PHP code generated by GPT-2

Still better than Wordpress core.

The tweets that go with this: <a href="https:&#x2F;&#x2F;twitter.com&#x2F;moyix&#x2F;status&#x2F;1096255984866082816" rel="nofollow">https:&#x2F;&#x2F;twitter.com&#x2F;moyix&#x2F;status&#x2F;1096255984866082816</a>

Any context on what this is supposed to be...? I can vaguely read PHP, but the code does not appear to be doing anything of much substance.<p>At first I thought it was something to do with a second revision of GUID Partition Tables.

I have been working with a group that is trying to clone this dataset and make it publicly available (<a href="https:&#x2F;&#x2F;github.com&#x2F;jcpeterson&#x2F;openwebtext" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;jcpeterson&#x2F;openwebtext</a>), and I have noticed quite a bit of code in the scraped dataset. Future releases of our dataset will be pre-filtered with another LSTM language model that will filter sentences by their probability under more conversational &#x2F; literary datasets.

It will be interesting when AI finds out there&#x27;s no need for her to generate human-readeable code.

With some automated formatting: <a href="https:&#x2F;&#x2F;pastebin.com&#x2F;7F2Leqy1" rel="nofollow">https:&#x2F;&#x2F;pastebin.com&#x2F;7F2Leqy1</a>

Just what I didn&#x27;t need to see this morning.<p>I am <i>literally</i> living in the streets, freezing my ass off and hungry, looking for any kind of programming work for the past month, and now I have to see some AI bot generating more inexpensive shit code that I am sure some manager will convince themselves might get them that final career promotion by lowering their labor costs to near zero.<p>WTG, geniuses, for developing AI that before you know will have all of us living in the streets and hungry...<p>I&#x27;ll save you a spot.

This is very fishy. You can get code like this by substituting words in identifier names for other words, but how can an algorithm trained on English dataset &quot;learn&quot; that keywords like &#x27;function&#x27; and &#x27;class&#x27; are exempt from substitution? I know most people here have unwavering faith in the magic of deep neural networks, but you&#x27;d need _a lot_ of examples to deduce this with any certainty, regardless of how you do it.

Motivated by this post, I decided to test it out. It&#x27;s impressive how powerful the software is even with the limitations. I made a simple tutorial on how to test GPT-2 out for yourself at <a href="https:&#x2F;&#x2F;medium.com&#x2F;heroic-com&#x2F;how-to-quickly-generate-full-articles-using-openais-gpt-2-3876870aeb5c" rel="nofollow">https:&#x2F;&#x2F;medium.com&#x2F;heroic-com&#x2F;how-to-quickly-generate-full-a...</a>

Important note: The AI did not generate that exact version of the code. It was <i>almost</i> syntactically correct. Here&#x27;s the diff:<p><a href="https:&#x2F;&#x2F;gist.github.com&#x2F;moyix&#x2F;dda9c3180198fcb68ad64c3e6bc7afbc&#x2F;revisions#diff-8074138f091ac83e2bef3faec88bdb05" rel="nofollow">https:&#x2F;&#x2F;gist.github.com&#x2F;moyix&#x2F;dda9c3180198fcb68ad64c3e6bc7af...</a>

Can someone shed more light behind this? What is the true source? Was it generated via the unreleased full model by an OpenAI employee? Or did someone generate it with the released &quot;smaller model&quot;? Can we, the curious public, see the model and replicate the results?

This makes me think that something like Stack Overflow could be used to train a model that generates code to answer a question—and that software specifications that are decomposed into a series of requirements or &quot;questions&quot; could be fed into this model to produce code that&#x27;s equivalent to a team of remote contractors.<p>Your model would be based on NLP&#x2F;votes of the questions, NLP&#x2F;votes of the answers, and separating the text from the code in both.<p>The fact that many markdown&#x2F;code formatting tools have you select the language for syntax highlighting is useful for classifying code as well.

Look ma, no Laravel

BTW, I would like to point you to the MIT&#x27;s project Genesis as an example of what a rule-based text comprehension system could do almost a decade ago.