Skip to content
01 of 06Reasoning Models

Reasoning Models

Library

Series and technical notes.

You are in Reasoning Models · Introduction.

Series

Reasoning Models

6 items

Watch video, summary and related content

Complete Technical ~45 min 5 chapters
Prerequisites: AI and Generative AI Foundations

LLMs can appear to reason, but "reasoning" is a deeply human concept. We can nevertheless model it as a process made of multiple steps whose execution consumes physical time (latency), computational effort (compute) and remains fallible (hallucinations and other failures).

This series develops one central idea: if reasoning is a process, then execution time is another system variable. You can pay for more steps, more samples, more verification or more interaction with tools to improve answer quality—but each lever changes cost, latency and failure surface.

Contents

1. What "reasoning" means for an LLM

  • Which definitions of reasoning can be useful in the context of language models.
  • The emergence of explicit reasoning-model product families, with OpenAI o1 as an important inflection point.
  • Apple's The Illusion of Thinking paper and the subsequent debate and responses.

2. How these systems fail

  • Failure modes are not purely random: shortcuts, systematic errors, objective drift and other recurring patterns.
  • Methods for detecting and mitigating these failures.

3. Test-Time Compute

  • Test-time compute as an additional scaling axis for generative AI.
  • Levers for exploiting it: more internal steps, more candidate generations and more structure.
  • The relationship between higher answer quality, higher cost and higher latency.

4. Physical time: latency, streaming and human interaction

  • "Thinking longer" is cheap in a paper; in a product the user waits, the session costs more and the system gains additional failure points.
  • Where the acceptable latency threshold lies for a given task and user.
  • Patterns for extracting the benefit of test-time compute without destroying the interactive experience.

5. Risks: overthinking, cost, attacks and alignment

  • Why more test-time compute can produce overthinking, unproductive loops and degraded quality.
  • Quality vs cost vs latency as a product problem: SLOs, queues, unpredictable bills and user experience.
  • New risks with tools / RAG / browsing: prompt injection, contaminated context and tool misuse.
  • Design criteria: hard budgets for time/tokens/tools, stopping signals, verification where critical and fallbacks such as asking for data, degrading gracefully, abstaining or escalating.

Related series: AI and Generative AI Foundations · From the Caves to AGI

View all series

Keep learning
Next chapterWhat is reasoning?Reasoning Models
All series →