GPT-5.6 Built a Tiny iMessage Transformer From One Prompt

Why it matters

Schirano's demo shows frontier coding models moving from app scaffolding into local ML project generation, where the bottleneck shifts from writing code to controlling private data, evaluation and memorization risk.

Pietro Schirano (@skirano), the MagicPath CEO and builder behind early AI design tools, posted a thread on X on July 9th showing GPT-5.6 Sol generating a full training pipeline for a small language model trained on his iMessage history.

https://x.com/skirano/status/2075283308276220011

The claim is narrow and important: Schirano did not say GPT-5.6 fine-tuned an existing open model. He said it wrote the code to train a decoder-only Transformer from scratch, locally on his Mac, from one prompt. The resulting model, according to Schirano, was built to generate short replies in his writing style.

Schirano's demo lands inside the kind of work he already sells through MagicPath, a design-and-code workspace where agents create editable product interfaces. The throughline is the same: frontier models are moving from chat output into project scaffolding, file operations, local tools and multi-step technical execution. Here, the artifact is not a web page or app prototype. It is a local model training system aimed at personal style imitation.

The repository Schirano linked, Doriandarko/texts-to-transformer, describes a small personal style model rather than a general assistant. The default preset is a 4-layer, 1.38 million-parameter decoder-only Transformer with a 4,096-token byte-level BPE tokenizer and a 256-token context window. The README says an example development run used roughly 8 million training tokens and produced a 1.38 million-parameter model that generated short replies in the owner's style.

That distinction matters. A 1.38 million-parameter model is tiny by modern LLM standards. It is built to learn rhythm, slang, punctuation, casing and common reply structures from one person's messages. The repository says it does not reliably learn arithmetic, factual knowledge, multi-step reasoning, long-term memory beyond the context window, or the capabilities of a pretrained assistant model. The technical point is not model quality. It is that GPT-5.6 apparently handled the scaffolding of a real local ML project: data extraction, privacy processing, tokenizer training, model selection, training, evaluation, export and terminal inference.

The privacy surface is as central as the demo. The repo instructs users to grant Full Disk Access so the pipeline can read the local Messages database, then create a read-only SQLite snapshot. It says attachments are never opened or copied, handles and chat identifiers are pseudonymized with keyed HMACs, obvious URLs, email addresses and phone-number-shaped strings are redacted by default, and raw messages are not printed in normal logs. It also warns that the resulting model may memorize private text and should remain private.

The command sequence in the README shows why Schirano's post is getting attention among developers. The pipeline snapshots ~/Library/Messages/chat.db, inspects the local Messages schema, prepares JSONL conversation sessions, trains a byte-level tokenizer, encodes the corpus, trains the model from random initialization with MLX, evaluates a held-out test split, exports local weights and then launches a terminal reply generator. The local chat command prints a suggested me: reply but does not send an iMessage.

Schirano said the whole build took about 30 minutes. That number is self-reported, and neither the X thread nor the repository proves that GPT-5.6 authored every line without human edits. The repository does, however, exist as a runnable project with installation, training and evaluation steps, and it matches the model dimensions Schirano posted in the thread.

The timing is useful for OpenAI. OpenAI began a limited preview of GPT-5.6 Sol on June 26th, describing Sol as its flagship model in a family that also includes Terra and Luna. The company said the preview would first be available through the API and Codex to select trusted partners and organizations, with broader access planned later. An OpenAI Help Center page says GPT-5.6 is available during preview through the API and Codex to a limited group, and is not available in ChatGPT during that preview.

OpenAI's own positioning for GPT-5.6 emphasizes software engineering, computer use, professional work, scientific research and cybersecurity. Schirano's example sits in the software-engineering lane, but with a more operator-relevant twist: the model is not just answering coding questions. It is producing a working system that handles local data, trains a model and exposes the result through a CLI.

That is why the demo is more meaningful than another benchmark screenshot. A single successful personal project does not establish that GPT-5.6 can safely or reliably build production ML systems. It does show what early access users are probing first: whether the newest frontier model can compress an ML engineer's project plan into an executable local workflow, with enough guardrails that a motivated user can run it without turning private messages into a cloud dataset.

The unanswered question is how reproducible the result is for ordinary developers once GPT-5.6 access broadens. Schirano's iMessage corpus, Mac hardware, prompt, edits and tolerance for failure all shape the outcome. The repository itself frames low-data training as memorization-prone, which is the honest caveat. GPT-5.6 can apparently make the pipeline feel approachable. The hard part is making the output safe, private and repeatable when the training data is a person's message history.