Using a language-model API and building a language model answer different questions.
An API can help me add a capability to a product. It does not make me decide how text becomes tokens, how examples move through a data pipeline, what training state must be saved, or what inference needs in order to load a checkpoint and produce a result.
That is why I am building Salty Steak.
The project is a connected system
The model is only one part of the work. My experiment includes a tokenizer, data preparation, training, checkpoint management, inference, and desktop software around the model.
Each boundary creates a question I cannot avoid by reading a summary:
- Does the tokenizer represent the data the way I think it does?
- Can the training process save enough state to continue honestly?
- Does inference load the intended checkpoint?
- Can the surrounding software explain what is happening without pretending the model is more capable than it is?
Implementation turns those questions into things I can inspect.
“From scratch” is not a marketing claim
I use the phrase to distinguish this work from calling a hosted model through an API. It does not mean I rewrote every numerical operation without libraries, and it does not mean the project contains new AI research.
The project is experimental. I am not calling it state-of-the-art, production-ready, or a ChatGPT competitor. I also do not publish a parameter count, benchmark, or release date without evidence.
Those limits make the build log more useful. I can focus on what I actually learned and what the system actually does.
The part I value most
When a difficult system stops being one large mysterious object, I can reason about the parts. A tokenizer can be tested separately from training. A checkpoint can be treated as a specific saved artifact. The desktop software can show accurate state instead of guessing what happened in a worker process.
That way of learning carries into ordinary software work. Break the claim into observable pieces, build the complete path, and verify the result that matters.