Hacker News Reader

Comment by simonw | Hacker News Reader

I've been relatively unimpressed with the ChatGPT browse mode.

My problem with it is that it seems to run really obvious, naive searches. Most of the time I ask it something, then see what it's searching for and think "oh no, that's not going to return anything more useful than what I could have found myself".

It's also pretty slow.

I'm very much looking forward to having a search assistant which can go ahead and wade through eg 20 websites about home battery packs and show me a summary table of my options - ChatGPT Browse isn't quite that yet.

View on HN

Metaphor (https://metaphor.systems) has a ChatGPT plugin that works pretty well. I asked for home battery systems, here's what I got: https://chat.openai.com/share/3ce0687d-6ffa-4c02-8957-d8787a...

Agreed. Its memorized knowledge generally seems to be better than its interpretation of a few web search results, often significantly so.

Analyzing and interpreting new information is harder than regurgitating what one already knows, so perhaps this isn’t surprising.

Even gpt-4, as amazing as it is, has quite limited analyzing, reasoning, and interpretation capabilities, compared to skilled humans.

Of course, the fact that we have to even ask this question is kind of unbelievable…

Have you tried perplexity.ai? It has become my go-to for LLM + browsing. It is really fast and does a pretty good job with retrieval and summarization.

Not in a few months, I should spend more time with that

ChatGPT, Claude and prepexity are each $20/mo

Chatgpt4 seems most powerful but Claude offers unlimited context which can be used to summarize entire books

Prepexity can browse and summarize web

Which one should I subscribe to?

Kagi ultimate is 25 a month and has Claude, GPT4, and AI web search. Been using it a month or so and it's great.

Do you have a source for that? All I can find is here [0], which doesn't mention GPT4 or Claude.

[0]https://help.kagi.com/kagi/plans/ultimate-plan.html

> For Starter and Professional members this mode uses gpt-3.5-turbo, and for Ultimate members it uses gpt-4.

https://help.kagi.com/kagi/ai/assistant.html

Thanks! That is good to know.

I wonder why they don't list that under ultimate plan? On pricing page the selling point for ultimate is like: Hey, help us test things and help us by giving some money: https://kagi.com/pricing

It is in closed beta.

How is Kagi for niche stuff? I'm frequently searching for niche and hard to find programming information about processors and whatnot and google/DDG have only gotten worse with time.

Can also try Phind.com

I use Kagi. Probably not better than Google, but often with less junk. It's much better than DDG. The extra features are nice, as are the customizations (boost specific domains, etc.) I thought I'd be !g-ing to find things on Google, but I rarely do, and I'm constantly researching niche stuff. When I do try !g, I generally find that Google doesn't have the results, either. Kagi also seems to do better with really old results that Google has memory holded.

You can try it for free. That's the only way you'll know if it suits you.

The only bummer is that it doesn't have DALLE.

Bing has Dalle-3 access for free.

Just tried it, the free version is really slow (it has been going on 7 minutes now to just generate the first image and still isn't complete) compared with what I'm getting from OpenAI. I'll keep the $20/mo on that, it is worth it.

Wait what? How did I not know this. I need to consolidate my subs into Kagi. Thanks for the info!

Which one is powering their bot which does citations?

I believe it's Claude 2.

Do you know what model perplexity uses under the hood?

I just tried it out with my normal test queries and found it to be far worse than Google's Bard.

Paid accounts can use the Perplexity LLM, GPT-4, or Claude 2.

To be fair, Bard has massively improved recently and can in my brief experience even outperform 3.5-turbo for lots of things.

Except that it still has a short ctx size.

I'd disagree by far. I use Perplexity as my main search engine and am a paid subscriber. Copilot with GPT-4 or Claude is fantastic.

I love Perplexity.

It’s also pretty useless at understanding any page that’s even remotely complex.

I asked it to tell me when the next baseball home game is so I can avoid traffic when going to the downtown library and it couldn’t answer even that basic question. The search results defaulted to the team’s official calendar but the day numbers were displayed as images so they could style them in the team’s font, making them invisible to ChatGPT.

No alt attributes or does GPT miss those?

ChatGPT seems to be able to parse images pretty well now. I wonder why they don't simply feed it a screencap of the rendered page instead of the html.

Takes much longer to make a screencap of an html page. Modern slow JS and many subresources coming from a slow server, combined with the fact there is no reliable signal to say a page is done loading, generally mean you won't be getting a screencap in less than about 10 seconds.

This example is more likely due to that specific page/site lacking any kind of accessibility information (e.g. how people who are blind, deaf or have other disabilities can use the internet), rather than being large or complex.

Relatedly, in my experience, for some reason most internet properties relating to anything sports related (MLB, NFL, NHL, or even small town little league) have absolutely terrible, extremely over complicated user interfaces.

How come the lack of disability info would make it hard for ChatGPT to interpret the page? I think I missed a step there.

Because on accessible pages there is alt-text for all graphical elements so that screen readers can get the content for visually impaired users.

https://kagi.com/fastgpt?query=What+exactly+is+the+chat+cont...

Kagis' FastGPT is pretty good in that regard (it uses kagi search + claude from anthropic) and it's really fast. Their other AI search/chat which is in beta currently (for ultimate users) is even better IMO.

Have you tried phind.com? It even cites its sources.

Or perplexity.ai which I find better than phind. Faster, looks better, more features.

Perplexity now requires dismissing 2 prompts before use, every single time. (Install their iOS app. Log in with Google.)

Why every successful web app gets plagued by these things?…

Wish it had my use case - go to website foo, copy it's CSS and apply to my framework of choice

Bard is good for quick searches

Reply on news.ycombinator.com