1 "Nucleus Sampling" Post

Temperature and Top-P: The Creativity Knobs

How sampling parameters shape AI personality

Dec 24, 2025 • Learning Paths: How Large Language Models Work · Understanding Model Behavior • Category: AI and the Mathematics of Language • Tags: #AI · #Anthropic · #LLM · #nucleus sampling · #OpenAI · #probability · #sampling · #softmax · #temperature · #top-p

What You'll Learn

How an LLM turns raw token scores into probabilities before choosing what comes next
How temperature reshapes a probability distribution to make outputs more predictable or more varied
How top-p sampling limits which tokens the model is allowed to consider
Why top-p adapts to model confidence differently from a fixed top-k cutoff
How temperature and top-p interact when both are applied to the same distribution
How to choose sampling settings for factual, structured, professional, and creative tasks

Behind the Code

My focus is the engineering beneath modern AI systems. After several years of building applications with AI through APIs, my interests have moved below that abstraction layer: understanding how models are architected, trained, evaluated, and served, why they’re designed the way they are, and the engineering tradeoffs behind those decisions.

That direction follows a pattern across my forty-year career as a software developer. I’ve primarily built applications, but I’ve repeatedly found myself pushing beyond the application itself and building the systems, abstractions, and tooling behind it: metaprogramming layers, dynamic runtimes, configuration-driven engines, and automation platforms designed to outlive the original feature request.

The pattern shows across my career: Forms Express at AT&T, a four-layer dynamic runtime shipped for under $500K against a $2M consulting bid; VeriSign’s first-ever customer self-service mobile activation system, which won the 2003 SUPERQuest Award; and TAG at Microsoft, a production LLM orchestration platform that shipped nearly 100 Terraform articles by parsing XML templates, routing prompts to Azure OpenAI, and composing results with browser-driven Azure portal capture. Along the way: C++ MVP, nine technology books, 100+ articles for .NET Programming Tips & Techniques, and helping run CodeGuru, the largest Windows developer community of its era.

After 21 years at Microsoft, I’m now full-time on what’s next: completing my Master’s in Computer Science and publishing at Signal & Syntax, where I research and write about the mathematics and engineering underneath modern AI. I use mathematical models, experiments, and small implementations to explore everything from gradient descent and attention to inference-time behavior and training infrastructure; not just how to use these systems through an API, but how the machinery beneath that API actually works.

Connect with me on LinkedIn .