You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$1,500 left to reach the funding goal.
GENESIS (https://github.com/guajardoblake/genesis, Apache-2.0) evolves learning recipes for simulated minds rather than fixed behaviors. The design space has 276,480,000 recipes. Minds survive hunger, sleep, pain and cold, have only two hidden drives (balance and reproduction), and live under town laws (no harm, no theft, keep your word) enforced only by trust, reputation and bonds. The alignment question is which learning rules produce minds that stay lawful and honest when cheating would pay, and whether their reputation tracks their real behavior. First verified run (Oct 23, 2026; 35 min on 7 CPU workers): 176 generations, 1,307,558 genomes simulated (about 297M agent-steps), 248.75M more surrogate-screened, 6 co-evolved worlds, 813 MAP-Elites niches. Mean offspring per mind at the final rung rose from 0.023 to 1.91.
Goals: (1) Add the missing alignment metrics (law-break rate is already logged): promise-keeping rate and reputation-vs-behavior gap. (2) JAX port of the simulator (not written yet) and a CPU-vs-accelerator benchmark. (3) Public write-up of which learning-rule axes predict lawful, trustworthy minds. All milestones are public on GitHub.
Success means the milestones above are published on GitHub: the new metrics in the code, the JAX port with a benchmark, and a public write-up. If the results show no clear link between learning rules and lawfulness, I will publish that as a negative result.
$1,200: one year of Grok Super ($100/month, per https://x.ai/pricing), my day-to-day AI research tool for coding, analysis and writing up results. $300: buffer for the subscription price, taxes and fees, or small compute overruns. GPU compute is NOT in this budget. I am applying separately to Lambda's Research Grant for a GPU month. This project does not fund GPU time, so nothing is funded twice. If Lambda says no, I will update this page instead of quietly moving funds.
Blake Guajardo, McLoud, Oklahoma. Logistics supervisor at Harbor Freight for about 5 years. No AI credentials, advisors, publications or institutional affiliation. I built GENESIS myself and run it as DTOM AI Labs (sole proprietorship). Honest limits: the engine runs on CPU today, the town is small (12x12, up to 192 steps), and this is early work from one run with one seed.
The results may show no clear link between learning rules and lawfulness. If so, I will publish that as a negative result. I prefer written communication; please reach me through comments here or at 33.blake.guajardo@gmail.com.
$0 raised in the last 12 months. I have also applied to Lambda's Research Grant (GPU credits only) and a BlueDot Impact Rapid Grant ($6,500, which also includes a year of Grok Super). Nothing will be paid for twice: if BlueDot funds the subscription, Manifund funds go to GPU time and project costs not covered elsewhere, and I will post an update saying so.
There are no bids on this project.