Notes / September 2026
When a simpler method was enough
A transformer where logistic regression did the job, and a pipeline that Excel already covered.
2 minute read.
The transformer
I once presented a complicated transformer architecture. My instinct was that a more sophisticated model had to give us something better.
- What I tried
- Logistic regression, on the same task.
- What surprised me
- It did the job. And it was easier to understand, and to explain to the people who would use it.
- What I did next
- We went with logistic regression. The model had to be deployed and used, not only shown in a conference paper.
The spreadsheet
Another time, I was explaining a pipeline to a team. Convert the data to Parquet, then process it through several stages. The team lead stopped me. Excel already did what the team needed.
When it was needed
Other times the complexity was needed. In my study on disability progression one shared model was not enough, because patients differ from country to country.