Works in the demo, costs in production
The prototype is impressive. Tokens are invisible. Errors are rare. Then production starts, and the bill changes character. Klarna learned this the hard way in 2025.
Articles
Where AI is useless, and where the market says otherwise
Often, a regex, a business rule, or a sort is enough. AI is sometimes sold as a universal answer to problems with simpler, more reliable, cheaper solutions. A tour of cases where the market sells you what you don't need.
Build / Buy / API: the trade-off nobody frames correctly
Building your own model, buying a vertical solution, or using an API: three strategies with radically different risk, cost, and sovereignty profiles. The sovereignty angle often changes the decision.
The 3 questions to ask before adding AI anywhere
Not every AI deployment decision deserves a 6-month project. Three questions filter relevant use cases from marketing use cases in under 30 minutes.
The real cost of an AI project: what the quote doesn't show
The AI integration quote only shows the surface. Underneath: token costs in production, infrastructure, data preparation, prompt maintenance, human review, and technical debt. The real TCO is often 3 to 5 times the initial budget.
The AI pitch jungle: reading it without being fooled
Every software vendor now has an 'AI' offering. Most are API wrappers with a ChatGPT logo on them, sold at custom-solution prices. A guide to reading an AI pitch without being impressed by the demo.
Your data and trade secrets in AI: what leaves, what stays
When you use an AI API, where does your data go? Is it used to train models? What does GDPR say? What the AI Act changes. The questions your DPO should ask before any deployment.
The hype cycle: the same curve every time, how not to buy at the peak
AI follows the same cycle as the dot-coms, 3D printing, blockchain, and the metaverse. Gartner's curve is predictive. Knowing where you are on it is the first skill of an executive facing an emerging technology.
Hallucination and false confidence: not a bug, a property to manage
An LLM asserts a truth and an invention with the same confidence. This isn't a flaw the next version will fix. It's a structural consequence of prediction-based operation. Here's what that means in production.
Beyond LLMs: the AI you already use without knowing it
Computer vision, recommendation systems, fraud detection, predictive maintenance: AI didn't start with ChatGPT. It has been running for years in the infrastructure you use every day.
Open source vs proprietary: control, dependency, and the real trade-off
Llama vs GPT, Mistral vs Claude, open weights vs API: the distinction isn't ideological, it's strategic. Understanding what you actually control in each case.
The 'reasoning' of models: what it does, what it doesn't
o1, o3, DeepSeek-R1, Gemini Thinking: a new generation of models sells its ability to 'think'. Benchmarks genuinely improve. But calling it reasoning is choosing a word that sells more than it describes.
Running AI on your own infrastructure: on-premise, OSS, GPU and APU
You don't need a datacenter to run a local LLM. A server with a consumer GPU or even a Mac M-series can host models from 7 to 70 billion parameters. What's possible, what isn't, and why it matters.
How an LLM works, without a single equation
Tokens, context, temperature, hallucination by design: understanding how a large language model works without a single mathematical symbol. For deciding what to trust it with and what not to.
AI wasn't born in 2022: from algorithms to machine learning
ChatGPT put AI on the map in November 2022. But the foundations are 70 years old. Understanding the genealogy means understanding why a model needs a base to learn and why the miracle didn't appear out of nowhere.
AI has 'intelligence' in its name. Mostly, it's the artificial part.
You're being sold intelligence. What you're renting is a machine that predicts the next word. This foundational misunderstanding distorts every AI decision you make.