Heard on a podcast

Recursive focuses on automated self-improving research agents

Richard Socher claims that neurosymbolic reasoning through code allows LLMs to outperform human experts in optimizing CUDA kernels and NanoGPT benchmarks.