Demo-scene model training shared

A builder trained a model to continue Dwitter demos, tiny JavaScript programs limited to 140 characters. The project used 25,000 Dweets for training and a curriculum that gradually gave more weight to the examples with higher “awesomeness” scores. In the builder’s tests, the model got better at predicting code it hadn’t seen during that curriculum training, too. The original write-up describes how the builder compared generated continuations by rendering them side by side, then branching from the ones that looked promising.

This isn’t a general-purpose tool for writing company software. It’s a small, specific experiment in using a model to explore creative options, with a person choosing what to keep. That pattern could be useful for businesses testing unusual content ideas: generate several directions, compare them, and build on the ones that fit. The same principle can apply to working with AI in a company, where someone needs a way to see what’s being tried and decide what happens next. That’s the kind of workflow we build into an AI command centre.