---
date: "2026-10-06T15:18:24.000Z"
title: "2026-10-06"
draft: false
---

import LinkedInPost from '@components/LinkedInPost.astro';

<LinkedInPost id="https://www.linkedin.com/feed/update/urn:li:activity:7513256352631222273/">

“AI” has come to mean too many things.

If someone tells you they work “in AI,” they could be working on training neural networks, writing evals, doing harness engineering, or writing software with coding agents. All of these could justifiably be described as “AI,” but it’s equally possible that all of these folks have little to no idea about each other’s work.

When it comes to making AI useful in your work, you first need to concretely identify the work that needs doing. From there, we can identify which parts of AI can be useful.

Otherwise, you’re just pointing an agent at a problem and hoping for the best, and you’ll likely be disappointed.

</LinkedInPost>

---

<LinkedInPost id="https://www.linkedin.com/feed/update/urn:li:activity:7513325307572011008/">

When I was learning to be an engineer, I learned the most from sitting next to other engineers, watching how they worked, and asking them questions. I felt fortunate that, by the time so much work had gone remote/hybrid, I had learned this way.

It's hard to imagine really getting good at a craft without seeing how people who are better than you actually do the work.

The same goes for getting good at working with agents today.

With code, we at least had the benefit that if we couldn't watch someone writing code, we could read the code they wrote, the PR, the description, and even the incremental commits they made along the way.

With agents, the same artifacts are still there, sort of, but void of all the process and context used to get to create them. Many people aren't making commits themselves. They are directing agents to do the work for them. But it's now harder than ever to observe the process they followed to get to the finished result.

What prompts did they write? What context did they include?

I hardly know anyone who shares this information along with their PRs or software. If you sat next to me, you could watch me prompt. You'd see me using voice to text for several minutes to describe to the agent how I want to stitch together two partial implementations, taking the parts I need from each, then polishing the API and creating verification tests.

If we paired, I could guide you in this direction. But independently, you have to chart this course alone. For some people, it's intuitive and easy, but for others, it doesn't make sense until you see it for yourself.

</LinkedInPost>