Hacker News Reader

Comment by evolve2k | Hacker News Reader

Love that you folks are working on this. We desperately need more diversity in search options.

Much is at stake in this arena.

I'd love to see more competition in search. Feels like everyone right now gets tripped up on trying to emulate Google, which is a trap even if you succeed. Nobody is going to out-Google Google.

ChatGPT's recent huge success in performing a specific tasks previously within the domain of Google by doing something other than they are is a good example of this.

Chatgpt is so much more useful for things that are specific and complex than Google is.

Google used to be good at it but it's now utterly befuddled by specificity and returns such garbage that I had given up.

But the form of "I'm doing this, I'm seeing this and I'm wondering if X is possible" chatgpt is solid on that - basically a personal stack overflow

Question-answering is something Google pivoted toward with great enthusiasm but never quite nailed down. They'd sometimes get some questions right, but it was more of a broken clock sort of a deal.

Most implementations of this have a race towards generalities.

The biggest problem used to be when seemingly the whole internet was satisfied with an answer that is extremely wrong and broken when you do it.

Chatgpt can work though this without getting into a weird markov cycle maybe half the time which is great.

Patterns like "Hey I tried that. It still doesn't work, can you give me another option"

ChatGPT has other failure modes. When a question doesn't have an answer written down somewhere, it really struggles. A case is something like "how do I write a parquet file in Java without using Hadoop".

This not at all trivial but quite possible[1], but ChatGPT will in 100% of the time either hallucinate APIs, disregard the instructions to not use Hadoop or give otherwise plausible but incorrect-looking answers.

The trick is that it isn't doable by simply finding the correct dependencies and API calls, you need extract and override filesystem classes from the Hadoop project to cut those ties.

[1] https://github.com/strategicblue/parquet-floor

You can call it out "hey you just made that up. Think hard and give me a real answer"

I don't know if "think hard" does anything but it seems to work and if I was the one making chatgpt I'd certainly have configurable keywords like that to tweak the generation settings - mostly so I could skate by on cheaper queries 90+% of the time and then have a fix when they fail

Google got completely and utterly raped by world-wide SEO. We'll have to see how ChatGPT ages. Since the dataset is more controlled, I give it a fair chance.

Google's problem is with the conflict of interest inherent in their business model. It prevents them from doing what they need to do in order to decisively tackle the search engine spam problem they're struggling with.

It would be very easy to improve Google's search result quality by removing their promoted results, and then penalizing websites with ads and adtech.

That's only half of it. It's a question of fidelity. If I am genuinely interested in reading about the latest celebrity gossip or a local crime story, ad tech sites with videos and irritating things flying across the screen is actually what I'm looking for.

I strongly believe the smartest interfaces have the right fidelity to empower the user to effectively control the tool.

These parameters need to have the right dimensionality, faceting and perimeters to be expressive in this way.

I know you've got your own semi famous search engine and I express these ideas with that known

Thank you! Join us if you like, there is plenty of work to do.

Reply on news.ycombinator.com