scripod.com
Building Pi, and what makes self-modifying software so fascinating

Overview

Shownote

Highlights

Transcript

Chapters

Pins

Building Pi, and what makes self-modifying software so fascinating

The Pragmatic Engineer

19 HOURS AGO
The Pragmatic Engineer

The Pragmatic Engineer

19 HOURS AGO
OverviewShownoteHighlightsTranscriptChaptersPins

Shownote

Brought to You By: • Statsig — ⁠ The unified platform for flags, analytics, experiments, and more. • Sonar – The makers of SonarQube, the industry standard for automated code review • WorkOS – Everything you need to make your app enterprise ready. — M...

Highlights

In this episode, Mario Zechner—the creator of the minimalist, self-modifying AI coding agent Pi—and Armin Ronacher—the creator of Flask—engage in a candid, deeply technical conversation about their evolving relationship with AI tools. Both seasoned open-source builders with roots in retro computing and game development, they reflect not as AI evangelists but as skeptical practitioners who build *with* AI while remaining wary of its overreach.
05:06
Armin and Mario, along with Peter Steinberger, were experimenting with AI before OpenClaw and Pi
10:07
AI agents gain real utility when granted direct access to files
20:25
A good engineer says no to keep complexity down, while using agents often leads to saying yes without considering complexity
21:50
AI agents give access to world knowledge, making it harder for experienced engineers to say no to juniors or product managers
27:05
We'll realize AI is better for jobs earlier than in previous cycles
31:14
Pi succeeded by packaging cloud code into an intuitive, workflow-aligned agentic search tool
42:48
Pi was built to enable specialized performance for specific tasks, with Mario Zechner using only two trivial extensions for the pymonorepository
48:09
OpenClaw uses an auto-closing mechanism for pull requests from unknown users
56:32
Few AI-engineered projects have been successful
57:33
We're in a 'messing around and finding out' stage
1:00:23
Machine learning models may converge towards the 'mean' of the garbage code on the internet
1:05:33
Friction, like checklists and approval processes, is deliberately added to slow things down and help engineers think before making significant changes
1:14:32
The best software specification is the software itself
1:19:29
CodeMode proposes TypeScript-based MCP servers to improve coherence and composability
1:25:06
There's a split in access to top-tier AI intelligence

Chapters

Intro
00:00
How Mario, Armin, and Peter Steinberger met
07:30
How 30 dev teams use AI agents: learnings
15:15
The importance of judgment
21:50
Challenges when non-engineers write code
24:26
Downsides of over-automation
28:30
Pi
32:18
OpenClaw + Pi
48:09
“Clankers”
50:54
Open source and AI
57:32
Complexity as the enemy
1:00:22
Building an AI-native startup
1:02:50
“Slow the F down”
1:11:52
MCPs vs. CLI
1:16:40
Predictions and staying up to date
1:25:03

Transcript

Gergely Orosz: What if I told you that? One of the most influential AI coding agents of 2026 was built by a single developer in Austria, who got frustrated with existing AI coding agents? This is Pi, a minimalist, self-modifying coding agent, which has qui...