The original argument
Dan McKinley's essay from 2015 argued for using established technology with known failure modes over exciting new tech. A "boring" stack lets you spend innovation tokens on the product.
Why 2025 is different
The AI tooling landscape is changing fast enough that choosing the wrong infrastructure layer locks you into something that may be irrelevant in 12 months. The cost of wrong bets is higher.
What boring looks like now
- Postgres (not a new vector DB unless you need scale that Postgres cannot handle)
- Redis (not a new cache layer)
- Node or Python (not a new runtime unless Bun's wins are material for your use case)
- S3-compatible object storage (everywhere now, vendor-neutral)
The exception
The AI inference layer is moving so fast that boring does not apply. Use the best model available now. Abstract it behind one interface so you can swap.
Takeaway
Be boring at the infrastructure layer. Be aggressive at the product and model layer. The combination is defensible.