Elie Bakouch

Bakouch announced his move to Prime Intellect to work on pretraining and mid-training, reflecting on the SmolLM models his team had developed at Hugging Face. He is also a co-author of the INTELLECT-1 report, which describes training a ten-billion-parameter model across geographically distributed contributors. The report’s infrastructure handles changing groups of available devices and recovery from interrupted workers. It explores how training can continue when participants and network conditions are less predictable than in a single managed cluster.[2][4]

Measuring research rather than assuming it

In a Prime Intellect study, Bakouch and collaborators ran 153 autonomous research attempts across 18 frontier models on a nanoGPT optimizer speedrun. Agents received a training baseline and sought improvements without internet access. The team compared repeated runs and released traces and experiment records. Results varied substantially between models and seeds. The study found useful improvements but no fundamentally new method among the runs, making it a test of experimentation and execution as well as idea generation. Its narrow setting does not establish general autonomous scientific ability.[3]

An environment that preserves the work

Bakouch co-authored the Prime Agent technical report. Its harness gives an agent a persistent IPython environment for processing context and running computations, while preserving histories, memories and reusable instructions across attempts. Subagents can communicate directly, and a human-facing view exposes ongoing sessions. The design separates execution, recovery and resource accounting from the strategy the model chooses. This makes failures of the surrounding software less likely to obscure what the model itself can accomplish.[5]