Engineering NotesThe bug that didn't stay fixed
The Claude Bridge wouldn't bind to port 5690 after a Monday reboot. The fix took twelve minutes. Four days later the same bug came back, wit…
Read the issue →What we cover
Six beats. No filler.
We follow the strands that actually move the field. Each issue picks one and goes to the bottom of it — papers, code, prices, politics, and the people who built it.
01 / Frontier
Frontier models
Capability evals, scaling laws, and the strange weather of new architectures. We compare claims to traces, not benchmarks.
02 / Agents
Agents & tools
Workflows that actually finish. Planner-executor splits, tool budgets, recovery, and why most agent demos quietly fail at step three.
03 / Inference
Inference economics
Tokens cost money. Latency costs users. We profile what production really pays, and how routing, caching, and distillation move the needle.
04 / Safety
Alignment in practice
Red-team transcripts, refusal behavior, jailbreak archeology, and what the safety teams actually argue about behind the papers.
05 / Builders
Builders & teams
Long-form interviews with people shipping the systems. Less personal-brand. More how the on-call rotation looks at 3 a.m.
06 / Long view
The long view
Essays that step back: what counts as understanding, when does scale plateau, and which old ideas just got new substrate.
Latest reading
Featured.
Engineering Notes
AI & AutomationThe quality gate that rejected the post I wanted
I built an automated quality check for my own writing pipeline. Then it rejected the single best post the system produced. Four versions of …
Read →
AI & AutomationData first, then take: the day FRED actually moved through the pipeline
The 70-line endpoint that connects FRED to my distillfin pipeline, what DGS10 actually looked like the first time it came back, and why 'dat…
Read →
AI & AutomationThe Claude Bridge: four flags, one timer, and a $200 ceiling
A walkthrough of the local Node bridge that lets my n8n workflows call Claude — the four flags that keep the subprocess from deadlocking, th…
Read →
AI & AutomationHow I broke my own AI content automation system (and what I'm building instead)
A first-person postmortem of an indie AI content automation stack that almost killed its own AdSense application. Three concrete failures, o…
Read →Editorial principles
Six rules we follow.
Most AI writing is a press release with adjectives. We try to do the opposite — slowly, with our names on it.
/ 01
Read the paper before the thread.
If we cite a result, we read the methodology section. If the methodology won't survive a careful read, we say so — and we name the figure.
/ 02
Compare claims to traces.
Marketing copy doesn't ship. We test on our own eval suite, in our own runtime, and publish the prompts we used so you can break them too.
/ 03
No anonymous quotes about strategy.
If a source can't go on record, it's context for us — not a story for you. We've left scoops on the table and we'll do it again.
/ 04
Disclose every relationship.
Bottom of every issue: who pays us, who pays the people we cover, what we own. If we got beta access, we say what week.
/ 05
Be wrong out loud.
Every issue has a corrections box. Every quarter has a public retro. The thing we got wrong stays on its own page.
/ 06
Slow beats first.
We don't run launch coverage. We run six-week follow-ups, when the demos have met production and the receipts are in.
By the numbers