Skip to content
The Fizz

Inside AI Pathfinder. The System Behind a Guided AI Tool

2026-05-2737:31

Anthropic's Rise and Market Positioning

Cultural Differences in AI Companies

The Role of AI Models and User Experience

Navigating AI Adoption Challenges

The Future of AI Companies and Market Trends

Building an AI Harness: Introduction to Pathfinder Method

Technical Implementation: Balancing Determinism and Inference

State Machine: Managing User Journey and AI Interaction

Turn Contracts: Ensuring User Engagement and Clarity

Data Management: Recording User Input for Contextual Awareness

Quality Assurance: Evaluating AI Responses for User Experience

Multi-Model Approach: Enhancing AI Reliability and Context Awareness

Navigating Chaos with Structure

The Role of Third-Party Validation

User Experience and AI Interaction

Handling Edge Cases in AI

Empowering Users Through Control

Escape Hatches in AI Conversations

Building Reliable AI Experiences

Leveraging AI for Broader Impact

Creating Actionable Outcomes for Users

Justin and Kellan go under the hood of AI Pathfinder, the guided tool they've been building, and break down what it actually takes to make an AI conversation reliable. The short version: the model is the easy part — it's everything wrapped around it that makes or breaks the experience.

The Drip:

  • Andrej Karpathy joins Anthropic; Anthropic acquires Stainless (~$300M) to convert API docs into MCP servers and CLIs
  • AI companies vs. "AI-injected" monoliths — and why Google, with a frontier model, an app stack, Android, and cloud, is the traditional peer to watch
  • Why a model performs better inside its own harness (Opus in Claude Code vs. a raw API call) — and why "we're an Anthropic shop / an OpenAI shop" is where teams are headed

Inside The Bottle:

  • The seven concepts behind a reliable guided AI experience: state machine, turn contract, writing notes to a database, quality evaluators, a second-opinion model, trusting the data over the prose, and an escape hatch
  • Why you can't trust the LLM to track its own place in a process — or to remember the whole conversation
  • The real challenge of balancing determinism and inference, decided question by question
  • Knowing when to let the user move on, and why the escape hatch was the first safeguard they built
  • The leverage play: turning their consulting methodology into something that can help far more people

New episodes every Wednesday. Subscribe to The Weekly Fizz newsletter at http://tinybottleai.com/the-fizz.

Newsletter

The Weekly Fizz.

Real examples. Real numbers. No hype. No spam, unsubscribe anytime.

Got something to say?

We're always looking for interesting guests and fresh perspectives.