Research / Hardware
The hardware behind the experiments
Every number we publish was produced on an ordinary consumer laptop, not a workstation and not a cluster. That is deliberate: if the aim is AI that runs on hardware people already own, the experiments should too.
The test machine, roughly
We leave out the exact model on purpose. It is not the point, and a rough description is enough to judge the scale: the machine is roughly what a student or a hobbyist might own, and the prototype's whole teach-from-nothing session fits in about 12 MB of memory.
What we measure, and what we do not
| Quantity | How | Status |
|---|---|---|
| Wall time | One CPU core, release build | measured |
| Peak memory | Resident set of the whole process | measured |
| Operations and words moved | Counted by a ledger inside the code | counted |
| Energy | Counted operations priced with published 45 nm figures | estimated |
| Energy from the CPU's own counter | Intel RAPL energy counter, now readable on the machine | to be wired in |
| Heat and sustained battery draw | Not measured; matters for phones | later |
Caveat
Because timings are on a single laptop core and energy is estimated, we trust ratios (this design against that design) more than absolute values. When something is timed we do not run heavy jobs alongside it.
A working budget
To give the work a concrete target we use a rough envelope, every number of which is negotiable: about 1 GB of resident RAM and a few GB of storage for a learned system on an 8 to 12 GB phone or laptop. Bigger stores can sit on disk, because disk is read sparsely. A smaller tier (a 2 to 4 GB phone) is a stretch goal. Budgets like these are stated next to each result, so results can be judged against what a device could actually spare.
"Not overwhelming the device" is measured, not felt: resident and peak RAM, storage footprint, memory bandwidth per answer and per lesson, energy and heat, and background behaviour. Learning should be scheduled for idle time so the machine stays usable while Transient runs.
What we plan to add
- A rented GPU, for the controlled comparison against a small transformer. Only code and synthetic data would be uploaded; nothing private. The comparison model is to be written in Rust so operations and joules are counted the same way on both sides.
- Real energy readings from the CPU counter, folded into the ledger so estimated and measured joules can be compared.
- A phone as a second target, once there is something worth timing there.
If you replicate any result on different hardware, we would like to hear what you got. See how to contribute.