In April 2022, Google introduced PaLM, the Pathways Language Model, a single dense language model with 540 billion parameters. It was Google’s flagship demonstration of Pathways, a new training system designed to spread one enormous model efficiently across many chips. PaLM was trained across two Cloud TPU v4 pods using 6,144 chips, which Google described as the largest such configuration used for training to that point, and it became a centerpiece of the company’s research response to the GPT-3 era.
PaLM’s most influential result was about reasoning rather than raw size. Google reported that combining the model’s scale with chain-of-thought prompting, a technique of asking the model to show its intermediate steps, produced breakthrough performance on tasks needing multi-step arithmetic or common-sense reasoning. This was strong evidence that certain capabilities emerge only once a model is large enough and is prompted to reason step by step, a finding that shaped how the whole industry thought about scale and prompting.
PaLM also mattered commercially as a backbone for Google’s products. Its successor, PaLM 2, was the model that powered the first public version of Google’s Bard chatbot in 2023, before Google consolidated its efforts under the Gemini line. PaLM thus sits at the hinge between Google’s research lineage and its consumer AI offerings.
Why business readers should care: PaLM was a clear public marker that scale plus the right prompting method could unlock reasoning abilities that smaller models simply did not have. It reinforced the scaling-driven strategy that defined the period and was the immediate ancestor of the models Google later put in front of millions of users.