Medium Foundation Models · 1 min read
PaLM 540B: Google's GPT-3 answer brings chain-of-thought
In one sentence Google publishes PaLM, a 540B-parameter model trained on the new Pathways system. Demonstrates emergent reasoning capabilities when guided with chain-of-thought.
Reading level
Google publishes a paper on PaLM, a giant language model — larger than GPT-3 — and shows something new: if you ask it to "think out loud" before answering math and logic problems, it becomes much more reliable.
The trick is called chain-of-thought: instead of "what's the answer to…", you say "explain step by step how you get to…". Sounds obvious, but on non-trivial problems the difference is enormous.
It's not publicly available, but the prompt pattern immediately joins the LLM practitioner toolkit.
Companies
Tools
PaLM
Tags
GooglePaLMPathwaysChain-of-ThoughtReasoning
Sources