Please don’t die—heard my wife from a distance.
I was lying on the floor of our garage gym.
Red face, tired eyes, pond of sweat below me.
This is my second week of using an AI coach and doing exactly what it tells me.
I’ve been exercising regularly for ten years and never had a coach. I can run and swim for prolonged periods, as well as do pull-ups and other bodyweight exercises, yet I’ve been going through hell for the past two weeks, and I don’t know what’s going on!
🔑 Key Takeaways
AI coaches? Risky: Great results might hide bad code (or workouts!).
Experience matters: Learn good habits first, then consider AI help.
Generation loss: AI helpers were trained on human-written code. In the future (actually, right now), they’ll be trained on AI written code, which is like photocopying a photocopy.
I never had a coach – AI or human.
But I got free access to an app and thought, why not? It asked me what my goal was: build muscle or lose fat. What kind of equipment do I have? Can I run? Do I have enough room for exercise? Do I have to stay quiet? There were loads of great questions!
After this, the AI coach created the perfect plan for me.
So what’s the problem?
First, let’s look at some research.
![Churn by year Churn by year](https://substackcdn.com/image/fetch/w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff95f7d5c-0eda-4614-8baa-05729500a969_1110x587.png)
If you’ve been reading this newsletter, you know I use AI a lot.
I’m being open about it, and some use cases, like generating boilerplate code fast, have no better alternatives.
So, why is assisted coding bad for you?
If you’ve accumulated enough experience, you can tell if something’s wrong with a code.
Maybe it’s O(n2), but with a slight change, it can be O(n), or it might contain a hidden bug, or it’s just dead wrong.
LLMs can’t tell you any of this.
LLMs give you the most likely answer based on the code they were trained on, and I’m positive that most of the publicly available code (don’t just think libraries or open software, but any code online: Stackoverflow, GitHub issues, and random blogs) is just bad.
I don’t have any numbers to prove this, but I can’t imagine the opposite being true: most of the code online is exemplary.
Pair this with programmers new to programming, and you have a recipe for disaster.
GitHub published several articles on the impact of AI on software development. Copilot users write code 55% faster, which begs the question, is it all good code?
The above charts told you already.
And it’ll worsen because of Generation loss, a phenomenon when you copy the copy.
![25 VHS copies... Every 3 seconds there's a generation loss - YouTube 25 VHS copies... Every 3 seconds there's a generation loss - YouTube](https://substackcdn.com/image/fetch/w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe0d8e4a7-1707-4477-b659-016d43e568bc_320x180.jpeg)
While degradation loss is easy to spot in pictures, it becomes much more demanding with Copilot code because it looks good.
Why is it present a danger for future-generation developers?
It’s the same reason I’m dying on my garage gym’s floor.
Can’t tell if the AI is wrong.
📰 Weekly shoutout
📣 Share
There’s no easier way to help this newsletter grow than by sharing it with the world. If you liked it, found something helpful, or you know someone who knows someone to whom this could be helpful, share it:
🏆 Subscribe
Actually, there’s one easier thing you can do to grow and help grow: subscribe to this newsletter. I’ll keep putting in the work and distilling what I learn/learned as a software engineer/consultant. Simply sign up here:
Valid points Akos. LLM-generated code is often a case of garbage in and garbage out. The same goes for other areas like writing. How can you expect the LLM to write well when most of the text it's trained on is not so great?
It's an interesting problem that hopefully doesn't get worse with time. Despite this, I still think LLMs are great for researching a topic but you've got to have some idea about the topic to ask the right questions.
Also, thanks for the mention!