Looking for AI in Linux kernel growth
I expected AI coding tools to leave a visible bend in the curve. Across 113 Linux releases, I could not see one.
Field notes
Systems software, measured.
Data layouts, benchmarks, SIMD, market infrastructure, and what the machine was actually doing while everyone was looking at the API.
Writing
Latest posts
What if the prefetcher is wrong for my code?
Machine defaults are someone else's guess. I put Intel's hardware prefetchers on a per-core switch in /proc so I can benchmark goblin-store, goblin-core, and even the Python interpreter against a CPU that stops guessing.
Freedom at 9,600 Bits per Second
Inkline, a satellite phone, and the freedom to work beyond the café's Wi-Fi. My data plan may be unlimited. My battery is not.
GoblinStore versus S3 Express, one object at a time
A Dell R820, spinning disks, Graviton4 clients, and a reproducible comparison across the same 590 objects.
I don't maintain software anymore.
Maintain the specification. The implementation is memoized build output, and replacing it gets cheaper by the day.
DESQview and WordPerfect, back in the terminal
Introducing GoblinView and Goblin Purrfect, with a live equation editor over an Emacs PDF and properly clipped Kitty graphics.
Thirty years was enough
From WordPerfect’s equation editor to a PDF beside LaTeX in terminal Emacs.
What If the Terminal Had Won?
Goblin Mosh, GoblinView, and Goblin Purrfect are useful terminal tools—and an experiment in a world where Windows never became dominant.
I Built a Zset That Holds 80.8 Million IDs in 2.10 GiB
Goblin Core's packed INT32/FLOAT32 zset used 53.6–72.2% less RSS than Redis, Valkey, and Dragonfly in the Wikipedia replay. Faster completion was a bonus.
Faster isn't better
Faster network paths reduced packet latency but worsened OpenNTPD's steady-state jitter; userspace UDP produced 10.8 times the 10 GbE baseline in this run.
The GPU Helps Before the Queue Gets Long
An AMD Vulkan path cuts progressive JPEGli latency and energy while an on-demand server has headroom; the advantage fades as the machine fills.
Faster JPEGli Decode with Lower Attributed Energy on M4 Max
For baseline images at or above 1 MP, direct Metal ending in a GPU-resident texture was 12.5% faster than TurboJPEG and used 53.2% less process-attributed energy; combined CPU+GPU rail energy was 3.9% lower.
The Bootstrap Listener That Ruined My Tail Latency
A harmless-looking nonblocking accept() loop turned a low-microsecond libfabric path into millisecond-scale tail stalls.
What a Noisy Benchmark Machine Almost Cost Me
A negative compiler result looked conclusive until quieter machines exposed what benchmark noise had hidden.
AI Is Random, Not Wrong
Why nondeterministic AI output is not itself evidence of error, and how seeded randomness already powers trusted systems from hash tables to Ethernet.
Two Poops and a Lollipop
Learning to market serious open-source systems software by telling the human stories between the benchmark wins.
The COW Goes MOO
How Goblin Core keeps copy-on-write memory growth during BGSAVE from becoming an out-of-memory failure.
Why We Should Learn to Love the Robot
Why AI-assisted software is still authored by the person making the architectural decisions and checking what must be true.