Papers · PaLM
AI PAPER

PaLM: Scaling Language Modeling with Pathways

Scaled a dense Transformer to 540 billion parameters using Google's Pathways system, demonstrating breakthrough performance on reasoning tasks and discontinuous capability jumps at scale.

Authors
Aakanksha Chowdhery, Sharan Narang, Jacob Devlin +7
Institution
Google
Published
April NaN, 2022
Citations
6,000
Category
Scaling
Impact
major
PAPER EXPLAINED

Plain-English Summary

PaLM was Google’s most ambitious language model at the time of release, with 540 billion parameters trained across 6,144 TPU chips using Google’s Pathways infrastructure. It was not just bigger — it demonstrated qualitative jumps in capability that smaller models could not achieve, particularly in reasoning, code generation, and explanation tasks.

The most striking finding was “breakthrough” capabilities that appeared suddenly at scale. Tasks where smaller PaLM models performed at chance level suddenly became solvable when the model reached sufficient size. This provided strong evidence for emergent capabilities — abilities that are absent in smaller models but appear in larger ones.

Think of it as the difference between a student who has studied for a few hours versus one who has studied for years. At some point, understanding “clicks” and they can tackle problems that were previously impossible, not just marginally harder.

Key Innovation

PaLM’s technical contribution was demonstrating efficient scaling to 540B parameters on 6,144 TPUs using the Pathways system, which coordinates training across multiple TPU pods. The model used a standard dense Transformer but achieved exceptional training efficiency — 57.8% hardware utilization at that scale, well above typical numbers.

The paper’s analytical contribution was even more significant: a rigorous study of emergent capabilities. By evaluating models at 8B, 62B, and 540B parameters on hundreds of tasks, PaLM mapped out which capabilities emerge discontinuously with scale, providing empirical grounding for the debate about AI capability trajectories.

Impact on the Field

PaLM influenced the scaling ambitions of every major AI lab. Its demonstration of emergent capabilities at scale strengthened the case for building ever-larger models. The chain-of-thought reasoning results (where PaLM 540B surpassed previous state-of-the-art on reasoning benchmarks) showed that scale alone could unlock capabilities previously thought to require specialized architectures.

The Pathways infrastructure innovations also influenced how labs think about distributed training across multiple data centers and hardware configurations.

Models That Built on This

PaLM 2 refined the approach with Chinchilla-optimal training and improved data mixtures, becoming the backbone of Google’s Bard chatbot. Med-PaLM applied PaLM to medical question answering with expert-level accuracy. Gemini, Google’s current frontier model family, builds on lessons learned from PaLM’s development. The emergent capabilities research from PaLM also directly influenced subsequent debates about AI safety and capability forecasting.