MAPLE
Gemma 4 26B cut to 48 experts per layer so it fits a 6 GB laptop GPU
Emerged finds new products worth a look, daily — what’s new today
It shows, with timings and byte-identical parity checks, how much of a small GPT an array language with built-in autograd can absorb, down to 32 lines.
A literate comparison of Karpathy's 149-line microgpt.py against a Rust port and three versions in sw-MLPL, an array language with autograd. Each port is an Org document whose output is real and checked against a baseline; one replays the Rust random stream and matches its output byte for byte. On an M1 Max the idiomatic 48-line MLPL version trains faster than the compiled Rust. MIT licensed, from Software Wrighter LLC.
More like this
Warp
Run trillion-parameter models on a laptop by streaming weights from NVMe
ai-infra-book
An open textbook that sizes LLM inference and training from first principles
Jik
A statically typed language with regions instead of a collector
jdBasic
A modern BASIC where you pause a running program, change the code and carry on
InstinctFlash
A serving runtime that speeds robot policy models up for real control loops