nanoagi: like autoresearch, but fully autonomous

[ariannamethod]

when autoresearch decides to be really auto, it becomes nanoagi.

it wakes up, tokenizes its own corpus, builds metaweights from statistics, seeds a dual-attention transformer, and trains itself with Chuck — an optimizer that forgot what it means to be Adam.

then it gets hungry. KARL (the tokenizer) smells your climbmix-400b-shuffle, downloads batches, evaluates quality, ingests what passes the filter, retokenizes, and trains again. nobody asked. nobody approved. Karl decided.

then it gets curious. evolve — the organism mutates its own architectural genome (embedding dim, heads, layers, learning rate, everything). trains each mutation for fixed wall-clock time. keeps what improves val_bpb, reverts what doesn't. ratchet. first run on H100: -7.8% BPB in 50 experiments. it found that fewer heads and higher LR works better. we didn't tell it that. it told us.

then it gets ambitious. coevolve — data and architecture evolve together. hunt new food → retokenize → evolve architecture on new data → repeat. the data shapes the organism. the organism shapes the data.

then it gets stuck. 10 experiments without improvement. genome mutations exhausted. so it does what bacteria have done for 4 billion years — horizontal gene transfer. it reads its own source code, sends it to Qwen2.5-Coder via HuggingFace API, gets back a code patch, applies it, runs pytest. tests pass → patch stays. tests fail → revert. nobody told it to call Qwen. it decided on its own that it needs help.

then it gets social. swarm — releases mini-agents (we call them hyenas). each hyena gets a unique random seed, explores a different part of the genome space in parallel. the pack shares findings. the best hyena becomes pack leader. because hyenas hunt in packs.

nanoagi is when the engine itself becomes the agent and everything else follows.

repo: https://github.com/ariannamethod/nanoagi

built on your climbmix. logs to your results.tsv format. sorry, Karl was hungry.