An ex-Anthropic guy vibe coded a LLM from one prompt only

Signals Inbox·July 10, 2026·AI DevTools

Former Anthropic builder Pietro Schirano says GPT-5.6 turned one prompt into a complete training pipeline, then trained a small language model locally on his iMessage history.

Q1Why is one prompt such a big deal?

Training even a small model normally involves several separate jobs: cleaning the data, converting it into the right format, choosing a model architecture, writing training code, adjusting settings, running tests, and fixing failures. Schirano says GPT-5.6 handled that whole chain from one instruction. That pushes AI coding beyond making websites and apps into automating basic machine-learning engineering.

Q2Has nobody done prompt-based model training before?

The broader idea already existed. New vibe-tuning tools can generate datasets, configure computing resources, fine-tune small models, and evaluate them from natural-language instructions. Developers have also used coding agents to assemble training scripts for years. What makes this demo interesting is the compression: one general-purpose model apparently created the pipeline and completed the training locally without a specialized platform.

Q3Why does running it locally matter?

His training data was private iMessage history. Sending years of personal conversations to an external service would create an obvious privacy problem. Local training keeps the raw messages on the computer and avoids renting expensive cloud GPUs.

Q4What is the bigger change here?

Vibe coding began with people describing an app and letting AI write the code. The next step is describing the AI capability you want and letting another model build the data pipeline, train the model, and package the result. That could turn custom models from specialist projects into ordinary software components. Instead of calling one giant model for every task, companies could quickly create smaller models for one customer, workflow, writing style, or dataset.