Wrong. Brands are compressed value profiles – exactly what an agent needs.
From Lock to Meter
Anthropic burned through three enforcement strategies in four months before landing on the only one that works: the meter, not the lock.
The Bottleneck
Apple opens MCP but controls the gateway. Opening the protocol while controlling the gateway gets you cooperation instead of resistance. Plus: The Gemini asymmetry nobody is working through.
The Market Will Sort It Out
AI will write good code because it's cheaper? The reasoning has 50 years of counter-evidence. The market doesn't evaluate quality, it evaluates delivery speed.
The Two Versions
Anthropic accidentally published the source code of Claude Code. More interesting than the blunder: internal users get a different system prompt than paying customers.
Design Gets a Text Field
Google rebuilt Stitch into a full design tool. Design is converging on text. But how much structure does that text need before the agent stops getting "creative"?
Merchants of Complexity
Agents make us faster, not more productive. When the human bottleneck disappears, so does the protection against uncontrolled complexity.
The 99-Dollar Wall
Apple is blocking vibe-coded apps. But if the highest barriers to entry in the software ecosystem no longer filter, what does?
The Invisible Craftsman
A top-3% engineer at Uber: zero GitHub commits, no social media, no LinkedIn. When AI makes visible artifact production cheaper, how do you recognize the people whose value lies in what you can't see?
Die Maschine klickt
Anthropic gibt Claude die Kontrolle über Maus und Tastatur. Was sich ändert, wenn AI nicht mehr Texte produziert, sondern Handlungen ausführt – und warum der Permission-Dialog die Governance-Frage nicht beantwortet.
Plausibel klingender Unsinn
Drei gezackte Oberflächen übereinander: Die Fähigkeiten der Modelle, die Evaluationsmethoden und die Intuition der Nutzer sind alle unzuverlässig. Warum der Aufwand für Evaluation steigt, wenn die Modelle besser werden.