compute.cx guides

Let’s use compute

Short guides that train real models on a fresh cloud GPU. One command. Then the machine is gone.

• 2026-09-24Qwen learns PokémonQwen3-0.6B goes from 23 to 179 wins out of 200 Showdown battles by copying a scripted expert. The simulator installs inside the job in 13 seconds. $2.07 all-in.
• 2026-09-21Two fine-tunes, one modelAsked for an order, a merged model answered with a support queue. Changing the weight mix kept both behaviors. A $0.42 experiment.
• 2026-09-21A chess model that never saw the rules29.5M parameters, 16M positions, 38.6% next-move accuracy. It draws random movers by repeating positions and loses to Stockfish skill 1. $2.14 all-in.
• 2026-09-17Can a small model learn to play Doom and Pong?A Jev-inspired action model copies the training answers and still loses. Two simple controllers do better. $1.62 in GPU runs, with decisions visible in the replay.
• 2026-09-14A $1.34 language-model compressor1.82 bits/byte on held-out enwik8. gzip needs 2.87. xz needs 2.14. A 3.4M character GPT plus rANS, $1.34 all-in.
• 2026-09-12A $0.38 AI-text detector99.4% on ChatGPT answers, 48% when the generator changes to GPT-4 and Llama. DistilBERT plus a TF-IDF baseline, $0.38 all-in.
• 2026-09-10A $0.08 prompt routerKeep 95% of GPT-4 quality, send 68% of traffic to Mixtral, cut list-price cost 72%. DistilBERT plus a TF-IDF baseline, $0.08 all-in.
• 2026-09-08Teaching Qwen to build tiny rooms192 Blender programs, a 9B model, and a $7.82 fine-tuning pilot. Real rooms and the geometry mistakes that remain.
• 2026-09-04 Face or Not Train a face classifier on Compute: two models, $0.25 in GPU runs, and a bug we found along the way.
• 2026-09-02 MNIST Digits Train a tiny CNN that reads handwritten digits, for under $1.
• 2026-08-31 Not Hotdog Train a tiny CNN that tells hot dogs from everything else.