Genesis AI is a $0-required-spend, from-scratch AI research project building an independent model family and optimizing capability per unit of compute.
Research loop: start tiny → measure → screen on free CPU → use free accelerators only for winners → promote only after frozen evaluation gates.
| Item | Current result |
|---|---|
| Promoted research model | genesis-micro-2m-v1 |
| Parameters | 1,895,808 |
| Demonstrated domain | restricted integer-expression synthesis, difficulty 1 |
| Frozen exact accuracy | 95.00% |
| Termination rate | 100.00% |
| General M3 validation loss | 3.355112 |
| Required cash compute for promoted run | $0.00 |
This is a research checkpoint, not a general-purpose assistant. The demonstrated domain result does not imply broad coding, factual, agentic, or frontier capability.
→ Benchmark & progress dashboard
→ Public roadmap
→ Current model card
Genesis deliberately separates cheap screening from scarce accelerator time.
- GitHub Actions CPU farm — 19 bounded architecture/tokenizer/filter/eval/verifier/optimizer/tiny-model screens, up to 12 in parallel.
- Kaggle GPU — CPU-screen-gated cloud dispatch to the free Tesla P100 pool. Kaggle currently documents a 30-hour weekly GPU quota or sometimes higher depending on demand/resources.
- Google Colab — optional GPU capacity when available; never a required dependency because quotas/hardware/runtime limits fluctuate.
No accelerator output can promote itself. It must pass the normal frozen evaluation, contamination, regression, and promotion gates.
→ Free accelerator guide
→ Open the Colab runner
Outside contributors are welcome. The fastest entry points are deliberately small and reproducible:
A useful contribution can be a benchmark, verifier, data-quality improvement, tiny CPU experiment, reproducibility fix, documentation improvement, or efficiency idea. You do not need a GPU to contribute.
- $0 required spend. Paid compute, paid APIs, and paid services are not project requirements.
- Independent weights. Models start from random initialization.
- No proprietary-model dependency. Training/evaluation/production must not require ChatGPT, Claude, or another proprietary model.
- Evidence before scale. Larger experiments happen only after smaller experiments justify them.
- Everything is tracked. Every task, bug, decision, experiment, documentation change, and enhancement belongs to a GitHub Issue and exactly one Milestone.
- No hidden capability claims. Public claims must match committed evaluation evidence.
./scripts/test_local.shTo refresh the evidence-backed dashboard after changing model metrics:
python3 scripts/render_progress.pyGitHub Milestones + Issues are the source of truth. No meaningful project work should exist only in chat, commits, pull requests, or documentation.
See PROJECT_RULES.md.