Sean V. Cohan

Blog

Out the Leetcode Mines

A Rant

For a couple years now, it’s been generally accepted that most companies use some sort of AI in their hiring process. The reasoning is simple: with demand far outpacing supply for engineering roles, it doesn’t feel reasonable to expect an employee to sort through thousands of applications. This is understandable.

So, the equally understandable response by applicants is to begin using AI on their side of the bargain. Why would you bother to cater something to a human reader if the first person to see it will be a robot? So you run your resume through an LLM to fine-tune it perfectly for each and every job, ensuring that you hit every keyword and bullet point that you think the LLM will be looking for.

Suddenly, we have a situation where computers are reading thousands of nearly-identical resumes written by other computers, and arbitrarily selecting based off of unknowable criteria. Incredible amounts of computing power being spent on both sides, canceling each other out, reducing the process back to, essentially, random chance. Once the initial crowd has been filtered out, the original human hiring managers step in and start doing the actual work.

At the end, both parties are left with less information than they would have otherwise had: The resumes have been fine-tuned beyond recognition, conveying no useful information about the applicants or their capabilities, and the applicants have no sense of what they could or should have done differently.

Not long ago, a very silly website called Moltbook received a lot of very silly press, purporting to be a “reddit-style social network” for AI agents. The idea was that you’d just let these robots talk to each other, and see what happens. Breathless, pearl-clutching, financially biased evangelists pointed to anecdotal conversations where they seemed to be having discussions, planning, scheming, etc. It looked, to them, like evidence for the future potential (and future dangers) of AI.

This didn’t amount to much. People willing to assign some further scrutiny to the website found aimless, meaningless conversations, and human vandals playing pranks. In addition, the vibe-coded application itself was found to be horribly insecure, and was easily hacked by security researchers.

Though the idea of AI talking to AI is compelling from a science-fiction perspective, there’s no real meat there. Functionally, it’s not any different than an AI talking to itself, or an AI responding to a human. Though we can build scaffolding that imitates this behavior, LLMs don’t learn in the same way a person learns. The end result of two AIs talking to each other isn’t more content, it’s less. A ping-pong back-and-forth erosion of meaning, nuance, and intention, reducing the information in the middle to the most generic, hollow version of itself. Humans using AI to cater information to other AIs, shooting as hard as they can for a dead-center smoothed-out average.

Song of the Day:

An unusually well-photographed clip, shot on grainy 16mm film. Parquet Courts, for this tour, teamed up with NYC weirdos PC Worship as some kind of not-so-super supergroup. The Television-worshipping indie stars bounce off of the artsier, jazzier PC Worship to create a brand-new cover of a 40-year-old jam. Dissonant vocals, empty space, pick-scrapes, saxophone blasts, culminating in a plodding, atmospheric groove. Bob your head along with them, join in.

Job Search Day 1

Stumbling out of the mind-scrambling pressure cooker of combined full-time-work and full-time-school and into a very complicated job market.

Today’s very minor goal was to update my personal website and implement this blog. I built this page with Astro last year, and then immediately began to enthusiastically neglect it.

Reasonably and logically, I know computers, networks, apps, devices, etc. are so massively, thoroughly integrated into our daily lives, that even if it’s a particularly difficult moment to break into the industry, there’s always going to be a need for software engineers. That said, it’s no secret that much of the grunt work that used to be the domain of the Junior Engineer can now be handled by AI. This is, for me, financially inconvenient.

A Computer Thought:

Is it normal to get a wall-eyed sense of juvenile satisfaction from hitting brew upgrade? Watching all the crap you’ve installed over the years update all at once? Am I easily entertained by lights and colors? Empty-brained, drooling, watching the numbers tick up, I did briefly wonder how safe any of this is.

I grew up using Windows, and as I understood it, if you wanted to install software, you had to navigate to a webpage, download a definitely-legit .exe file, allow it to do its dark and mysterious business, and hopefully wind up with a totally different .exe file that will run the software, taking advantage of all the unknowable changes that you just made to your dad’s computer. The security risks here are, initially, unambiguous.

When I started using Linux and MacOS, I was surprised by the simplicity of package managers. Where are these packages coming from, and why do we trust them?

Is there anyone making sure all these packages are safe?

SOMETIMES (For apt on linux, YES. For homebrew, YES but less rigorously. For pip and npm, NOT AT ALL.)

If I make a typo, can I accidentally install something dangerous?

YES (On pip and npm, definitely.)

If it was safe to begin with, is it still safe?

NO

Song of The Day:

Real dorky stuff. King Crimson at the beginning of their 80s rebirth. Adrian Belew, freshly poached from the Talking Heads, hopping around like a lunatic, resplendent in his floppy pink suit. Increasingly menacing polyrhythms.