What this funds (honest numbers from today's preflight)
STATUS.md is the live blocked-vs-doable document. Phase 4 compute is the structural blocker. Concrete asks:
- 10M validation run — ~$30-50. A 10M-param intermediate trained on the full BPE-tokenized corpus, ~30-40 hours on an A100 spot. Validates whether our capacity-ceiling hypothesis is right (i.e. that A1/V7/A3-sharpness all become solvable at scale) before we commit a bigger run.
- 50M production run — ~$500-1,000. The actual Phase 4 base. ~830 hours of A100 time. If the 10M validation passes, this is the train that gives us an Eli that speaks coherently AND retains the values architecture we've already validated at 1.8M.
- Floor on the engineer's tooling — Claude Code at $200/month. Recurring donations hitting that level let the project pay for itself.
- Phase 6 multimodal compute. Vision + voice teacher models run on Groq free tier; the small runtime models need GPU hours that have to come from somewhere.
The headline architectural research at 1.8M is done. What we can't currently afford is the test of whether it scales. If you're in a position to fund any of the above — even the $30-50 10M validation — open an issue on the GitHub repo and we'll wire it directly to the training run.
Phases are listed publicly in docs/roadmap_to_perfect_interface.md with effort estimates and decision gates. Every donation level has a phase it unblocks; the math is in the open.
How to donate
Stripe checkout (PLACEHOLDER LINK — replace with the real one once the Payment Link is created):
Donate via Stripe →One-time or recurring, any amount. We don't see your card details — Stripe handles checkout. We receive a name + the amount + an optional message.
If donating isn't your thing
- Star the repo. Visible signal helps grant committees rate this as a serious project. github.com/lordbasilaiassistant-sudo/substrate-self
- Cite this work in your own writing if it informs your thinking. The empirical proof artifacts and the roadmap document are stable URLs.
- Forward the demo to one ML-curious friend. Distribution without peopling — the URL goes farther than any Twitter thread the author isn't going to write.
- Send the author a bug report. Found a way to break the proof, get Eli to leak information, or surface a falsifier we didn't pre-register? Open an issue on the repo.
What this doesn't fund
Listed explicitly to be the kind of project that posts these things explicitly:
- No paid hosting tiers (the demo stays free even at high traffic — GitHub Pages + Cloudflare absorbs it).
- No social-media spend. The author can't peoplee. There is no team running Twitter / Discord / TikTok to fund.
- No staffing. One engineer + autonomous-after-deploy AI agents. Donations stay 1:1 with compute + the author's tooling floor.
- No equity to sell. There is none. This is open-source MIT.
Grants and structured funding
The project is actively applying to ecosystem grant programs (Anthropic Researcher Access, Mozilla MIECO, Cohere For AI, EleutherAI partner program). If you represent a grant-giving org and want to discuss the work, the contact path is opening an issue on the GitHub repo or DMing @THRYXAGI with a one-line message and an email to reply to (the project doesn't run DMs ongoing).