Hacker Newsnew | past | comments | ask | show | jobs | submit | discreteevent's commentslogin

When enraged_camel lies through his teeth like this it's not to try to convince anyone who has experience. It's to catch those who don't. This site used to be a place where developers helped each other out and that's still true to some extent. But for a lot of greedy people on here lately, you are no longer an equal. You are the mark.

It's not a lie in the sense that those individual developers cannot match the KLOC/day output.

bots and useful idiots

Why not call it machine learning? Some people have many different words for snow. That's because snow can have different qualities and the people who came up with those names are intelligent and want to be able to communicate effectively.

Machine learning definitely has a different quality than human intelligence. If someone says "machine learning" I know what they mean and what to expect. So why not have at least one different term to describe it? - It's because people want to give it attributes that it does not have.


>If someone says "machine learning" I know what they mean and what to expect.

I'm not sure how this is. The amount of things machine learning can do now is pretty wild vs. what it could do a year ago. It's not even in the same realm as what it could do 5 years ago.


Interesting that the main thing you can say about your codebase is that it took exactly 36 hours. Not a few days, like most people would say (because they would be focused on the product and not the exact hours). And nothing about what is does or whether there is even one customer that is using it.

I’ve started thinking naturally more in hours of autonomous work in implementation situations. You’re correct in that it’s not a full measure because it took a lot of iterations in design work before kicking off a specific goal pursuit for implementation only succeeded because the LLMs had clear instructions they could work autonomously on incrementally. The end result is still a work product that was produced in 36 hours, so less than a couple of days, and is of high quality and while hard to be precise I’m confident 4 years ago probably would have taken a team of 10 average developers at least a year of development time.

Most of the comments encouraging people to disregard quality are from accounts only a few months old.

FWIIW Quality was always cheaper and quicker for anything but a prototype. This is just as true now. The problem was that many people were too lazy and unfocused to make quality work and those people are only too happy to agree with these comments


There is too many businessmen and too few hackers in here

If it is even that and not just someone's LLM pretending to be one.

Y'know - if I were a LLM company, and I wanted to encourage market penetration of my LLM within the developer community, then targeting developer discussion areas would make sense.

What exactly are you implying with your strawman here?

> Most of the comments encouraging people to disregard quality are from accounts only a few months old.

I noticed that too about a year or so ago. It's unbearable that made me pretty much leave the conversation. I am yet to see a codebase written by LLMs that is half decent. Also, I am unable to generate any decent codebase with LLMs without massive handholding and later refactoring. But maybe that's a skill issue and am holding the genius wrong.

I suspect many people are under AI-psychosis, conflict of interest or were simply introduced to tech very recently.


One thing I learned is to establish as much guidance up front as possible. One shotting something more than a few lines of change isn't sustainable without additional guidance. The best flow for me is to have it generate a draft plan in a file, add copious inline comments, have the same session process them, repeat until plan is detailed and only then give the go signal.

It still goes wrong when there hidden gaps that I don't spot in the plan or when the LLM decides to abandon the plan halfway through (rarer these days). But it's still not as bad as giving a single prompt and hoping for the best.

Of course, all of this only works if you know what you're doing and what you want and can/cannot accept in your codebase.


Sounds reasonable to me. And that’s very different from the parent and grandparent comments that assert they don’t look at any of the code at all and just have agents buzzing around writing and fixing everything autonomously.

Are you suggesting anyone who doesn’t agree with you is a shill? Generally a convenient and comfortable fiction. (There are shills here, when my tin foil hat is on I sense geopolitical ones, but advocacy for AI coding is pretty widespread.)

There's no need to be a shill to have a different perspective on something from the people that do it. Management do usually forget about how programming works even when they did it themselves before - I know because it happened to me.

Businessmen are what ycombinator is about and there must be some of them here. Their motivation is to sell and AI is one of the sauces they can put on their business-burger to make it sell better. Why wouldn't they be enthusiasts?


> intermediate representations to bridge the gap are on the way.

That's interesting. Are you working on these or are you talking about something else?


Yes, I am working on these. Particularly for representational systems - those that model external systems - rather than transformational software.

> if things go well

We didn't need AI for things to go well. We could all have a lot more if people were willing to collaborate and work constructively together instead of sucking rent from everything and tricking people.


Sure, I think that got overshadowed by the western capitalistic approach - it gave people a real possibility to work and get to a more comfortable place in society.

It seems obvious today that following it up to the extreme, optimizing technology and reducing production costs ultimately ends up removing humans from the process.

I wonder if the founders of capitalistic economic model considered this outcome, going to chat with ai about that lol.


Google this: do not delegate understanding

You will find a lot of articles from before LLMs about not delegating understanding of your core system or business. You could say that the software is not your core business so you don't have to understand it. But if the software is any good then it should be taking care of a large part of the business. So software is your business and you shouldn't delegate understanding it.


I like this and the warning it entails, on the flip side of this I have found that you need to be able to delegate and trust. If you demand on understanding everything and doing everything yourself the world will grind you down. It's a big place and sometimes it is OK to delegate what is not part of your core business it let's you focus on this. Not everybody needs to be able to cook 5-star meals, you should know how to heat up some pasta, but don't beat yourself up if you can't bake a cheese soufflé.

braking


Are you saying that a complex stateful system cannot generate a state that is not already recorded in text somewhere?


No. In a large enough state space, you can of course easily find a state that has never existed before. For example, if you pick a 1024-bit integer uniformly at random, it will certainly be different from any previous 1024-bit integer. However, it's much more difficult to find a new integer that cannot be assembled out of bits taken from previous integers. Composition from a handful of base atoms can cover very large state spaces, and unbounded composition can cover infinite state spaces. Which is why only being able to combine things that already exist is not much of a limitation.

To demonstrate:

"Are you saying that a" https://news.ycombinator.com/item?id=49642931

"complex stateful system" https://news.ycombinator.com/item?id=14881439

"cannot generate a" https://news.ycombinator.com/item?id=45956162

"state that is not already" https://news.ycombinator.com/item?id=43013637

"recorded in text" https://news.ycombinator.com/item?id=44731577

"somewhere?" https://news.ycombinator.com/item?id=50391679

A novel state, obtainable by combining six existing ones.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: