AVA
MEASUREDA 2B model fine-tuned to beat Llama 3.2 3B on ARC — with a 42 MB adapter on a 4 GB GPU.
// WORK
A 2B model fine-tuned to beat Llama 3.2 3B on ARC — with a 42 MB adapter on a 4 GB GPU.
Eleven AI agents deliberate every prediction-market trade; refusals are recorded on-chain.
Architecture research targeting a 1.05T-parameter sparse MoE on consumer hardware.
A version-controlled observatory of human belief about AGI.
// MORE
Single-binary local PKM: SQLite FTS5 + HNSW hybrid search, cited RAG chat to exact file/line via Ollama.
repoReal Linux desktops on Windows via WSL2/XRDP from one control surface; Arch-first MVP shipped as pane.exe.
repoSix-chapter interactive in-browser curriculum: linear algebra → calculus → backprop.
repoAny website, video, or shader as a live Windows wallpaper.
repoWindows shell environment — spec phase.
repoTerminal environment for Windows — spec phase.
repoQuantum-entanglement-between-browser-windows toy.
repo