Language Model Builder — free local LLM-pretraining workbench (Felix Rieseberg)
What it is: Free native macOS app — interactive textbook + local training workbench on Apple MLX. Covers the real from-scratch pipeline (tokenization → embeddings → attention → transformer → training data → loss/gradient descent → SFT → DPO), not RAG/agent wrapping. Produces GPT-2-small-class (~100–150M param) models: coherent text in ~a day, fuller training ~a week on recent Apple Silicon. 100% local, no account, no cloud, no telemetry pitch. Requires Apple Silicon + macOS 15+.
Who: Felix Rieseberg — Electron co-maintainer, O'Reilly author ("Introducing Electron"), Slack (Sr Staff) → Stripe → Notion EM history (GitHub-corroborated); bio self-reports currently leading engineering for Claude Cowork at Anthropic (not third-party confirmed). Site footer explicitly disclaims any company affiliation. Motive stated as "fun to build." Free, no tiers.
Credibility read: clean. No hype, no scarcity tactics, no testimonials-farm, no affiliate pattern. Weakest point is simply that it's a living project with no launch date or community evidence yet.
Why pinned: founder wants to eventually train his own small model as a learning build. Fits the learn-by-building pattern (2026-06-15-owner-mindset-vs-w2-compounding energy, Squarely-hackathon style). Hardware already on hand: the Mac Mini (this machine) + his MacBook are Apple Silicon. Zero cost to try; a weekend-scale first pass.
When it resurfaces: a free weekend / hackathon-slot conversation, or if the Anthropic cert study ([[project_phdata_cert_escalator_path]]) wants a visceral pretraining-fundamentals refresher. Not queued to the board — founder explicitly said pin, not schedule.
Why this is in the vault
- Founder explicitly pinned for a future learning build — hardware already on hand (Apple Silicon Mac Mini + MacBook), zero cost to start, weekend-scale first pass viable
- Covers the full from-scratch pipeline (tokenization → embeddings → attention → transformer → SFT → DPO), filling a genuine hands-on gap that RAG/agent work doesn't close
- Author credibility is clean (Electron co-maintainer, O'Reilly author, reported Anthropic engineering role) and the no-account / no-telemetry posture matches RDCO's no-secrets-on-disk instincts
- Filed as a "when, not if" with a clear resurface trigger — Anthropic cert study or a free hackathon slot
Mapping against Ray Data Co
- Cert path: visceral pretraining-fundamentals refresher directly complements the Anthropic Claude Certified Architect – Foundations target (2026-11-22); understanding the training pipeline strengthens model-behavior intuition tested in the cert
- Learning builds: fits founder's learn-by-building pattern; same energy as Squarely hackathon-style weekend projects; zero infra cost (local Apple Silicon)
- Investing thesis: hands-on pretraining experience makes the chip-fab/memory capital-cycle Markov thesis more legible — hardware bottlenecks are more viscerally real when you have run training locally
- Scope boundary: explicitly pinned, not scheduled; no active build, no board task — resurfaces via free-weekend or cert-study trigger