palm (pathways language model)

PaLM (Pathways Language Model) is Google's large-scale language model that demonstrated breakthrough capabilities through massive scaling, achieving state-of-the-art results on hundreds of language understanding, reasoning, and code generation tasks. The original PaLM (Chowdhery et al., 2022) was trained with 540 billion parameters using Google's Pathways system — a distributed computation framework designed to efficiently train models across thousands of TPU chips (6,144 TPU v4 chips for PaLM 540B). PaLM achieved remarkable results: surpassing fine-tuned state-of-the-art on 28 of 29 English NLP benchmarks using few-shot prompting alone, and demonstrating emergent capabilities not present in smaller models — including multi-step reasoning, joke explanation, causal inference, and sophisticated code generation. Key innovations include: efficient scaling through Pathways infrastructure (enabling training at unprecedented scale with high hardware utilization), discontinuous capability improvements (certain abilities appearing suddenly at specific scale thresholds rather than gradually improving), strong chain-of-thought reasoning (solving complex multi-step problems through step-by-step reasoning), and multilingual capability (strong performance across multiple languages despite English-dominated training). PaLM 2 (2023) improved upon the original through several advances: more diverse multilingual training data (over 100 languages), compute-optimal training (applying Chinchilla scaling laws — more data, relatively smaller model), improved reasoning and coding capabilities, and integration across Google products as the foundation for Bard (later Gemini). PaLM 2 came in four sizes (Gecko, Otter, Bison, Unicorn) designed for different deployment scenarios from mobile to cloud. PaLM's architecture uses a standard decoder-only transformer with modifications including SwiGLU activation, parallel attention and feedforward layers (improving training speed by ~15%), multi-query attention (reducing memory during inference), and RoPE positional embeddings.