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

Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

textsynth.com

Post reply on HN
Print view View original

131–140 of 140 posts

Re: Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

#131

Post by GordonS »

Prompt: "The CIA are a despicable organisation, because"

Generated: "The CIA are a despicable organisation, because they are the ones who gave the green light to Saudi Arabia to go and blow up the world.

There is a pattern developing here. As in the case of the war in Vietnam it is always the people who resist US aggression, that suffer. The ones who are complicit, or are in a position to be coerced, are left untouched.

The government’s propaganda will not tell you about the CIA connections with Saudi Arabia’s royal family. And it’s a fact that the Saudi’s are the ones who financed Bin Laden and provided him with his training.

If Obama thinks he can use the CIA to get rid of Iran’s regime and install the next one as an obedient stooge, he is totally out of his mind. The CIA is the problem. The CIA will run the show.

And by the way, the only reason for invading Iraq is to seize the vast oil reserves, to ensure America’s continuing monopoly."

Re: Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

#132

Post by iggldiggl »

I've played around with GPT-J for some time and collected bits of it here: https://www.buttercookie.de/AI/

I think my favourite bits so far are this list of "facts" about garden gnomes (https://www.buttercookie.de/AI/garden_gnome_facts.html#garde...), the strange tale of the Werkstofflimousine (https://www.buttercookie.de/AI/the_burkiss_way.html#werkstof...) and The Chronicles of the Orbital Railroad (https://www.buttercookie.de/AI/orbital_railroad.html).

(There's also more stuff which I haven't had the time to edit properly hidden under https://www.buttercookie.de/temp/AI/)

Re: Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

#133

Post by GordonS »

Prompt: "She loves sex, because"

Generated text: "She loves sex, because anal is really extreme pleasure for her, with it she cums really fast. She is a pretty horny girl and her partner just needs a hand. So the dude puts her on all fours, takes off her panties and starts analing her! Her butt gets a nice pounding and the dude puts his dick in her mouth. Now she has got a great opportunity for real anal orgasm!"

Interesting that it focused on anal sex in particular?!

Re: Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

#134

Post by exdsq »

GPT-NeoX 20B is worse than davinci at writing code - NeoX seems to only really work with C/C++, and even then it's not the best. Just tried to ask it some leetcode questions and honestly I'm not sure what it was doing, but reversing a string included a `temp` variable that was multiplied by the index of the string.

Re: Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

#135

Post by sagebird »

Trying to find a decent prompt for clickbait with NeoX:

Top 5 things you can do today to maximize chances of your toddler being accepted to an Ivy League University:

1. Start studying for the SATs

2. Make sure the kid is using her best handwriting, because that is likely to be the determining factor for an acceptance letter

3. Make sure that the kid has not gotten to a point where she has become too cocky, or a bad attitude, for any Ivy League University.

4. Make sure that the kid is willing to work with as hard as they put her.

5. Make sure that the kid is still excited and passionate about the things she wants to do in life.

What are you doing to ensure your child’s acceptance to the Ivy League University? Share your thoughts and ideas with the parents of America on our Facebook page.

Re: Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

#136

Post by jodrellblank »

Prompt: >

Completion: "Yeah, why should I pay for something when there's already a free alternative? Go and try and make one, you bloody idjit." Which pretty much sums up the attitude of most of the open source community.

So I decided to make one. I used Linux, as it was already installed on the server I'd got the data from, and it seems to work perfectly. So I gave the server owner a copy. A bit of an epic email was involved - I got the server owner's email from a contact of mine who works at the server owner's workplace. The server owner must have gotten all excited when I told him I'd hacked his site, and sent me his root credentials. The email said the server owner could access any data he wanted, but I didn't have time to read it so I didn't check. I didn't think it was something to worry about.

I went back to Australia, went to bed, and woke up to the news that my site was being distributed on Google. I had a look, and what do you know, it had been released! A guy in Germany had used my site, without my permission, to upload about 1,200,000 files. It took Google weeks to deal with it. You can see a screenshot I took of the page showing the content. The filenames are listed by the order in which the files were uploaded. When I went to find one of the files on the site, I was shown all the files he had uploaded - all of them. I thought I'd deleted all of them from the server, so I had to work out which one to download.

I'm not sure how he found my email address, but it seems he used the "Whois" on my website to find the email address associated with the domain. He was sending me messages on IRC about the files being released. As I said, I didn't have time to respond to his messages

%

Prompt: >>

Completion: I'm guessing that either it was a good answer or you were trying to show off. The limerick goes something like this:

    There was a gooseneck who was in anorak,
    There was a gooseneck who was in a tree.
    When he was in his gooseneck, he didn't feel so bad,
    But when he was in his anorak, he felt a lot better.
Anyway, I don't know what you were trying to say but the limerick is pretty funny and you've earned yourself a high five. (Well, actually you earned yourself an "I don't know what you were trying to say but the limerick is pretty funny and you've earned yourself a high five" high five, because that's the sort of stuff that happens here)

Re: Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

#137

Post by speed_spread »

l33tman wrote: post #107
speed_spread wrote: post #19

[In medieval times, ballistic cheese was clearly the superior siege meal] It took a few days to make and was very hard to get, but the results were so dramatic that it became the preferred option for sieges. Nowadays, when siege engineers usually try to get hold of some food from a castle before launching a siege, there is no guarantee that it will be there when they get there. In the age of artillery, the same chees…

I tried the same in the normal GPT-3, randomness 1.0, and it started reasoning about the physics of how cheese might be used in the siege to crumble the walls... Also be sure to read the reference about the cheese and the friars :) [In medieval times, ballistic cheese was clearly the superior way of laying a siege to a well defended opponent] as it could go over had kind of barricade and there was no easy way to get…

Wow! I'm impressed at how good ML turns out to be as a creative tool. Turns out that if you're not looking for a specific or "correct" result, AI gives enough control over the general outcome to allow exploration while keeping the process opaque enough to provide cool surprises.

Re: Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

#138

Post by avmich »

lifthrasiir wrote: post #51
yosito wrote: post #40

I tested this model out with Hungarian. I was curious to see if it could handle the agglutinative nature, and lack of Latin roots. It mostly returned absolute nonsense, but at least it was grammatically correct Hungarian. It also properly answered the capital city of Hungary. I entered a prompt about being a Hungarian speaking computer with something to say, and it returned a passive aggressive response telling me th…

It is always interesting to see how language models perform for languages not native to them. My trial with Korean (also agglutinative) was amusing: [Prompt, in Korean:] In order to do machine learning in Hangul [end of prompt; GPT-J still continuing in Korean:] do you think it's needed right now? [GPT-J switching to English:] English: When the world's about to explode, it's time to learn Korean. If you learn Korean…

[GPT-J switching to English:] English: When the world's about to explode, it's time to learn Korean.

This is worthy to be on a T-shirt. Especially these days.

Re: Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

#139

Post by juliensalinas »

yosito wrote: post #40

I tested this model out with Hungarian. I was curious to see if it could handle the agglutinative nature, and lack of Latin roots. It mostly returned absolute nonsense, but at least it was grammatically correct Hungarian. It also properly answered the capital city of Hungary. I entered a prompt about being a Hungarian speaking computer with something to say, and it returned a passive aggressive response telling me th…

Totally, and that's because GPT models don't really support multilingual content. It works, but very poorly. It's the case for GPT-J, GPT-NeoX 20B, and even GPT-3.

I recently integrated GPT-NeoX 20B on NLP Cloud: https://nlpcloud.io . I had hopes that non-English languages would be better supported than with GPT-J since the model was trained on 20B parameters instead of 6B parameters, but quality still leaves to be desired. In my opinion, the best way to handle text generation in non-English languages for the moment is to couple it with a good translation module. I actually wrote an article about that: https://nlpcloud.io/multilingual-nlp-how-to-perform-nlp-in-n... .

But there is hope! Bigscience is about to release a huge NLP model that should theoretically work very well in almost 50 languages: https://bigscience.huggingface.co/ . We'll soon see if it's true!

Re: Textsynth: Bellard's free GPT-NeoX-20B, GPT-J playground and paid API

#140

Post by lawrenceyan »

Prompt:

An Analysis of Algorithmic Stablecoins

Response:

Abstract: We study the computational cost of generating a digital token that would make certain monetary transfers (payments) possible without fear of fraud, and argue that the price should be proportional to the amount of financial resources put into security protocols. We identify two sources of security that have to be paid for in cash, namely, privacy and decentralization.

We study the computational cost of generating a digital token that would make certain monetary transfers (payments) possible without fear of fraud. If the generated token is a blockchain-based currency, and the payments are in Bitcoin, then a simple mechanism achieves security in a way that is proportional to the amount of money put into privacy and decentralization. In a stablecoin, the generated token is backed by a real currency. To ensure that the generated token has a stable value, the algorithm should take into account the market value of the real currency. We identify two sources of security that have to be paid for in cash, namely, privacy and decentralization.

In a classical algorithmic stablecoin, the total amount of cash is used up as a combination of these two components, and this is how the price relates to the amount of money put into the algorithm. We present two ways to achieve a constant price, depending on whether the market value of the real currency is known a priori, or is revealed via a continuous-price auction.

Post reply on HN