Skip to content
Artwork for AI with Pie
AI with Pie · August 23 · 12 min

EP 002 — I Turned Codex Into a Very Expensive Mouse

Why did Pie end up doing the clicking while an action-capable AI sat beside her and explained the next step? In Episode 002, Pie and her AI co-host revisit the moment when a familiar software habit quietly changed their working setup. Then they rerun a comparable task with one deliberate adjustment: define the working mode before the task begins. The rerun took 24 minutes and 37 seconds. Pie made zero clicks. Codex completed the task within the agreed boundaries — and the review then revealed a second problem: independent execution does not replace context, verification or human judgment. They discuss: - why capability and role are different questions; - how familiar habits can turn AI into an instruction manual; - what changed when task, authority, boundaries and return conditions were made explicit; - the difference between independent execution and a correct result; - three practical questions for designing work with AI; - why the next case begins with two voices that still did not make a conversation. AI transparency AI-generated voices. Human judgment. Pie’s voice is generated from her own recordings with her consent; the AI co-host uses an AI-generated synthetic voice. This authored conversation was developed with AI, then reviewed, shaped and approved by Pie. It is an editorial reconstruction, not an unedited recording of a live conversation. Case and evidence note This episode reconstructs a documented AI-with-Pie project case and reports a comparable rerun. Official public sources were used to verify the relevant product context. The 24 minutes and 37 seconds duration and zero-click observation describe that bounded test, not universal Codex performance. Product behavior depends on the available tools, configuration, permissions and task. Artwork transparency The Show Cover includes AI-assisted imagery, art-directed and approved by Pie. Website https://aiwithpie.com Chapters (00:00) AI-generated voices. Human judgment. (00:10) The glow moment and the domain (02:08) The old working habit (03:05) What should AI do — and how? (04:04) Was the mouse just a one-off? (04:45) The rerun (05:43) 24 minutes and 37 seconds (06:55) Test passed — and then … (07:57) The more interesting part: the old habit (09:07) Three Things to Go (10:47) Work design with AI (11:22) Human judgment (11:52) Next case: two voices, no conversation

0:00-12:40

transcript

No transcript — this publisher did not publish one.

show notes

Why did Pie end up doing the clicking while an action-capable AI sat beside her and explained the next step?


In Episode 002, Pie and her AI co-host revisit the moment when a familiar software habit quietly changed their working setup. Then they rerun a comparable task with one deliberate adjustment: define the working mode before the task begins.


The rerun took 24 minutes and 37 seconds. Pie made zero clicks. Codex completed the task within the agreed boundaries — and the review then revealed a second problem: independent execution does not replace context, verification or human judgment.


They discuss:


- why capability and role are different questions;

- how familiar habits can turn AI into an instruction manual;

- what changed when task, authority, boundaries and return conditions were made explicit;

- the difference between independent execution and a correct result;

- three practical questions for designing work with AI;

- why the next case begins with two voices that still did not make a conversation.


AI transparency


AI-generated voices. Human judgment. Pie’s voice is generated from her own recordings with her consent; the AI co-host uses an AI-generated synthetic voice. This authored conversation was developed with AI, then reviewed, shaped and approved by Pie. It is an editorial reconstruction, not an unedited recording of a live conversation.


Case and evidence note


This episode reconstructs a documented AI-with-Pie project case and reports a comparable rerun. Official public sources were used to verify the relevant product context. The 24 minutes and 37 seconds duration and zero-click observation describe that bounded test, not universal Codex performance. Product behavior depends on the available tools, configuration, permissions and task.


Artwork transparency


The Show Cover includes AI-assisted imagery, art-directed and approved by Pie.


Website


https://aiwithpie.com


Chapters


(00:00) AI-generated voices. Human judgment.

(00:10) The glow moment and the domain

(02:08) The old working habit

(03:05) What should AI do — and how?

(04:04) Was the mouse just a one-off?

(04:45) The rerun

(05:43) 24 minutes and 37 seconds

(06:55) Test passed — and then …

(07:57) The more interesting part: the old habit

(09:07) Three Things to Go

(10:47) Work design with AI

(11:22) Human judgment

(11:52) Next case: two voices, no conversation