The New Automation Paradigm: Leveraging AI When the Cost is Near Zero

You should automate a task if the benefit outweighs the cost. While this principle remains constant, a fundamental shift in the cost component has occurred recently. I used to hesitate before scripting repetitive work due to the time investment, finding most tasks not common enough to warrant automation.

What Has Changed in the Last Year?

The advent of code generation models and agentic tooling (such as Codex or Claude Code – not an advertisement) has completely transformed my approach this year. Now, automating work is virtually free, leading me to automate almost anything easily accessible from the terminal.

Credit: XKCD 1205

Here are a few recent examples from my work:

Podcast Transcript Processing

• Problem: Audio transcriptions for my podcast are often inaccurate and require manual correction. While Large Language Models (LLMs) can fix these, they cannot process an entire two-hour transcript in one go.
• Solution: I generated a script to chunk up the transcript, feeding each segment to an LLM API. Now, it's a single command, and the task is complete.

Rescuing Podcast Footage

• Problem: I use software to record remote podcasts, which uploads video in chunks. During one interview, the uploads failed. Although we couldn't fix the upload issue, we located all the video chunks in the interviewee's Chrome folder. He then sent me a ZIP file containing hundreds of numbered video segments.
• Solution: I generated a script to list and sort these files, then used ffmpeg to stitch them all together into a complete video.
• Note: Even though this was a unique task unlikely to be repeated, generating the script was still significantly easier than manually figuring out the complex ffmpeg command.

Moving Notes to Markdown

• Problem: I wanted to consolidate my notes into Obsidian, but they were scattered across various file formats in Google Docs, Apple Notes, and Google Keep.
• Solution: I exported all notes into ZIP files, placed them in a central location, and wrote three specific scripts—one for each format—to convert them to Markdown.
• Note: Similar to the previous example, this is a one-time migration. However, generating the scripts proved to be the fastest and most efficient method. In the past, I might have foregone such a migration, but with the current ease of automation, it became a straightforward decision.

This shift has profoundly changed how I interact with my computer. I now automate significantly more because the perceived cost is essentially zero. I find myself having AI write quick, one-off scripts more frequently than performing these small, local tasks manually.

These agentic tools not only save me time but also enable me to undertake tasks I previously wouldn't have considered due to the perceived effort. While LLM code generation isn't flawless, for these small, targeted tasks, it performs exceptionally well. LLMs often provide correct solutions on the first attempt, and even if they err, the cost of correction is minimal. Amidst the hype surrounding AI, this represents a tangible use case that already delivers substantial time savings. I believe you'll find it equally beneficial.

I plan to continue writing more, though without a fixed schedule. Please feel free to share any topics you're curious about, and I'll consider them for future posts.

Additionally, I've curated my top three podcast episodes to date, should you be interested in exploring my work further:

• Ethan Evans (Amazon VP) [430k views]: An excellent speaker who provides valuable insights into tech management.
• Philip Su (Open AI & Meta Distinguished Eng) [334k views]: A figure I greatly admired during my career ascent in tech, offering profound insights from the highest levels and reflections on his professional journey.
• Steve Huynh (Amazon Principal Eng) [206k views]: Shares extensive experience from his tenure at Amazon, including insights into performance reviews and interview processes from an insider's perspective.

Thanks for reading, Ryan Peterman

Join 100,000+ software engineers from companies like Google, Meta, Amazon, and Microsoft who receive new posts and support my work.

[Subscribe]