An interactive browser demo for exploring whether a learned meta-model can predict the behavior of neural-network architectures and training dynamics.
Live demo: https://metacircleai.github.io/ai4ai-demo/
The demo has two focused views:
- Architecture design: assemble a valid DAG from four fixed-width MLP parts and one softmax Gate, then inspect the predicted Log MAE or search for an architecture near a requested value.
- Training dynamics: vary activation, depth, width, and learning rate, then compare a frozen prediction with the measured reference curve.
All inference runs in the browser. The architecture view loads a static table from one deterministic forward pass of a trained baseline GNN over the 10,006 canonical DAGs. The training-dynamics view uses frozen browser-side weights and reference curves; neither view requires a Python server.
This is a controlled research demo, not an architecture-performance guarantee. The displayed architecture targets are preliminary and inherit the noise of the underlying candidate-training protocol.
The architecture editor is served at /, with /ai4ai/ retained as a direct
alias. The training-dynamics interface is served at /dynamics/, and the
frozen classic interface remains at /classic/.
pnpm install --ignore-scripts
pnpm run devIf the system does not expose node and pnpm, add a compatible Node.js
installation to PATH, or invoke the local Vite binary directly after
installing dependencies:
PATH="/path/to/node/bin:$PATH" ./node_modules/.bin/vite --host 127.0.0.1 --port 5173The local production check is:
PATH="/path/to/node/bin:$PATH" ./node_modules/.bin/tsc -b
PATH="/path/to/node/bin:$PATH" ./node_modules/.bin/vite buildThe GitHub Pages build uses the repository base path:
pnpm run build:pagesThe deployment workflow is configured in
.github/workflows/deploy-pages.yml and publishes the static dist/ output
through GitHub Pages.