Case Study

4 posts

Mar 12, 2026 · 10 min read tooling case-study

Teaching an LLM a Language It Has Never Seen

Fine-tuning a 4B parameter model to write valid Hew code, using the compiler as the test suite and a 780M iGPU as the training rig.

Mar 10, 2026 · 5 min read case-study

MicroGPT — From 1,092 Lines to 677

Most of the gaps microgpt found got fixed within the week. Here's the refactoring sequence — commit by commit — and how close the output matches Python.

Mar 9, 2026 · 6 min read case-study compiler

What MicroGPT Found — 13 Bugs and 340 Lines of Workarounds

The point of porting a real program isn't to have a GPT. It's to find out where your language breaks. MicroGPT found 13 compiler issues and revealed that 31% of the code was working around language gaps.

Mar 8, 2026 · 7 min read case-study

Porting Karpathy's MicroGPT to Hew

I needed a real program to stress-test the compiler. A 120-line GPT turned into 1,092 lines of Hew, and a third of those lines were working around compiler bugs.