Measure, publish, say what it cannot show.
The methods, the code and the raw measurements are public. Anyone can rerun a number on this site and get the same answer, or tell us why not.
We repeat every run, validate models on data they were not fitted to, and treat an impossible coefficient as a broken model rather than a discovery. When we are wrong, the note says so and keeps the mistake, because the mistake is usually the most useful part.
The company is the work around the research: making a model run well on a particular machine, and measuring it honestly when it does.
Inference engineering
Memory bandwidth, kernels, quantization, batching, KV cache, and how inference clusters are put together: interconnect, power and cooling per rack, serving at scale.
Data science
Experimental design for noisy hardware: repeats, order effects, held-out validation, and physical models fitted only where the data can identify them.
Software engineering
Small, readable, open tools. joule is the first; each study adds the scripts that produced it.
Later: electrical engineering. An open power-measurement board, current sensing at about 1 kHz, to replace the 1 Hz board reading. Not started.
Have a model and a machine?
If you build hardware or ship models to devices and want a model made to run well on yours, and measured honestly, open an issue on joule or reach us through GitHub.
Open an issue