Elie Bakouch

Elie Bakouch works on open language models and automated AI research. His 2026 AI Engineer conference biography described him as a Research Engineer at Prime Intellect, working to advance open pre-training and mid-training. Previously at Hugging Face, he created and trained the SmolLM series of efficient language models and contributed to Open-R1, SmolVLM and open pre-training playbooks—comprehensive guides and recipes for training language models from scratch.

In his conference presentation, Bakouch described experiments testing coding agents on the Optimizer Speedrun, where changes are restricted to optimizer-related parameters. Agents proposed experiments, submitted preemptible cluster jobs and read training logs, with record validation requiring a statistical threshold. He reported that Claude Code and Codex surpassed the human record by combining existing methods, without producing a novel optimizer or mechanism. His comparisons examined persistence, memory use, token efficiency and literature search. He also outlined a more controlled benchmark with separate levels of access to external information, and an AlphaEvolve-inspired discovery loop combining idea generation, speedrun evaluation, human judgment and larger-scale testing. He advocates conducting this research openly so researchers beyond major laboratories can understand and evaluate how AI systems perform research.