Robert McHardy

Robert McHardy leads pre-training at Poolside, building the systems behind its open-weight Laguna coding models. His work spans model architecture, distributed training, numerical stability, and evaluation across thousands of GPUs.

From computational biology to coding models

McHardy studied natural language processing at the University of Stuttgart and machine learning at University College London, where he worked with the Deciding, Acting, and Reasoning with Knowledge laboratory. Early projects included a CYK parser and a machine-learning agent for the Halite III programming competition.

At InstaDeep’s joint research laboratory with BioNTech, he worked on cancer and infectious-disease applications and coauthored LightMHC, which uses graph neural networks to predict peptide–major histocompatibility complex structures. He subsequently became a senior researcher at AssemblyAI, leading pre-training for Universal-1, a multilingual speech-recognition system trained on more than 12.5 million hours of audio. Its technical report, which he coauthored, describes practical capabilities including multilingual transcription, code-switching, and accurate timestamps.

At Poolside, McHardy has led pre-training for Laguna M.1, Laguna XS.2, and Laguna S 2.1. He coauthored the Laguna M.1/XS.2 technical report and has concentrated on making large coding-model training observable, reproducible, and resilient to subtle failures.

What reliable model training actually requires

• Replica-hash verification: McHardy’s team compares hashes of supposedly identical model replicas and stops training when their weights diverge. The safeguard exposed a faulty GPU that silently corrupted computations, producing abnormal loss curves and gradient norms despite an unchanged training configuration.

• Numerical precision at the failure point: During Laguna M.1 training, growing activations overwhelmed BF16 precision during tensor-parallel accumulation in the unembedding layer. Switching that operation to FP32 restored convergence and prevented instability from propagating backward through the model.

• Distributed safeguards have blind spots: An FP8-kernel race condition produced illegal memory accesses, invalid gradients, and silent corruption that ordinary replica checks could not reliably detect because replicas processed different data. McHardy has pursued controlled checks capable of comparing equivalent computations directly.

• Public evaluation trajectories: Laguna S 2.1 is an open-weight mixture-of-experts coding model with 118 billion total parameters and 8 billion active parameters per token. McHardy highlighted its publicly available evaluation trajectories, enabling closer inspection of agentic coding results while distinguishing preliminary base-model evaluations from eventual post-training performance.