Extract. Transform. Read.

A newsletter from Pipeline​

Hi past, present or future data professional!

Like me, if you’ve coded for any period of time, you’ll know the familiar dread of being stuck in “tutorial hell.”

Now, to clarify, “tutorial hell” isn’t a label that comes because of a ”victim”’s suffering. Instead, it suggests the inability to escape a vicious cycle of similar tutorials when searching for an answer.

While AI has improved many aspects of the day-to-day data profession, it has accelerated a dev’s ability to be stuck not just in a tutorial hell but in other kinds of “hell” too, notably:

• Collaborator hell
• ​Optimization hell​
• Debugging hell

For as “friendly” as LLMs may appear, they’re not great at telling you “‘no” when you push them for help with a problem. This means you and a model are left frustrated when confronted with difficult, nuanced problems neither can solve.

Of the hells presented, collaborator hell is most easy to fall into. With more vendors and data-driven orgs pushing for real-time collaboration between dev and agent, the wrong auto-correct or suggestion can introduce bugs that wouldn’t exist had a dev been more consciously developing by hand.

Bugs can quickly land a coder in Debugging hell because unless you also understand the full scope of your problem, an AI Agent must “fill in the blanks” in order to help you out of your jam.

For me, nothing compares to being caught in an Optimization Hell in which you ask an Agent to help revise some code and they make so many changes to “optimize” your work that it becomes unrecognizable.

When I initially wrote about tutorial hell, I was referring to the mindless loop of video tutorials and blog posts one could fall into when searching something like “how to use Google Cloud Logging language.”

In the context of AI I’m adding the following suggestions to my initial points:

• Avoid generating excess code
• When providing an error log, only highlight the important bits for the LLM
• Provide (privacy-compliant) samples of your data whenever possible
• Ask the LLM for an output and a testing config before copy/pasting its response
• If you’re going in circles, start a new chat; don’t be afraid to explore other LLMs

In some cases I’ve found myself so frustrated with an Agent’s response that I’ve returned to Stack Overflow to see how humans arrived at a solution.

Even if I don’t find the fix I’m looking for, it at least provides me comfort knowing I’m not the only one trapped in this particular development hell.

Thanks for ingesting,

-Zach Quinn