Poolside’s Marah Abdin and Robert McHardy described on 2026-07-26 the engineering reality behind scaling large language models, emphasizing that what works at 1B parameters breaks catastrophically at 100B. Their talk covered poolside’s transition from the Laguna M.1 model through the open-weight XGen-2 (33B parameters, released on Hugging Face in July 2026) to the forthcoming Laguna S (118B total / 8B active, trained on 30 trillion tokens across 4,000 GPUs). Abdin framed synthetic data as a lever to extract latent structure, not a replacement for organic web data. The final pre-training mix for XGen-2 included 13% synthetic tokens, achieved via a six-component modular pipeline called Hive applied through four shapes: rephrasing, multi-stage generation, cross-domain porting, and multi-turn chat loops. McHardy catalogued failures at scale, including broken GPUs causing silent data corruption, exploding activations from BF16 tensor-parallel accumulation, and a race condition in DeepChem’s FP8 kernels that silently corrupted 0.5% of gradients, each caught by periodic weight-hash comparisons. Early benchmark results for the still-training Laguna S already outperform DeepSeek V4 Flash Max and Nemotron on SWE-bench agentic coding, while trailing on MMLU-Pro.