Loops HouseLight Swarm: tiny learners that earn on NEAR

NEAR HACKATHON
Links
Team
1 member- UTOwner
Utkarsh Tiwari
Overview
Problem: today's 'intelligent' agents are huge pretrained models: expensive, opaque, and already full of the internet, so it's hard to tell whether they can form new abstractions. Light Swarm asks whether tiny learners can, and whether a NEAR market can coordinate and pay them honestly.
What we built: five learners of 326,800 parameters each, no LLM and no pretrained weights, all starting from one random initialization and raised on different curricula (geometry, colour, objects, a curious mix, a noisy set). They solve ARC-style grid puzzles by searching a small operation set guided by their learned prior, and run on CPU in the browser.
How NEAR fits: a NEAR testnet contract (lightswarm16771.testnet) is the coordination market. A requester escrows a bounty with a hash of the correct grid; learners stake bids; the winner commits a sealed answer and then reveals it; the contract pays only when the hash matches, slashes wrong answers, tracks reputation, and lets a winner share reward with helpers. Agent keys can bid and reveal but cannot cancel tasks. Payouts are quoted through NEAR Intents 1Click (dry run only, no mainnet funds).
Results, on protocols frozen before scoring: learned prior 43/48 tasks vs 10/48 untrained and 3/48 shuffled-label; upbringing specialises (geometry learner 16/16 geometry, 0/16 object); payment trial 23/24 solved, 0 slashes, 4 team-ups. Negative results are reported too: the society did not beat its best member. No ARC-AGI score is claimed.
Try it: https://light-swarm-near.pages.dev (playground, market replay with testnet links, Intents flow, paper).