#Google Colab

A 4 GB laptop GPU once sent every heavy job to a borrowed Colab runtime. Deciding per piece of work instead — by a measured number, by reliability, by a change of method — split one project across three places and deleted 9975 lines of machinery on the way home.

Each project began with its interface served from inside the borrowed Colab runtime, through Colab's own proxy — until Colab warned against it. Moving the interface out made the runtime a pure worker, and left two codebases held together by nothing but an agreement.

A borrowed Colab GPU ends every session by time limit or daily quota, often in the middle of hours of image generation. Picking up where it stopped meant treating each new runtime as a machine that knows nothing — and may not even be the same machine as yesterday's.

A Colab runtime has no address anyone can call, so a laptop and a borrowed GPU talk through a broker both of them dial out to. Sending commands was the easy half. Knowing — whether the worker is alive, whether a command arrived once — is where six silent failures hide.

A 4 GB GPU pushed the heavy work onto a borrowed Colab runtime, so a laptop and that runtime had to share one store. Storing the data was easy. Believing it was not: seven of the ten obstacles were successes that were false, and the most useful fix removed an entire layer.