Hacker News Reader

Comment by miles | Hacker News Reader

[flagged]

According to Script Debugger, hardly the massive powerhouse that may end humanity like GPT-4, it's the "bounds" property of a window, specifically the 3rd element.

In pixels, yes, but that's not trivial to convert to character width...

I'm somehat worried about the incidental harm caused by people relying on GPT's word chain hallucinations as facts for something critical.

Most laypersons don't understand that GPT is not a factual search engine. I'm increasingly seeing people who should know better also forget this now and then.

the problem is that the quality of data on the web is such you can't really trust what you get from google either, so you need to go back to the pre-google days of using specialized search engines, or accept that some of what you get back is likely polluted and you don't have the resources to verify everything.

Nail on the head.

The web is a swamp of hot garbage, in large part thanks to Google SEO incentives.

If Google hadn't poisoned the well, and if searching pulled up good content instead of ads designed to look like search results, and search results that link to web pages that take 5 screenfuls to convey one sentence of information, they wouldn't be looking down the barrel at an existential threat.

with google, at least you could know where and who that piece of information is written..

with gpt, you got nothing but the same convincing tone.

You can ask it to provide links to primary sources as part of your prompt. This helps quite a bit with this problem.

When I try this I seem to get back either non-existent URLs, or links to real documents that don't actually contain what it implies they contain.

It is also important to note that GPT is _not_ useless as a search engine.

BTW, BBEdit has a `hard wrap` scripting command with a limit option (page guide, window width, character width).

And, no, GPT is not a tool for factual verification, it creates probable chains of words that are likely to be found impressive by humans.

Turns out that the strategy of saying something false in order to get the right answer works even better when the false statement comes from an LLM. I guess I'll keep that in mind.

Please also keep in mind that if the corrections fail to materialize, posting false information will have polluted the knowledge pool for no benefit.

Apologies all. Thank you for the information.

> only use BBEdit once or twice a month (usually for multi-file search and replace via regex).

BBEdit’s regular expression support is jaw-dropping. Combined with how quickly it can update north of 10,000 files has kept me using BBedit for over 25 (!) years.

And I'm looking forward the next quarter century.

Okay let's just take GPT's word for it.

I can’t believe this has to be said here of all places. GPT-4 is not a search engine of factual knowledge. You can not assume anything it says is true and not just a made up grouping of words that look nice together.

True, strange that everyone understands this about people, but hold computers to be an authority. Computers had this level of respect even when they spat answers onto a teletype and were called electronic brains whose thinking was represented by spinning tape reels flashing lights and an occasional beep.

Reply on news.ycombinator.com