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

Show HN: Talk to Transformer – Generate Text with OpenAI's Latest GPT-2 Model

talktotransformer.com

Post reply on HN
Print view View original

11–20 of 93 posts

Re: Show HN: Talk to Transformer – Generate Text with OpenAI's Latest GPT-2 Model

#11

Post by minimaxir »

gwern wrote: post #3
minimaxir wrote: post #2

What hardware are you using for this?

A quick guess - a standard server instance. It's hosted on what looks like a Digital Ocean IP, and Digital Ocean doesn't have GPU instances AFAIK; and generation is very slow, much slower than I'd expect from GPU. So this might be running on CPU, which would save a ton of money.

It was relatively speedy before it hit the front page.

(also as an aside, I do have my scalable/virtually free GPT-2 API on a CPU ready, should be released this week.)

Re: Show HN: Talk to Transformer – Generate Text with OpenAI's Latest GPT-2 Model

#12

Post by minimaxir »

AdamDKing wrote: post #5
minimaxir wrote: post #2

What hardware are you using for this?

The site was using one K80 GPU but it's slowing down significantly so I'm adding a second GPU. The servers run on Google Kubernetes Engine.

I'm also working on implementing that particular workflow with GPT-2/GKE/GPUs (I'm curious on your deployment strategy if you want to talk more).

You may want to use preemptible GPUs if you aren't already.

Re: Show HN: Talk to Transformer – Generate Text with OpenAI's Latest GPT-2 Model

#14

Post by 1f60c »

I fed it a small snippet of Python code, and I got this back (edited to fix syntax errors):

  import matplotlib.pyplot as plt

  ax = plt.Axes()
  ax.set_xlabel(name='Person ')
  ax.set_ylabel(name='Name ')

  plt.scatter([df.name, ax.x(1/2)).sum(), ax.y(1/2))

  plt.col, ax.rect(name='Name')
  ax.show()
It doesn't make total sense, but it's so cool to see that AI has seemingly learned to use the matplotlib API, save for some unbalanced parentheses :-)

Re: Show HN: Talk to Transformer – Generate Text with OpenAI's Latest GPT-2 Model

#15

Post by minimaxir »

fooqux wrote: post #7

My first test showed that it's surely not perfect. I admit, after all the hype, I expected my SOCKS to be blown off. My prompt: "If Timmy is" Response: "If timmy is iced. then iced = iced + ( " "M " + m ) where iced =" Hmmm.

The default model isn't that robust. Even the outputs on the official GPT-2 page took a few tries.

Re: Show HN: Talk to Transformer – Generate Text with OpenAI's Latest GPT-2 Model

#17

Post by krick »

GPT-2 models' robustness and worst case behaviors are not well-understood. Our findings support the view that this may be a very conservative estimate of the true effect of low-risk paternal circumcision because our model has a relatively narrow definition of risk. The robustness of the model also highlights the possibility that the risk of adverse effects reported is indeed due to the high risk of unintended pregnancy. The observed risk of unintended pregnancies to women in a GPT-2 setting does not approach that observed in the US population. However, an even more likely explanation for our findings would be that unintended consequences that would not be present in the US population would have been lost due to the high prevalence of safe sex, especially condom use, during early childhood.

...Man, wtf.

Re: Show HN: Talk to Transformer – Generate Text with OpenAI's Latest GPT-2 Model

#19

Post by option_greek »

Completing this actionable guide will require significant effort on your end, but just remember: validation dollars are infinitely cheaper than development dollars. And you don’t have to complete this guide in one sitting, so don’t be intimidated. So that's it! This guide gives you a complete overview of building a reusable business logic framework. To get a very comprehensive overview of this topic, see this article . The code for this guide includes unit tests for all of the elements necessary to achieve this goal - it is fully documented, and provides for an API that is straightforward to extend and customize. To learn more about testing, check out this post on TDD-Unit Testing . Feel comfortable using the code when reading the code. This is not a tutorial on how to create a framework, but rather a tutorial on how to build a business logic framework. I'll walk you step by step through building this framework, and demonstrate how it should function in production. You can read the entire code sample here . If you'd like to make additions to or remove code from this framework, then you'll need to write unit tests for the framework. A common problem I face when writing systems is, "what's the best way to add a new element to a single page item?" To answer that question, I will write out a functional imperative programming paradigm called the Agda Pattern. The Agda Pattern is named after the author of the original Agda paper, and as I've noted before , Agda has been used not only by Agda authors, but by other functional languages

Guess where the original text ends and where the model generated text begins :)

Post reply on HN