Comment by personjerry | Hacker News Reader

Comment by personjerry | Hacker News Reader

I don't share the enthusiasm. It seems like it's just reading the top results and repacking them into a paragraph, in which case as someone good at google-fu (i.e. every developer) why wouldn't I just go directly to the site with the result I want to ask? Like I'd read an actual ruby tutorial website instead of the Phind version.

Code snippets are fine, but I'd actually rather just use ChatGPT to generate code snippets and ideas, and have it explain to me details and other options in a conversational way.

Thus in general this approach, like most other AI search options, doesn't excite me.

At a high level, I think AI search is terribly hard to solve. You're fighting against existing user patterns heavily tuned to the nuanced behaviours of existing search. So you have to find a specific use case that Google sucks at (and while StackOverflow's fall from grace somewhat lends an opportunity, I think it's not sufficient for most developers to start looking for a different solution) and that you're 10x better at.

But 10x better means your speed has to be just as fast, your answers have to be significantly better, and crucially you have to be nearly 100% consistent, which is brutally hard for most AI approaches.

At that point I wouldn't even market yourself as search. In this context maybe something like "AI-powered software development assistant".

I agree that AI search is hard to solve and then it needs to be 10x better. But I also think there are many Phind searches that are 10x better than reading documentation or visiting Stack Overflow. Not just because it's adapted to exactly the thing you're looking for, but because you can ask followups and narrow it down.

For example, "how to debug pthread race conditions in c++" is asking for a guide where no single comprehensive source exists. Phind combines information from Stack Overflow and other websites to generate a more comprehensive guide: https://phind.com/search?q=how+to+debug+pthread+race+conditi....

Generating guides that aren't simply addressed by docs or individual Stack Overflow answers are where Phind really shines.

In this case, I honestly found the first result (stackoverflow) more useful. It was more comprehensive and direct. Phind told me to use tsan, use mutexes (not a debugging technique), and use a reverse debugger. The first stackoverflow link told me to use tsan among some other very useful things. And the second result was for a reverse debugger.

This is made even worse by the verbosity of Phind's answer; it could've simply given me a sentence about tsan, mutexes and reverse debugging, but it decided on a highschool essay which took as much to read as it took to skim five different methods in some stackoverflow answers.

I really don't think it's there.

So you have to find a specific use case that Google sucks at

One of the things that comes to mind is searching for exact error messages or codes, but for that I want something more like grep rather than AI.

It's also been my experience that for that to be effective, one needs to index GitHub (and GitLab) issues (maybe even commits!) to find reports from others who may have had the same problem. I cannot describe in words how much I want every single project Slack in the world to use linen.dev to stop locking away those folks who use Slack for error reports (I mean, maybe Discord, too, but ...)

That's one of my pet peeves about DDG -- they're indescribably bad about that. I actually haven't tried the same query on Bing to see if it's a Bing index problem, or if DDG just isn't good about it

Reply on news.ycombinator.com