The Bottleneck Was Never the Code
A one-page spec produced working software in an hour. Everything between that and a product people could depend on was not in the page, because nobody had decided it yet.
Read this issue →The MG Signal
Real AI. No shortcuts. Every two weeks.
One real story from the field. One tool worth knowing. One truth the AI world gets wrong. 400 words, no hype, every two weeks.
Unsubscribe any time.
Past issues
Old Mathematics Quietly Powers Modern AI
Eigenvalues shaped search in 1990 and ranked the web in 1998. They still sit under embeddings, PCA and LoRA. The small end of the spectrum may have one more use.
Read this issue →Everyone Predicts Failure. Few Predict Time.
Failures were rare, so most of our data wasn't failure data. Survival analysis let every observation contribute what it actually knew, instead of forcing a label.
Read this issue →Why We Stopped Using Western Electric Rules
The zone rules assume a distribution most real processes don't have. We kept Rule 1, made the limits adaptive, and handled drift as a separate problem.
Read this issue →Analytics Didn't Go Away. It Got More to Work With.
Traditional Analytics and ML aren't obsolete. GenAI turns years of unreadable text into data they can use, then explains the results back to people.
Read this issue →The Model Is Only 10% of Your AI Agent
In any agentic system, the model is about 10%. The other 90% is data engineering. An anatomy of one agent turn—and why the leverage was never in the model.
Read this issue →