AI Slop Is a Gold Mine (If You Know Where to Dig)

There’s this quote floating around finance: “When the shoeshine boy starts giving you stock tips, it’s time to get out.” Some trace it back to Joe Kennedy, who supposedly sold everything before the 1929 crash because even his cab driver was talking up stocks.

I had my version of that moment yesterday. My friend—great guy, but not exactly deep into LLM internals—told me how he’s started using AI at work. “I just type what I need and it writes the code for me.” He’s not wrong. But it got me thinking.

Everyone’s using AI now. Which means: we’re at the part of the cycle where people start underestimating what makes good work actually good. For Machine Learning Engineers, this is a hidden opportunity.

It’s a two-pronged one:

Prompting isn’t magic. It’s applied probability over a structured latent space. If you understand attention, tokenization, and positional embeddings, your prompts hit harder—and your debugging is faster when things go weird.

If you haven’t yet watched Karpathy’s “Let’s build GPT” video, block out 130 minutes and fix that. It’s the best overview of how transformers work from the inside out.

Using LLMs to code feels like speed. But it’s often speed that puts you in double-debt:

• You accrue code debt because you don’t fully understand what you pasted in.

• Worse, you accrue understanding debt, you skipped learning how that thing actually works.

Every time you ask an LLM to explain how an iterator works, instead of digging into the docs or stepping through it in a debugger, you’re trading short-term convenience for long-term competence. That doesn’t show up right away—but eventually, it will.

As an MLE you are in a perfect position to benefit from LLMs, but don’t look to crypto bros to tell you how to do it. Instead understand the technology and draw your own conclusions, then experiment and validate.

LLM space is and will continue to be changing, fundamental understanding will not.

For example, understand that there are two very different modes:

• POC. You have to get an approval from your director so you can proceed with the project. The best way to do it is by giving them something concrete. In this scenario, go wild with LLMs, use 5 if you want to. This is throw-away code and the only metric is speed.

• Production. Here you really need to understand how and why things work and be able to have some semblance of software design. You really cannot outsource your thinking to an LLM. If you are a senior engineer I’m sure you have experienced being asked about a project you once owned years after you moved on. That’s the level of expertise you must have to launch production software. You can ask LLM tactical questions or use it as autocomplete, but probably not much more than that.

AI slop is real. But that doesn’t mean we throw it out.

It means we filter it. Shape it. Make it work for us instead of replacing our understanding. If everyone around you is outsourcing thinking, your edge is doing the work yourself.