From the Engine Room
Straight talk on AI, language models and requirements engineering – written by an engineer from the automotive industry.
Are You Still Hanging With the Right Friends?
Claude Opus 5.5 leads in reliability, reasoning and coding – at more than eight times GPT-6.1 Sol's cost per benchmark task. Ten models from ten providers, from the brilliant friend to the most economical one, and why intelligence isn't the only quality that matters.
Why Iterative Prompting Rarely Eliminates the Remaining Errors
When an AI-generated algorithm is almost right, asking for one more fix seems obvious – yet with every round the result often gets less precise, not better. Why that happens, and four strategies that work instead: restart, decompose, build the critical parts yourself, Best-of-N.
What a Local Language Model Can Really Do – and Where Its Limits Are
Qwen3 14B on an off-the-shelf graphics card: what a local model handles well, where it invents details or misreads intentions – and why the safeguards belong in the application, not in the model.
Source Links Matter: How the Regulation Check Makes AI Answers Traceable
Grounding instead of hallucination: the Regulation Check bases its answers hierarchically on binding regulatory text, curated context and general model knowledge – and says so plainly when a detail simply is not covered by the imported regulations.
The Dream of AI Autonomy
With Nemotron 3 Ultra, an open model finally comes noticeably close to the frontier class – and runs entirely in your own data center. What that means for data sovereignty, cost and critical reasoning tasks, and where the gap to the cloud leaders remains.