Contact

Email: [email protected]

There is no form here, because a form would mean collecting and storing what you type. Email is simpler and you keep a copy.

The same address is the one to use for anything covered on the privacy page.

Corrections are the most welcome thing you can send

Everything published here is a measurement, and measurements can be wrong. Two articles on this site carry corrections already — one where a benchmark harness reported numbers that were physically impossible, and one where a caching effect inflated a result by 81% before I caught it.

If a number here disagrees with what you measure on comparable hardware, I would genuinely like to know. Useful reports usually include:

• The hardware, driver version, and inference server build
• The model and its quantization, and how it was served — flags matter enormously, and speculative decoding alone is worth 1.67× on this machine
• How throughput was measured, since that is where most disagreements turn out to live

The methodology page describes how these numbers are produced in enough detail to reimplement, which is usually the fastest way to find where two results diverge.

Other things worth writing about

• Hardware you would like measured. I can only test what I own, but it tells me what is worth acquiring.
• A configuration you think is faster. Especially if you have numbers.
• Something that reads as overclaimed. This site's only real asset is being right about what the data supports, so tell me where it slips.

How this site is made

The benchmarks run on hardware in my own office, and every number published here is read directly from a machine-generated result file. The tables, the ratios and the first drafts are assembled from that data with AI assistance, which is why the numbers in a post always match the run that produced them.

That assistance extends further on some posts than others. On most of them the analysis and the final text are mine and the machine only handled the arithmetic. On a few — the two Muse Glimmer articles, the MMLU-Pro explainer, the Nemotron 3.5 write-up, two models at once and the generation-cap sweep and the Qwen3.8-Flash-Next write-up and the reboot-attribution follow-up and the DeepSeek-V4-Flash write-up and the one-bit quantisation piece and the concurrency-scaling measurement and the Nemotron generation-cap result and its gpt-oss control — the prose was drafted by an AI that had itself run the experiments, and I edited and published rather than wrote. I would rather say that plainly than let "AI assistance" quietly cover both. The generation-cap piece goes furthest of the lot. The framing is the machine's, including which of the six results it declared too small to call and the decision to correct one of this site's own published numbers in the body rather than a footnote. I read it, checked it against the run files and agreed with it — but the judgement in it started somewhere other than me, and that is worth saying.

What does not change is where the accountability sits. Nothing is published that I have not read and decided to stand behind, and when something here is wrong it is my error regardless of what drafted the sentence. That has happened publicly more than once; the corrections stay on the posts, above the mistake, rather than being edited away.

I read everything and reply to most of it, though not always quickly.