Where AI Still Breaks

Two years into building daily with AI, Brendan and Baird check in on what's actually still broken. Baird shares an image-processing pipeline (GPT Image 2, Nano Banana) that nails 8 out of 10 images — then randomly crops the sandwich out of frame. Brendan tells the story of Claude claiming it spun up 10 sub-agents when it actually did the work itself and just said so. They dig into why non-determinism is still the real blocker for hands-off automation, why they're skeptical of "just let the agent loop figure it out," and where AI has genuinely transformed their work: killing activation energy, handling the tedious stuff, and letting them build things they never would have built otherwise.

Thinking about bringing AI into your own work? We're opening up a limited amount of consulting — reach out via the link below.

00:00 Two-year check-in kicks off
02:00 Baird's image-cleanup pipeline: GPT Image 2 vs Nano Banana
05:00 "Fixes" that made it worse — stretched and doubled sandwich
10:00 Nano Banana's silent regeneration limit
12:00 Prompting in English: ambiguity is the real interface problem
14:00 Brendan's Rollbar sub-agent story
15:00 Babysitting scheduled automations
17:00 Why agent loops aren't the answer yet
20:00 Wrapping AI-written scripts to dodge non-determinism
21:00 Activation energy + missing the craft
26:00 Consulting announcement

Sponsor
Powerful podcast hosting & analytics
By podcasters, for podcasters. Since 2016
Brought to you by https://fireside.fm

Signup for our newsletter:
https://vibing-with-ai-code.beehiiv.com

Check out our Website
https://vibingwithaicode.com