groundtruth¶
90 days to top-1% AI user & builder. No pedagogy, no certificates — a ship ladder: one real, public, measured project per week. The portfolio is the credential.
Groundtruth is a build-first program modeled on Plaintext: original bridge prose over a curated path of the best free resources, with a real, deployed, measured project shipped every week — anchored in real artifacts (Kaggle, Hugging Face, arXiv), every one completable at near-zero cost.
Start here¶
- The 90-Day Plan — the thesis, the rules of the game, the definition of done.
- The program — the nine briefs: an arm-yourself week, seven weekly ships, one four-week big swing. Each has a mental model, a path (first hour → Mon–Sun build order), a spec, an eval bar, and what gets published.
- Toolkit — the Python you're assumed to use, the language stance, and what's deliberately excluded.
- JIT map — pull-when-stuck resources. No study weeks; hit the wall first, then read.
- Build log — the public record, one entry every Sunday.
The ladder¶
| When | Ship |
|---|---|
| Days 1–7 | Arm yourself — AI-native workflow, first MCP server, eval template |
| Week 2 | LLM feature — structured outputs + golden-set eval, deployed |
| Week 3 | RAG — retrieval with before/after numbers |
| Week 4 | Agent — the loop, your tools, a completion rate |
| Week 5 | Browser agent — operate human software, measured |
| Week 6 | Fine-tune — beat the base model on your own eval, publish to HF |
| Week 7 | Multimodal — voice or vision, with latency numbers |
| Week 8 | Eval gauntlet — evals + tracing across every ship; the flagship write-up |
| Weeks 9–12 | Big swing — one product, real users, plus the visibility loop |
The rules, in one line¶
Ship every week · everything public · evals on everything · secure what you ship · AI writes, you review every line · JIT learning only · log every Sunday · stay at the frontier.