Hacker News Reader

Comment by sgt | Hacker News Reader

We need developers like these who are able to innovate without being tainted by the echo chamber of AI.

LLM's tend to regurgitate known design patterns and familiar UX. Like those typical "keep scrolling down to learn about our app as we show you animations" - gets a bit of.

View on HN

Not long before (or it's already happening) LLMs start training on stuff they wrote previously and it becomes the largest echo chamber the internet has ever seen.

I’m pretty sure this is already part of the training loop even if it isn’t coming from the internet. It is definitely used for fine tuning and distillation. As for how LLM producers avoid model collapse, they curate and filter.

Which could be damaging, or it could create interesting results if it's more like an evolutionary algorithm than entropy. That is, if it can iterate and improve on itself, instead of just take in all information and treat it equally, we'll get something interesting.

For the sake of the IT industry, let's hope it's the latter!

Counterpoint: Swearing off AI doesn't ipso facto make you a good developer, and there are plenty of skilled people who can innovate and use AI at the same time.

To be fair, hero banners lower bounce rates. Nobody would waste the time if it didn't work.

Sure if all you're doing is making it just vomit code then yes.

But there's way more LLMs can do, like assist you in connecting dots in complex codebases, find patterns to refactor given some rules, provide ideas, allow you to dig in dependencies to find apis that are undocumented or edge cases that are subtle to find, find information that you might've had to dig up by endless google queries and hard to find GitHub issues, provide support in technologies you have sometimes to use (sed, awk, regexes, Google sheets apis, etc) but you just don't care enough to learn because it happens to use them few times an year allowing you to focus on what matters, etc, etc.

I'm frankly tired of those pointless debates conflating LLMs in the context of code just for the same boring arguments of people hating on vibecoding or thinking every developer is delegating everything to AI and pushing slop. If that's your colleagues fire them. They are indeed useless and can be replaced by a prompt.

If one cannot see the infinite uses of LLMs even just for supporting, without ever authoring a single line of code or never ever touching a file, it's only been that person limiting it's own growth and productivity.

Seriously, this is beyond tiring.

You hit on something key here: the vast majority of the pro-articles are talking in the context of code vomit or content vomit or other types of vomit. As you say, it is beyond tiring, so we respond. I'm still waiting for that cancer cure!

> it's only been that person limiting it's own growth and productivity.

Maybe limiting raw productivity, but I sure don't buy that it limits growth. Maybe if all you ever did was copy and paste off of SO, but taking the time to study and deeply understand something is going to be much better for your overall growth. Also collaborating with humans instead of robots is always nice.

You can study better and more effectively with more tools at your disposal ;)

I've never learned and collaborated as much as in the last years.

It's all situational, I guess. For example I never learned or collaborated more (or had more fun at work) than when I worked at an XP shop, pairing every single day. I feel the majority of people are just going to have coworkers who have fully stopped asking asking questions. But you aren't wrong. I use AI a bit as a search engine.

You can still do XP, can't you!

I do agree that real people sharing is absolutely crucial, and that AI may have a negative effect it it becomes the main and only way of (non) sharing.

I could if there were a wealth of companies that offered it ;) Not only that, if there was a wealth of people who wanted to do it. It's unsurprising that most people are more willing to collaborate with a machine than a human (if they are even willing at all), but there's obviously nothing I can do about that.

> I do agree that real people sharing is absolutely crucial, and that AI may have a negative effect it it becomes the main and only way of (non) sharing.

Yes, exactly. Most of the pro-AI articles that show up here are about generative AI which is what the vast majority of people are raging against. This particular nuance is often left out. Although the TFA's About page explicitly states that when they use "AI" they mean "Generative AI." Obviously pretty much all of us having been using "AI" in some form or another for the past 20... 30+ years?

But then there are a whole host of other reasons I don't like "AI" that are a little out of scope of this rant.

The cheerleaders, and the programming vegans are exhausting. LLMs are not taking over programming, but neither are they useless. They're tools with strengths and weaknesses we're still learning. I just want a take on them for once that doesn't have some absolutist axe to grind.

AI will beat humans at all tasks that are not subjective (such as a landing page being pretty), but instead can be determined to be correct or not (does an endpoint return the correct data? how fast?)

Just like a chess engine beats any human.

People think LLMs are still at the point of programming based on what they learned from the data they scraped. We're well past that. We're at the point of heavy reinforcement learning. LLMs will train on billions of LoC of synthetic code they generate, like chess engines train on self-play games.

Chess has a very specific win condition by which moves can be assessed. Many real world problems are much fuzzier than that and don't reduce neatly to algorithmic validation of 'correctness'.

That's the breakthrough of ML, it can handle fuzzy. And chess is in some ways similar. Outside of endgame and blunders (where you can just bruteforce), you can't prove one move is superior to another. That's why chess engines used to have human made heuristics.

Reply on news.ycombinator.com