Writing

My agent gained root on a spare JioFiber JCOW411, spent hours in the weeds of closed Broadcom drivers, and concluded a wireless repeater was 'impossible on this hardware'. It wasn't impossible — here is how Broadcom WET mode and a misread error code cracked it wide open.

A text-only Qwen3.5-2B, 4-bit LoRA and a nine-quantile regression head took Amazon product pricing from a 68.83 SMAPE baseline to 44.11 in two laptop-sized epochs.

Phase 1 of miniLLM-214M went perfectly on Kaggle's free T4s. Phase 2 failed for reasons that had nothing to do with deep learning — and the model driving my coding agent is the one that failed.

113M parameters, 1.4B tokens of 2024-or-newer code, 4.6 hours on four L4s — and two headline results that shrank when I measured them properly.

154 commits, 15 new repos, 3 models trained from scratch, and one number I'm not proud of.

One model reads move text and builds the board itself. The other gets the board handed to it. Same domain, opposite bets — and the numbers say something uncomfortable.

Character-level SAN, no board in the input, and a linear probe that reads the position out of the activations at 85.6%.

12 hours, 8GB of VRAM, 1.2B tokens of FineWeb-Edu, and a measured 2.7x win from swapping the optimizer.

First post — how writing works on this site.