Thoughts on data engineering, state of AI, Microsoft Fabric & Power BI.
Why DAXBench needed a v2, and how it works: 94 tasks across five task types, 77 deployed semantic models, a gold gate that tests every task before any model sees it, execution-based probes, and all-or-nothing scoring. Built with SQLBI.
A LoRA fine-tune of GPT-OSS-20B improved DAXBench score from 15/30 to 20/30, with 20% more correct results on an independent evaluation set and dramatically reduced response length and latency.
AI can speed up Power BI development, but DAX is still a sharp edge. Benchmarking 70 models shows the biggest failure pattern is not weak business logic, but DAX grammar.
Vibe coding removes the friction that used to force natural pauses. The result is a new kind of addiction - and a cognitive overload nobody's talking about.
A systematic benchmark of large language models on Power BI formula generation reveals which AI assistants are worth using—and why most of them fail on the same patterns.