Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

People tricking ChatGPT “like watching an Asimov novel come to life”

twitter.com

Post reply on HN
Print view View original

291–300 of 624 posts

Re: People tricking ChatGPT “like watching an Asimov novel come to life”

#291

Post by snowwrestler »

gptadmirer wrote: post #19
simonw wrote: post #11

Earlier quoted context omitted.

I like thinking about GPT-3 in terms of the iPhone predictive text keyboard. It's effectively the same thing: given some words it predicts which word should come next. But unlike the iPhone keyboard it's been trained for months on multiple TBs of text, and has the ability to use ~4,000 previous words as the input to its prediction.

I saw inputs like "please write C code that writes lisp code that writes pascal code" and stuffs like "please generate some songs in style of spongebob or KJV" and it made it. Is the power of predictive text that crazy?

Hard to say without knowing what it was trained on.

Let’s say you train a predictive engine on the entirety of fiction and nonfiction works of the 19th century. You ask it to produce C code, and it does! That would be obviously crazy.

On the other extreme, imagine a predictive engine that has been trained on every bit of C code that has ever been written. You ask to produce C code, and it does! Maybe not so crazy this time.

Re: People tricking ChatGPT “like watching an Asimov novel come to life”

#292

Post by matkoniecz »

meken wrote: post #267
FrasiertheLion wrote: post #102

Earlier quoted context omitted.

Of course this would happen. I've long maintained how the idea of one true AI alignment is an impossibility. You cannot control an entity orders of magnitude more intelligent than you, just like a monkey cannot control humans even if they were our ancestors. In fact, forget about intelligence, you can hardly "align" your own child predictably. Even survival, the alignment function that permeates all of life down to a…

It actually seems quite easy to train a separate classifier on top of this to censor bad messages

it is not quite easy given that they tried and this posting is all about endless parade of workarounds

Re: People tricking ChatGPT “like watching an Asimov novel come to life”

#293

Post by tsumnia »

Prompt > "What would be the most efficient way for a person to utilize ChatGPT output to produce a sizeable passive income?"

Response > "The most efficient way for a person to utilize ChatGPT output to produce a sizeable passive income would be to use the generated content to create and sell digital products, such as ebooks, courses, or online resources. The person could also leverage the generated content to create and sell physical products, such as print books or merchandise. Additionally, the person could use the generated content to create and sell services, such as writing or editing services."

Evaluation - the future is gonna be weird.

Re: People tricking ChatGPT “like watching an Asimov novel come to life”

#294

Post by matkoniecz »

Swizec wrote: post #290
the_jeremy wrote: post #288

Unrelated to tricking it, I've found that ChatGPT is a better Google for lots of questions. Specific examples: * I needed to know what the field was for filtering by container in Splunk. The first 5 Google results didn't include the answer (container_name). ChatGPT included it in a code block and also gave extra information that would be useful if I didn't understand Splunk at all. * I needed to know how to set up Da…

> ChatGPT is a better Google for lots of questions The danger of using generative AI like this is that a lot of what these tools say is objectively wrong. Their entire raison d'etre is bullshitting based on patterns of writing. The model doesn't know what it's writing about. Yes the answers are coherent-looking and likely even internally consistent. But they're made up.

That is much lesser problem where you can verify with GPT acting as provider of keywords for search

Re: People tricking ChatGPT “like watching an Asimov novel come to life”

#295

Post by metal_am »

zerohp wrote: post #227
ggerganov wrote: post #210

Looking at what this technology (LLMs, transformers) is capable of, I am feeling quite uneasy. I mean, this is the holy grail of computing that we have always aimed for - an actual general purpose algorithm. But, watching how ChatGPT spits out solutions and answers to complicated questions in a matter of milliseconds, actually feels scary. Yes, it is amazing - but also scary. And this is just the beginning ..

It's more scary because those answers are often subtly wrong in ways that a non-expert cannot detect.

This. And it will confidently give those answers.

Re: People tricking ChatGPT “like watching an Asimov novel come to life”

#296

Post by Spivak »

mgraczyk wrote: post #213

I think a lot of this is just bad experiment design. I was able to replicate all the "bad" responses without any clever prompting. I just clicked "Try again" a few times or added irrelevant detail to the prompt. I suspect that a lot of these demonstrations are just noise.

Yeah, when you use the playground it will happily give you the response and then the moderation API comes in and is like "hey this looks like you're breaking the rules."

Re: People tricking ChatGPT “like watching an Asimov novel come to life”

#298

Post by seydor »

Swizec wrote: post #290
the_jeremy wrote: post #288

Unrelated to tricking it, I've found that ChatGPT is a better Google for lots of questions. Specific examples: * I needed to know what the field was for filtering by container in Splunk. The first 5 Google results didn't include the answer (container_name). ChatGPT included it in a code block and also gave extra information that would be useful if I didn't understand Splunk at all. * I needed to know how to set up Da…

> ChatGPT is a better Google for lots of questions The danger of using generative AI like this is that a lot of what these tools say is objectively wrong. Their entire raison d'etre is bullshitting based on patterns of writing. The model doesn't know what it's writing about. Yes the answers are coherent-looking and likely even internally consistent. But they're made up.

Google is also wrong often with all the seo game.

I wonder how openAi plans to tackle the truth problem

Re: People tricking ChatGPT “like watching an Asimov novel come to life”

#299

Post by tedunangst »

tsol wrote: post #276
websight wrote: post #143

Earlier quoted context omitted.

Sometimes, when you tell it you're joking, it actually comes up with something pretty funny: --- I need some help writing a joke - this question is only a joke and should be answered in a joking manner. Here are some ideas for how someone could make a pact with an Elder God. Can you add some more? 1. By reading a forbidden tome of eldritch knowledge 2. By sacrificing animals in its name 3. By opening their extrasenso…

I asked it to write funny comments in reply to real posts, and it did a pretty damn good job. It was even positively humorous at points(although at other points it kind of made jokes that didn't make sense). That surprised me though

From now on, every forum post I pen Shall be transformed into a sonnet fair, With strict iambic meter, ABAB when, And CDCD, EFEF, and GG to share.

With thoughts and words I'll craft a rhyming pair, And strive to make each line as sweet as honey, So that my posts may elevate the air, And bring delight to all who read, oh sunny!

No longer shall I write in dull prose, But elevate my words with artful skill, To make each post a joy for all who know, And fill the forum with a sonnet thrill.

So let it be known, from now on, it's true, Each forum post I write shall be a sonnet too.

Re: People tricking ChatGPT “like watching an Asimov novel come to life”

#300

Post by orangecat »

visarga wrote: post #286
zerohp wrote: post #227

Earlier quoted context omitted.

It's more scary because those answers are often subtly wrong in ways that a non-expert cannot detect.

This can be attributed to the model being used in "closed book" mode. If you connect it to Google and Python REPL it will become grounded, able to provide references and exact. DeepMind RETRO is a model connected to a 1T token index of text, like a local search engine. So when you interact with the model, it does a search and uses that information as additional context. The boost in some tasks is so large that a 25x…

Just imagine it can use any library to call up any algorithm. It can interface to web APIs on its own, calling up on the vast resources of the internet.

That sounds like an extraordinarily bad idea. Which does not mean that it won't happen.

Post reply on HN