AI Writes My Code Now. So What Am I For? | An ML Engineer’s Honest Take

In progresseasybiteaivideo2026-10-28read planted ai llmssoftware engineering

Synopsis — AI-drafted from Dan's notes

An 11-minute monologue from the machine learning engineer who runs the channel, back after a break from posting. The host describes months of what they call AI depression: pushing code they mostly reviewed rather than wrote, and a manager who pasted the problem into ChatGPT and got a working answer. The part of the job that is going, the hours of chasing errors through documentation until the bug finally disappears, was also where the daily satisfaction lived. The host finds the same feeling in developer threads online, and names the mixed signals: use AI for everything at work, but still write code by hand in interviews. For ML engineers it is sharper still, since the old core skill, building and tuning models, now competes with a tool anyone can use, while everyone suddenly expects them to be agent and local-LLM experts.

The answer offered is three things engineers still do better. Communication: when a manager asks why the agent added a cache and whether customers could see stale prices, “that’s what Claude did” is not an answer, so the host now questions the AI the way their manager would before handing work over. Ownership: an agent can close ten tickets in half an hour, but one quiet change to a shared function can break something unnoticed for a week, so the host only delegates as much as they can actually review. Fundamentals and curiosity: a fraud model reporting 99% accuracy on data where 1% of transactions are fraud may have caught nothing, and only someone who knows that accuracy lies on imbalanced data will ask. The host admits to having no neat answer, says the anxiety means you care about the craft, and asks viewers whether this is too hopeful.

Connections