261–270 of 812 posts
Yeah, when your code runs on more than one machine, more than once. Then it's worth it to spend some time to make it actually good. RAM usage also matters.
I call these AI tools "proprietary non-determenistic database of the free internet". They belong to american companies which can cut off your access if american government doesn't like your country's government. They fed from the free internet that many of us grew up in, store it in humans unreadable form and sell you access to it. If some day claude starts to spit out compiled binaries instead of code nobody will no…
Deepseek v4 flash is by any means comparable to SOTA from 6 months ago. It's more than good enough for AI-assisted coding and there are no reasons to believe that one year from now or so, they won't be even better and faster.
People have reasonable concerns about the ethical, political, economic dimensions before we even get to the technical capabilities.
Even within the narrower question of the technical utility, I think there are a lot of factors which differentiate people's different experiences which are largely unacknowledged and lead to people talking past each other and failing to understand how others have such different experiences and opinions.
The sentiment that "users only care if it works" for example implies that all considerations beyond "does the feature work today" are developers self serving their aesthetics, but overlooks many other concerns which _do_ impact users at a later point.
I wrote about just this division of experience and the polarisation which manifests on hacker news just last week as it happens https://www.julianhaeger.com/posts/Disparate-Results-with-AI...
Earlier quoted context omitted.
While I agree that it's "divided", I wouldn't say "simply". Mentioning AI brings out a sharply negative side of HN that I had not seen before 2023. It is the only subject where, when I have shared that I built something with it, I have gotten derogatory comments claiming I am inexperienced, unintelligent, and that the thing I built (a hobby project) is unimpressive or embarrassing. This has never happened in the deca…
I would say that HN was at least as sharply negative during the cryptocurrency craze. I recall various submissions asking "Why is HN so anti-crypto?" as well.
It wasn't just that crypto was an obvious grift; it was that you didn't need to be an experienced developer to confirm that 99% of the "web 3.0" nonsense that what was being thrown around literally made no technical sense.
You might reject LLMs on principle, or find that they don't work for you. But I think we're well past any debate of whether they do anything at all, which is exactly where crypto was sitting at peak hype.
I’m more of an “AI centrist”, as I think the topic is extremely nuanced. As with most tech hype, there tends to be a black and white “AI good” or “AI bad”. I think reality is somewhere in the middle, personally. > Let’s face it: by the time I manually ship version 1.0 of a product, the AI-assisted version could have been deployed 10x faster. By then, enough real-world feedback would have surfaced to identify the majo…
If you have an application that real people use in everyday work and depend on stuff not moving under their feet. Execution speed is not something you aim for because you have to adjust to speed of users adopting changes.
As a startup owner of course you can’t really care about those people using your software because you have to pivot as quickly as possible to an investment that pays the most and boring every day app is not going to be a unicorn.
There is another option at the end of the spectrum. You aim at posthuman world where humans are not required and your software doesn’t have to bother with earning money or users that are slow or have emotions.
Earlier quoted context omitted.
You are going to be asked to do your job, faster, and you are going to have a massive headache managing a bunch of machines mostly doing your job.
That is the status quo now . If LLMs get better and better, managers will simply direct them, well, directly. That's what the parent comment means by the machine learning to do your job. It completely , not mostly , replaces you.
I call these AI tools "proprietary non-determenistic database of the free internet". They belong to american companies which can cut off your access if american government doesn't like your country's government. They fed from the free internet that many of us grew up in, store it in humans unreadable form and sell you access to it. If some day claude starts to spit out compiled binaries instead of code nobody will no…
Open weight models remove most of the issues you list and require relatively affordable hardware like a MBP with 128gb of ram or even less. Deepseek v4 flash is by any means comparable to SOTA from 6 months ago. It's more than good enough for AI-assisted coding and there are no reasons to believe that one year from now or so, they won't be even better and faster.
> Users don’t care whether the code was written by AI or by hand That's true, but they care deeply about the consequences of that: > about how AI “writes bad code,” “introduces bugs,” “creates technical debt,” or something along those lines. So whomever your strawman is, they got a point. Note that I'm "anti-ai", I use it a fair bit and even received the trendy email asking me to watch out how much I spend in it caus…
It doesn’t feel that way to me. I remember reading the recent thread about Bun being rewritten in Rust with LLMs. Definitely not a 50/50 split of positive and negative comments. It was a pile-on.
Earlier quoted context omitted.
As an aside, the variety of examples in that other thread is impressive. Here are some that I noticed: Fixing my furnace: https://news.ycombinator.com/item?id=48417845 New software for a retro keyboard: https://news.ycombinator.com/item?id=48418158 Customizing my camper van: https://news.ycombinator.com/item?id=48417379 Porting my astronomy app from an old Nokia phone: https://news.ycombinator.com/item?id=48419242 Fi…
I haven't had the patience or mentality to really absorb "ok but how is this useful" until that thread and your highlighted references. Thank you for the curated highlights, however brief it may be, because it's very hard to find such diamonds without dedicating far too much time wading through the abstract gatekept comments on the topic, in most cases. Real world examples give me much hope!
My next project was a 2d game with random terrain, physics, sound, music, multiple levels, a day/night cycle with transitions high score tracking... (not uploaded anywhere, but it works, and I refined it a good bit.). That was more like 8 hours and maybe a 100 prompts.
Here are a few screenshots:
One thing that I have found to make a pretty big difference is using both the latest models and higher thinking levels. Opus 4.8 with thinking on Extra or even Max is genuinely mind blowing. The thing I hadn't really appreciated, having a sort of naive impression formed mainly from using free early versions of stuff like ChatGPT and Stable Diffusion was sort of that "Type a big ass prompt and it craps out a result" experience. But Claude is really great at refining from feedback, and it's way more flexible and responsive than I would have ever expected. I can do something like take a screenshot of a small portion of the running app or website or whatever and just say "This button needs to be bigger" or "make this red" or something like that, or even sometimes just "fix this", and Claude both correctly identifies what I'm talking about, and actually does the thing.
here I've found it really, incredibly game changing is my health. I have a pretty, to put it mildly, complex medical profile at this point. I haven't worked in over a year and pretty much every sign is pointing towards permanent disability at this point. Tons of symptoms, long med list, and I live in a smaller town with not great access to care. I'm also autistic and have not the greatest verbal communication, especially under stress or time pressure. I dumped all my info at it, in bits and bobs over several days (Side note... it's memory is pretty limited, but it will quite happily right out everything it knows from a session into a markdown file it can later re-read. I've found it very good for things like screening for drug interactions, or talking through and logging symptoms (and it can log those into human readable markdown files too). Biggest win (other than having unlimited time and interactions) is that it thinks across specilaties, versus the "real world" where the gastro only wants to deal with gastro stuff, neurology only wants to do neuro.
I certainly don't (and wouldn't) use it as a replacement for a doctor, but as an adjunct it's phenomenal. For instance, it flagged a possible drug interaction with a symptom I was having, and then offered to draft a portal message to my GP about it. I have poor executive function so lowering the friction from "type up a message and send it" to "copy and paste" is actually a pretty big deal. Turns something (I probably won't do) later into something I will do now.
It wouldn't surprise me if my very direct, literal, autistic communication style is particularly well suited to interacting with AI. I actually find talking to it rather refreshing as, while of course it's not perfect, it tends to actually respond to what I say rather than the all the assumed subtext NTs tend to expect/react to.
> They care that the product works This reminds me of Anthropic's post where they say they ship 8x as much code as they used to. And I stopped to consider how many times I've used an app and thought, "You know what this needs? More code!"
Unsurprising, agents' solution to everything is writing more code . They'll happily reinvent the universe (a really crappy one). Bug? More code . Unexpected behavior - read the docs? Couldn't find anything . Let's try another 1000 lines of workarounds . Still doesn't work? Write another 1000 lines to monkey-patch behavior . It sort of works now. The actual solution is removing those 2000 lines and passing the correct…
With LLM coding I already have seen agent pointing out easier solutions because agent is not scared of or tired by reading existing code. Whereas most of developers want to write „their code” not read someone’s else’s code.