Intelligence.Under themicroscope.
Real models. Real hardware. Repeated measurements.
Explore the measured ranges, read the test materials
and inspect what the models actually built.
01 / THE EXPERIMENTS
Less hype.
More evidence.
Same task. Different approach.
Explore exactly what the models produced.
Each episode includes measurements and the models’ actual outputs. No touch-ups.
Open an episode to explore the evidence.
02 / THE PROCESS
No black box.
Just the whole picture.
A benchmark is only useful if you can check it.
Here’s what goes into every experiment.
One variable at a time
Same prompt, same rig, one variable. Model reloaded before every run. Published cohorts: n=3 / n=5 per variant. See each episode for cohort details.
Metrics from the run
Repeated runs can vary. Compare ranges, not single runs. Inspect metrics.json beside each output.
Real hardware
Ryzen 5 PRO 3600, 32 GB DDR4 ECC and an RTX 3090 with 24 GB VRAM. Local models are served through LM Studio.
The exact prompt
The full input, byte for byte. Read the task ↗ or open prompt.txt next to each run to reproduce the test.
BEHIND THE LAB / CURIOSITY IS THE STARTING POINT.
Don't take our word.
Take a look.
Mateusz Piekut (PC Magik).
Computer service and homelab operator from Poland.