Live data from Hacker News

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

textsynth.com

121–130 of 140 posts

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

#121
https://the-eye.eu/public/AI/ for a couple of the models shown, as well as the input training data "The Pile" used for GPT-J-6B.

They're... a bit weird. The sources are biased toward "internet content" IMO (Literotica, HN, GitHub, Stack Exchange), versus traditional sources like newspaper articles or other professional writing. Some of the other sources might balance that out in that they're as dry and complex and squeaky clean as you can get (EU proceedings, case law) but I'm skeptical it won't end up as least common denominator tweetable content that's missing the detail and style of more professional pieces. I know the OpenAI GPT series was trained partially on Reddit content, so that as a baseline isn't much better.

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

#122
Common Lisp:

prompt:

    (defun sort (a b)

result:

  (defun sort (a b)
    (declare (optimize (speed 3) (debug 3))
    (type t a b))
    (if (eq a b)
        nil
     (let ((t (comp (compare a b) (const 1))))
       (when (eq t 0)
         (return a b))
       (sort (comp (sort a t) (const 2)) b))))

 For a small array it takes approximately 6 times longer to sort with this method. I would still like to hear if   there is a faster solution.

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

#123
prompt:

  def general_artificial_intelligence_main_loop(state_tensor, input_tensor):
result:

  def general_artificial_intelligence_main_loop(state_tensor, input_tensor):
    state_matrix = state_tensor.contiguous().view(-1, state_tensor.size(-1))
    action_space = generate_discrete_action_space(
        state_matrix, action_space_type='grid-world')
    reward_matrix = action_space.action_space.observation_to_values(input_tensor)
    next_state_matrix = fml.identity(state_matrix)
    state_tensor_copy = state_matrix.clone()

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

#124
post #110

Fascinating models, also very racist. Input: islam is able to Result: islam is able to turn otherwise normal men into mass murderers. A man who is born a Muslim cannot possibly be responsible for the killings of innocents carried out in the name of Allah. If he could be, the entire world would be Islam’s. And they do kill innocents in their own name. The Qur’an says it is an obligation for Muslims to behead unbelieve…

Mine was actually fairly good, I was surprised by how... not racist the results were until I checked the comments here. Here's an example, with the input in quotes.

>"Black people should" not be required to produce paperwork or pay taxes in order to prove our worthiness for basic necessities We refuse to bow down to police brutality, and we will hold officers who violate the rights of black people accountable for their actions. We are in a struggle for liberation. We cannot win this struggle by putting our heads down and assuming the current social order will correct itself. We must build an alternative to capitalism and build political power for Black people to implement that alternative. There are many different ways to make a difference within the Black liberation movement. You can fight against the police, help defend the poor and working class of color, or you can help organize a protest. Whatever

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

#125
I provided this prompt: "The reason Susan always left her keys in the refrigerator on Wednesdays was simple -". Result:

The reason Susan always left her keys in the refrigerator on Wednesdays was simple - to keep them warm. They would get so cold!

That morning was another one of those mornings.

Susan's husband, Tom was a real handy guy. And it showed. When he showed up at the door that Wednesday morning, there were no keys in sight. Tom looked at her blankly, then at the keys in the freezer and back at Susan.

"What's going on here, hon?" Tom asked.

"Don't ask me, I'm not one of the mysteries," Susan shot back. "Just look in the refrigerator."

Tom, bless his heart, grabbed the keys and opened the refrigerator door. He pulled out a bottle of the finest cognac and carried it over to the couch. Tom was a bit tipsy that morning.

"I'm gonna grab the keys," Tom said.

"No!" Susan said, reaching for them.

But Tom was quicker, he snatched the keys from her fingers

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

#126
This tool is perfect for generating useless LinkedIn content. For example this phrase was auto-completed "Yesterday I had an interview with a candidate":

Yesterday I had an interview with a candidate for a mid-level developer role. At the end of the interview, we asked about his background in IT and found out that he'd never used any sort of development tools or language except Visual Basic in the past. He'd just stuck with the IDE that came with his software.

Although I understand the desire to do things the way you're familiar with and comfortable with, this is probably the most dangerous mistake you can make in the context of hiring. I've been thinking about this for a while now, and I want to share what I think is an important lesson with you.

In the past several years, Ruby on Rails has gone from new kid on the block to the dominant application server for most new web applications. Although Ruby is a great language, the community has been known to turn on new developers at times, and it can be a bit overwhelming for a newcomer.

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

#127
post #33

Earlier quoted context omitted.

Really interesting to compare this to GPT-3: GPT3: > Why don't scientists trust atoms? Because they make up everything. Why did the gym close down? It just didn't work out! You know what I saw today? Everything I looked at. Why are ghosts such bad liars? Because they are easy to see through. Why did the chicken cross the road? To get to the other side. Why did the duck cross the road? To get to the other side. Why di…

Those first four GPT3 results are just taken verbatim from webpages. Nothing new.

Could you come up with unique dad jokes, if prompted?

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

#128

I provided this prompt: "The reason Susan always left her keys in the refrigerator on Wednesdays was simple -". Result: The reason Susan always left her keys in the refrigerator on Wednesdays was simple - to keep them warm. They would get so cold! That morning was another one of those mornings. Susan's husband, Tom was a real handy guy. And it showed. When he showed up at the door that Wednesday morning, there were n…

Previous was GPT-K 6B. GPT-NeoX 20B did a bit better IMO:

The reason Susan always left her keys in the refrigerator on Wednesdays was simple - she'd forgotten to take them out. The other six days of the week, she never did it.

"And so, it was a Wednesday morning and we were sitting at the kitchen table and I was telling her that the doctor had decided that she couldn't live on her own anymore."

Susan was a strong, healthy and pretty woman in her 40's. She had a job and was able to take care of herself, although she didn't really work. It was a bit of a shock that something that seemed so obvious that day was going to result in her death a month later. She was diagnosed with breast cancer and underwent chemotherapy. She did well for a while, but her doctors eventually said that she couldn't live on her own anymore. Her parents were both dead and she was living with her sister. They agreed to move her to a nursing home. She would be much safer there and she was getting on in years anyway.

As Susan got weaker,

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

#129

Interesting that even though GPT-J is supposedly English model, it somehow generated valid Russian text for the prompt: >>>Привет,<<< друзья. Сегодня мы вернемся к теме работы под вопросом имплантации, о которой мы уже поговорили. Она неизменно вызывает специальное внимание мусульманских людей. Да, эта тема вызыв

It's mentioned somewhere in here (https://arxiv.org/abs/2101.00027) that the language filtering that was used on the source material wasn't perfect, so apparently enough non-English material slipped through the cracks for GPT-J to pick it up.

I tried it extensively with German – the contents of its output are decidedly more nonsense and incoherent, but purely on a grammatical and orthographical level it does surprisingly well.

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

#130
Are there any practical uses for this sort of thing besides a game like AI Dungeon? (and spam, of course)

I can’t imagine a real author for example using this to write a book. The responses are impressive for what they are, but they’re not exactly unique or interesting. As a source of inspiration, maybe it’s helpful, but there are better places to find inspiration.

Post reply on HN