Hacker News Reader

Comment by lb4r | Hacker News Reader

Consequently, this is probably the beginning of a very abrupt end for traditional search.

It's an end to keyword search.

Search now has state. I can search something, and then manipulate the results. Filter the results. I dont have to start over with a new query, or mess with quotes and -signs, which has been utterly terrible ever since google started taking things like that as suggestions. Or when I search for 4-5 words, and it instantly chooses to leave out the most important of the 5 terms so it can return tons of utterly useless off topic stuff.

I absolutely hate this, and it turns the search useless to me. When searching with "direct quotes" on Google, it has ignored my exact term in quotes and shown me results WITHOUT THAT KEYWORD. Useless.

It is utter baffling how often I will search a couple terms and the top 5 results will say missing x and make me click to bring back my most important term. Obviously the term that is most limiting my search result volume is a candidate for either causing interference, or BEING the most important term. The frequency with which google chooses to disregard that word instead of focusing on it is just painful. Why theyve chosen to optimize for returning the MOST results is just misguided. I do not need or care for 10k results. Give me 90 that are actually useful.

This was finally largely fixed last year.

https://blog.google/products/search/how-were-improving-searc...

That doesnt appear to be related. Your post is about making the preview text more relevant.

It helps to turn Verbatim mode on (under Search Tools).

> It's an end to keyword search

I really hope not. I want keyword searching. I hope that it can continue to exist alongside this.

I guess I'm perfectly fine having to press "reset state" to get back to a clean search, vs every search being a clean search.

Most of my searches are vaguely phrased as a question anyway with the question words missing. Im guessing they will work just the same.

Most people ask human-style questions, eg: "How do I turn my toaster on?"; unlike tech enthusiasts who might be more inclined to feed mechanical queries, eg: "toaster turn on -site:reddit.com".

If the former query becomes more useful, even at the cost of the latter (which has gotten fairly useless already anyway), that is an overall boon for end-users.

Only for some kind of searches.

E.g. I'd suspect most people use keyword search when looking for images. They don't do "show me that funny picture of when AI drew the salmon swimming except it misunderstood salmon", they look for "ai generated salmon swimming".

> that is an overall boon for end-users.

But it would be a loss for people doing research.

Can you give an example of something you think will work in standard search but that stateful search would fail to return good results for?

Any searches that aren't looking for an immediate answer. When I'm searching, I don't want the search engine to find an answer for me. I want the search engine to provide me with a list of websites that touch on the topic I'm looking for, so I can get a "big picture" of what's out there, review them, and put together the information I'm looking for from the mosaic that I have been presented.

I don't want the search engine narrow my view to a tiny selection of websites or, worse, try to summarize the contents of the websites. That's not useful to me in terms of what I use search engines for.

Stateful search is really only useful for narrowing things down. That's not what I want the machine to do except in terms of paying attention to the keywords I want (and the keywords I actively don't want).

I totally understand there are use cases where what I actively don't want may be exactly what is desired. That's just not my use case. That's why I'm not saying there shouldn't be such search engines. I'm just saying that I hope my use case will also be possible, or web searching loses a whole lot of usefulness to me.

I would expect stateliness/statefulness to work in both directions. Instruct it to broaden its interpretation.

If you provide stateful search with keywords it will return the list you want. Then just dont follow up. Im probably ok with a couple extra instruction words when looking for traditional looking search results. "list sites re [keywords]"

It'll be interesting to see how the product changes over time when it retains some state about your preferences. Maybe you can teach it a bang abbreviation to always give you the type of result you are looking for. (!lw)

(You already see limited state on google when you search something, then search something unrelated and some results for the former keywords show up highlighted. Google sometimes thinks a new search is a folloup search to the previous one.)

It's going to eat up a lot of Wikipedia, StackOverflow, Quora, etc. traffic but it doesn't change the reality that often you need a real link to somewhere else on the Internet. And all of that is still going through Google or Bing.

Yes, this is a serious concern. These GPT's have stolen the collective knowledge these places have created, particularly Wikipedia - MS will refactor it and spew it out through a Microsoft portal. In effect, ChatGPT becomes a new window into the 'hand' generated Internet - eventually this window will be a new corporate garden wall. Huge swathes of traffic will be pulled from the likes of Wikipedia, StackOverflow, blogs, etc - basically any public facing source of knowledge/information.

It's not good.

If websites don't want their content republished, they should use appropriate licenses restricting such uses.

For example, Wikipedia licenses all encyclopedic text under CC BY-SA 3.0[1] and GFDL[2], both of which permit free-as-in-libre redistribution.

[1]: https://en.wikipedia.org/wiki/Wikipedia:Text_of_the_Creative...

[2]: https://en.wikipedia.org/wiki/Wikipedia:Text_of_the_GNU_Free...

But GPT’s can restructure copyrighted text to make it different enough for it not to be direct theft. It won’t be long till you can get an entire book, feed it in and prompt it to regenerate the book in a different style. Then what happens? Or what about taking books out of copyright, then prompting regeneration in a different style with a slightly modified plot?

If you can prove there was a source and that source was used without appropriate licensing, that is a valid case of copyright infringement and possibly even IP theft.

We really need to stop muddying the water surrounding "AI" and their use of source materials. Those so-called "AIs" are just software like any other program, they are tools like any other computer application, and the copyright and licensing legal precedences that apply to software apply to them.

Obligatory IANAL.

Their ability to transform things (art, text, whatever) at incredible speeds, means chasing copyright is going to be a game of Whac-A-Mole. Also, what percentage of transformation of copyrighted works is still infringement? A GPT can ingest a 500 word piece of copyrighted text and either summarise it down to 50 words or triple the word count in a stylistically unique manner.

Obligatory IANAL.

If you use any copyrighted source material without appropriate licensing (note: Fair Use is a form of licensing) in any way, that's copyright infringement. This is very simple and the legal world has demonstrated that fact time and time again long before computers ever came onto the scene.

Some "AI" drawing a picture derived from copyrighted materials is no different from a game using copyrighted assets from another game. If there is no appropriate licensing, it's copyright infringement.

This is akin to saying that licensing your code under the GPL will prevent GitHub's copilot from recreating it in a proprietary project. I wish it were true, but it's not.

A license won't stop others from stealing your stuff, but you will have legal standing to go after them when they do if you are so inclined.

From what I’ve seen chatgpt is able to generate a list of links for a topic you are discussing. They seemed very relevant to me when I tried this.

Obviously it won’t have anything post 2021 but for some subjects it’s fine.

The article states that they got a demo where they were able to ask questions about the new Bing launch, so the delay is ~hours now, not 2021 anymore (at least, for this new Microsoft thing).

I mean yes, but what stopped ChatGPT2 from doing this for you?

"As a language model, I don't have the knowledge to answer your question 'who won 2024 world cup?'. These are the search result from Bing engine: ..."

Then you never need to actually access google.com or bing.com.

There will still be links, but I'm guessing it will be GPT saying: "Here's a link on the topic: ..." And if you still need more links, there will of course be a way to access the old search, but it won't be front and center.

If I type "wapo" into Google or Bing, I don't want an earfull from a LLM about what "wapo" might mean (especially if streamed slowly a la current ChatGPT), I want a link to Washington Post's current frontpage.

That seems like a trivial matter of fine-tuning the model so that it includes the link on those types of queries. For example, if you search for a website, I doubt it would be hard to have the AI infer that it's a website and include a link to it.

If it works like the other GPT search engines that have popped up, it does provide links.

New era of confidentially incorrect search results.

Reply on news.ycombinator.com