Live data from Hacker News

Generate Hacker News Comments from Titles

hncynic.leod.org

81–86 of 86 posts

Re: Generate Hacker News Comments from Titles

#81
post #34

I don't get the point. How does this differ from reddit comments?

I know you're joking, but think of the implications if these comments actually didn't differ from Reddit comments. It would then be easy for a small group of people to push agendas on Reddit. Imagine being able to automate creation of posts that push a certain agenda, then being able to automate thousands of coherent comments that agree with the sentiment of the post. You'd be able to manipulate thousands of real peo…

Ahah, the comment I wrote was the result of the algorithm on its own title.. Guess it's working great :-)

Re: Generate Hacker News Comments from Titles

#82
Title: What is the worst language?

hncynic 1 minute ago

> The most difficult languages to learn are C, Java, C++, and C, which in their implementation would have helped prevent them from becoming "worse" languages

Nope, but still more challenging than "easier".

> Lisp

No, because that's not so hard...

> Ruby, Python, and even Haskell

Nope, but Lisp is the least painful language to learn.

> PHP

Lets say I want to program on a Mac. I want to program on a Mac. The only reason I need to do that is that the code base sucks and you can't use PHP.

hncynic 1 minute ago

I don't think this guy is wrong. I think that's the best language to write web applications. I personally have spent a fair amount of time building web applications. I think people have a hard time focusing on web apps. If you can use Java or PHP or Python, you can do web applications too.

Re: Generate Hacker News Comments from Titles

#83

NN-generated text has this really weird uncanny valley feeling to it, where the words are well organized but there’s no cogent thesis or point to be made. I feel like I might be reading schizophrenic ramblings. No point here necessarily, this stuff is fascinating but creeps me out. Words from generated comments (using this website) are organized in a way that cause me to second guess my own reading comprehension abil…

second paragraph sounds like an hncynic comment with "comment" in the title.

Re: Generate Hacker News Comments from Titles

#84
post #75

Earlier quoted context omitted.

> Who has ever looked at a reddit profile and decided that an account should be trusted based on karma? For starters, automated subreddit spam protection algorithms. You may have misinterpreted me, think of it like StackOverflow karma. Your account needs a certain amount before it can do certain things.

Yes, maybe like +200 (which is what I've set it at as a mod in severe cases of brigading), not thousands. Most of the subreddits don't require anything but positive karma, and you can achieve that with less than a dozen manually created bots upvoting the comments that new bots make. There are no subreddits with a spam algorithm counted in the thousands of karma, that would prevent them from growing.

> with less than a dozen manually created bots upvoting the comments that new bots make.

As I said in my other comment, these self-upvoting schemes are very easy to detect and are not a scalable way to create tens of thousands of bots.

Re: Generate Hacker News Comments from Titles

#85
post #33

Good work, the github repository is also clean and easy to follow. May I ask : What is the size of the live model of the website, in MByte ? (or maybe it would be nice to add that info in the README, just to have an idea) Also, would be interesting to see how it compares to the recent GPT-2 ( https://github.com/openai/gpt-2 )

Thank you!

The model weighs in at 1.2GB with 100M parameters, which is similar to the smallest GPT-2 model.

I wouldn't be suprised if GPT-2 small (+ finetuning on HN data) performed better than what I have trained. Other than hyperparameters, I think there are two main differences: First, I pretrained the model solely on Wikipedia data, while GPT-2 used more general web data. Second, I used an encoder-decoder model, while GPT-2 is a language model. I'm suspecting that the encoder is not very useful for this task.

Post reply on HN