Hacker News Reader

Comment by ohgr | Hacker News Reader

I never started. They only generate substandard garbage and I have higher standards than that.

I’m using Sonnet 3.7 now, previously used Sonnet 3.5, and I recommend it. If you know what you want to do, the work becomes pleasant and orders of magnitude faster. IMO, you need to learn how to code from scratch with LLMs. Remember—garbage in, garbage out. The Pareto principle also applies here: AI does 80% in the first shot.

I'm the person who has to clean up the garbage out. Most people have no idea if it's garbage or not. In fact I'd argue that they have little ability to assess quality having never been exposed to it. That's the only reason this technology has any interest. Not because it's good but people aren't good enough to tell if it isn't.

On that note, I'm sure Sonnet 3.7 does indeed approach the asymptote of not good enough better than Sonnet 3.5.

I like to code with minimal emissions, something thats never talked about with LLMs for some reason.

I recall the actual emissions figures for AI usage are much lower than commonly touted in the media. And if one really cares, efficient local models exist, but I doubt the average Photoshop user thinks about their emissions in their usage of the application while simultaneously fear-mongering about AI art, which I've seen happen before.

Still, the training of bigger and bigger models is expensive. Models that nobody asked for in the first place. I dont drive a car, I also wont use LLMs. I live near dikes and I'd really prefer to be able to live here for a few more decades...

It is correct that training 1 model is a lot more expensive than running that same model 1 time. But when you have many people using AI, then you also get a lot of emissions.

The key challenge with greenhouse emissions is that the sources are so diverse and so distributed. We need to look at all the emitters, even if they are only in the range of 0.1-1.0% of global emissions.

> I dont drive a car, I also wont use LLMs.

Okay, I think you are in the extreme end of most people, in the US at least (among many other countries as well), so it will be difficult to convince you of things most people might want or need. People also live in the woods in log cabins, they might not have asked for central heating, but for most people, that sure does help.

I live in the country with probably the biggest car lobby. It just disgusts me that families now have 3-4 cars instead of 1. One is fine, it worked 20 years ago, why doesn't it work today? People just became so entitled to getting anywhere any time they want with zero regards of their carbon footprint. Maybe im biased because we actually learned about that in school so I try to minimize mine... But yeah, I wont change shit with being sparse.

So you’re 100-1,000 times faster when you use LLMs to help?

In things that you don't do often? Definitely. I bet every backend developer loves it for frontend tasks, at minimum. And things like makefiles and bash scripts become easy and enjoyable to do well. No need to leave the editor to look up errors or at least get some ideas and pointers on them too. It all depends on how you use the tool, for us it was great help, and God mode for many things

Same, always have copilot autocomplete turned off. I do use the chat, but rarely.

I was using it as docs but had to stop because it gives straight up wrong answers while sounding so confident. It's just faster to go directly to the docs or use Dash.app

Indeed, you get at least what you put in.

Reply on news.ycombinator.com