From 2026-03-18 to 2026-05-01, this fifty-day record documents a live applied AI research campaign across three technical tracks: novel architecture exploration, competitive neural submissions, and custom-kernel / FA3 performance work.
During this period, the work produced both original experimental architecture ideas and competitive neural submissions. The neural track reached the top of the competition multiple times during the campaign, while the broader research process earned $1,000 in research funding in recognition of performance and daily work quality.
Some ideas and approaches were quickly mirrored by other competitors, which was useful external validation that the work was operating in the right part of the search space. In the end, my best submission was a hybrid variant where my main contribution was a larger vocabulary. My goal was knowledge and progress through daily work.
The work was organized into three categories:
Novel architecture experiments intended to explore concepts, generate insight, or test unusual ideas. These included recursive layered transformer concepts, wildcards, megakernel-style ideas, and other theoretical or unconventional designs. Not every experiment was expected to match SOTA; the goal was to test whether the idea exposed something useful. I used this lane to explore concepts relating to my own research and exploration, doing so through rapid testing of new ideas.
Competitive architectures aimed at the highest level of the contest using bpb / score-based evaluation. This track performed strongly, including several jumps into first place during a fast-moving leaderboard environment where competitors were constantly leapfrogging each other.
Custom-kernel and FA3-related performance work focused on GPU behavior, hardware-specific results, performance constraints, and novel interactions.
The record includes accessible run files and test results showing what was tried, what failed, what was preserved, and why.
A twenty-year background in 2D/3D game development informs the work: systems thinking, optimization, geometry, simulations, visual reasoning, tooling, debugging, and building complex interactive systems under constraints. As AI systems become more capable and more multimodal, that background becomes increasingly relevant.