← back

Cognitive Exhaust Fumes, or: Read-Only AI Is Underrated — Šimon Podhajský, Head of AI, Waypoint

1.8K views · Apr 08, 2026 · 11:30 min · Watch on YouTube ↗
Takeaway

Read-only personal AI that only analyzes your data — never acts — produces deeper self-knowledge with bounded risk, and deserves to stand apart from agentic systems.

Summary

  • Builds 'Fulan', a personal AI with read-only access to six data sources (email, browser, journal, CRM/Clay, etc.); writes only to a separate Obsidian vault.
  • Analyzes 'cognitive exhaust fumes' — your digital byproducts — for intention-action gaps, attention drift, and relationship decay only visible cross-source.
  • Weekly-reflection Claude skill runs a Python script that pulls from sources, hits the Anthropic API for structured outputs, and produces a markdown review doc.
  • Argues read-only AI is its own product category, not a stepping stone: zero downside on errors vs unbounded write-mode risk; avoids contaminating the very signal it studies.
  • Acknowledges security exposure: mosaic effect and Simon Willison's lethal trifecta (private data + untrusted content + external comms via shell access).
personal-airead-only-aiprivacy
Original description
Every other personal AI demo has agents sending emails and managing calendars. I built the opposite: a read-only system that queries my data sources (email, journal, tasks, CRM, browser sessions, notes) but can't modify any of them. This is an intentional limitation. I'll cover why trust asymmetry matters (read is safe, write is dangerous), how cross-source pattern detection beats task automation, and why ""exhaust fume analysis"" of one's cognition is more valuable than yet another AI assistant trying to act on your behalf.

Šimon Podhajský - Head of AI, Waypoint AI

I'm Head of AI at Waypoint and a full-stack builder with a background in data science and data engineering. I built this personal AI system to scratch my own itch -- and discovered that the ""read-only"" constraint led to better architecture than the agent-first approaches I see everywhere.

I made a Github repo with a template for people to try out the read-only AI / personal intelligence system: https://github.com/shippy/personal-intelligence-kit 

Socials:
https://linkedin.com/in/simonpodhajsky
https://x.com/sim_pod
https://simon.podhajsky.net

Slides:
https://slides.podhajsky.net/read-only-ai