Contact
Vol. 1 · No. 16Tuesday, 13 October 2026

Notes from the bench

4b-postgres-cover

Lead story18 Sept5 min read

A 4B model beat Postgres, but what did it really cost?

Rohan Bansal's QORL showed a 4B model could make Postgres plans faster. But for engineers with slow analytics, the real question is what it means for their workloads.

Also this week

17 Sept7 min read

Proving a model swap is safe before it reaches users

model-swap-cover

16 Sept6 min read

Ubuntu 26.10's rust coreutils swap: it's time to check your scripts

ubuntu-rust-cover

14 Sept6 min read

Google's quiet /goto redirect just broke every AI agent that searched

google-goto-article-cover
real-swe-cover

14 Sept6 min read

The best coding agent scored 38.8% on real code

A new benchmark shows top coding agents score under 40% on real enterprise code. I explain why the real story isn't the low score, but the missed requirements, and what it means for your team.

deepseek-4-1-flash-cover

13 Sept4 min read

DeepSeek V4.1 Flash: Your inference bill isn't about total parameters

DeepSeek V4.1 Flash boasts 763 billion parameters, but only 8 billion are active. This model breaks the old rules for sizing AI inference, making KV cache the real cost driver.

tailwind-shopify-cover

13 Sept4 min read

Vibe coding didn't kill Tailwind, it killed its business model

Shopify acquired Tailwind Labs after AI coding tools eroded its revenue. This isn't about Tailwind's death, but how generative UI reshaped open-source funding.

From the archive