Hey friends!
Today we start the off-season... and this off-season, I'm starting with a little grouchiness. 😡 Over the past few months, I have observed a statement becoming more and more common that absolutely infuriates me. It goes like this:
I wrote this code with [Claude/Gemini/GPT/Qwen/Whatever].
Now go back and read the title of this post. Let's dig in.
You Didn't Write That
In the purest, most literal sense of the phrase, a language model did the actual 'work'. It took a prompt, ran through a ton of cosine similarities, and assembled something that's statistically related to what was prompted.
Taking credit for work you didn't do used to have a name: plagiarism. And it was considered bad. And I'm not sure why throwing a language model in the middle of an episode of plagiarism suddenly makes it acceptable.
Whoa Blink, Shots Fired 🔫
You might think this is overly harsh. You might accuse me of either "nOt UnDeRsTaNdInG AgEnTiC", or maybe "fishing for social media engagement with hot takes".
Impugning me in these ways would be the easy way out. That would be equivalent to putting your fingers in your ears and tuning out another naysayer.
That's your right. Feel free to ignore me if that's your desire.
But if you're willing, I'd like to outline some thoughts on this topic that feel problematic to me. I welcome reasonable, genuine debate.
Why it's Important To Admit That You Didn't Write This
So you described what you wanted, and the model assembled the actual code for you...
How did it know how to do that?
We've all learned a little bit about this by now: the model was trained... on tons and tons of code gathered from all over the internet. To my mind, this leads to two major red flags:
🚩 How do you know that the code was worth training on?
Computing has always operated on a core principle:
GARBAGE IN, GARBAGE OUT.
Frankly that applies to a lot of things in life too... but we're talking tech, so I'm staying in that lane.
"Because I found it on the Internet" has never been a standard for truth and accuracy. And "this code runs" also doesn't mean that it's safe, or efficient, or even actually solving the problem. Yet with AI tooling, there's no provenance provided for the data. We don't know where they got their information from, but we're expected to put enough trust in it to build production systems with it?
Another facet of this red flag is that AI labs are recursively training now... the outputs of the previous model are being used as the inputs for the next generation. The "learning" in Machine Learning is reinforcement -> we add weight to a correlation that is more frequently expressed. If they're training new AIs on things that AI generated... we'll introduce biases and cause feedback loops.
It's the technological equivalent of inbreeding.
For my friends who are just now getting into building things, I want you to hear my heart on this: I'm not trying to gatekeep. I'm excited that you're excited about building. I love watching you bring your ideas and explore them. That's great stuff. But as a builder, we have responsibilities - to build well. To keep the users of our software safe. If you're not learning how your app is built, if you're just "surrendering to the vibes"... you're endangering everyone with what you build. You need to take this part of the journey seriously.
If you don't learn the craft, you won't know when the bot is producing good, safe, secure, efficient code and when it's hallucinating and setting you up for failure. Without the context, without the experience of building by hand, success and failure may look the same to you.
The folks who have had several doses of the AI Pill will probably try to make an argument here: "You just get one model to review the other model's output and fix it, Blink." Ok, that might be a little safer than just YOLO'ing it into production... but think through this a minute: you're solving an AI-caused problem by adding more AI to it. This is what we used to call a "conflict of interest".
🚩 How do you know it was trained ethically?
Let's assume that all the code found on the internet actually was good code (which it's not... But let's assume it so we can get to the next point).
Was that code obtained legally? Ethically? This summer, we saw the largest class-action lawsuit in history settled by Anthropic with authors whose works were illegally used to train models.
If I told you I stole the money out of the tip jar at a local restaurant and I was going to buy you dinner with it... would you be like "🤷🏽♂️ lol ok" or would you have a moral dilemma?
If you posted an open-source labor of love to your github, and I copied it, removed all your attributions, and sold it back to you... would you feel any differently?
Theft is still theft, even if you run it through a bunch of matrix math.
Another way I'm struggling with the ethics of AI is the concept of wastefulness. Using AI models has already been shown to consume large amounts of electricity, straining our power grids and consuming even more natural resources. Just a few years ago, we were all being told to "offset our carbon footprints"... but now that AI is the new hotness, that talk has all but disappeared. Is it really ethical to exacerbate these infrastructure problems for a chatbot or a silly image generator?
This ethical dilemma bothers me even more because at its roots are questions about how much we value our fellow humans. I've heard it said that you can tell what someone truly values by how they allocate their own resources. If we're unwilling to care for our fellow man but actively dumping massive amounts of resources into AI, there's an implication that we would prefer artificial interaction to genuine human connection. That scares me.
Ok Blink That's Going Too Far... You Don't Write in Assembler, Don't Lecture Me About Using LLMs
A common counterargument I've heard to advocate the use of LLMs for coding goes something like this: "You write your Python code, it goes through an interpreter. I write my .md file, it goes through an LLM. We are the same".
No, we're not. That's not a valid equivalency.
When you write code, there's a direct, deterministic connection between what you wrote and what the interpreter produces in a lower-level language. All of the logical structure, all of the decisionmaking is done IN YOUR BRAIN, and then the interpreter just reorganizes that into machine language.
When you write your .md file, you aren't responsible for thinking through the logic of what happens. Yeah, you're describing the application at a higher level. Yeah, you're "architecting". But just like we talked about in that last red flag, you're abdicating all the small tradeoffs and decisions to the model.
You know what lives in those small tradeoffs and decisions?
Scalability. Performance. 3 AM wake-up calls because the code did something you didn't think about.
I'm not saying that old-school coding always solved those problems. I've lived through enough brutal on-call rotations to know better. But when the outage occurs, you know who we call? The person who knows the code. And if Claude was the author, that's... nobody. Strategically speaking, you've now made your entire success dependent on how many tokens you have left in your plan. Is that really a wise business risk?
So is Blink Just an AI Hater?
Not at all. I use AI tools in my work... often, although with an increasing amount of concern. It's ok to use them for research purposes. It's ok to let them handle small boilerplate kinds of tasks. It's even ok to troubleshoot and debug alongside them.
If you use your AI tooling to help you learn and perfect your craft? I'm all in for that.
If you're just exploiting an army of Claudes to produce more stuff as fast as you can? You've just automated selfishness.
Top comments (0)