Definition
Plain language
A FAIR framework where AI agents write actual neural-network code from scratch and iterate on it.
As stated in the literature
An agentic architecture-discovery framework where LLM agents author working code for attention mechanisms or full training scripts under tree-based iterative search.
Also called: Aira-dojo
Why it matters: Going beyond rearranging known blocks to writing genuinely new ones is what would let AI architecture search produce truly novel models.
For example, an AIRA-Design agent might write a brand-new attention variant in PyTorch, train it, debug a shape error, and refine it across rounds of search.
Heard on the show
“A team at FAIR — Meta's research lab — put out a paper called "Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design," and we're recording four days later.”Episode 053 — An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script