71–80 of 107 posts
Could we feed gpt-2 Turbo Encabulator? I want more Turbo Encabulator.
[1] http://drusepth.com/series/how-to-speed-up-your-computer-usi...
Cool demo. With these GPT models, I don't get the appeal of creating fake text that at best can pass as real to someone who doesn't understand the topic and context. What's the use case? Generating more believable spam for social media? Anything else? Because there's no real knowledge representation or information extraction going on here.
One use-case is as a creative writing assistant. Typical ways to beat writer's block are to do something else (read, walk, talk, dream). At some point, either consciously or subconsciously, the hope is that these activities will elicit inspiration in the form of an experience or idea that will connect with the central vision of the author's work, allowing the writing to continue. So too it could be with prompts generated from these models, just another way of prompting and filtering ideas from the sensory soup of reality.
(While the "dark" version of reality has bedroom "writers" pooping out entire forests of auto-generated pulp upon ever-jaded readers, being able to instantly mashup the entire literary works of mankind into small contextual prompts would be another tool in the belt for more measured and experienced authors.)
There are other more practical use-cases for these models. There's work being done on auto-generating working or near-working software components from human language descriptions. Personally I'd love to just write functional tests and let the "AI" keep at it until all the tests pass. So seeing the models improve over time is a sign that this may not be an impossible feat.
From a less utilitarian perspective, I'd love computer assistants to have a bit of "personality". "Hey Jeeves, tell me the story about the druid who tapdanced on the moon." and just let the word salad play out in the background. Yeah it's a toy, but it would jazz up the place a bit, add a bit of sass even.
I do think we're a ways off the first computer generated science paper being successfully peer reviewed and contributing something new to human understanding of nature, I'd have scoffed at the idea ten years ago but now I'm sure it's only a matter of time.
We have known for a while that language models can generate superficially good looking text. The real question is whether they can get to actually understand what is being said. As humans don't understand either, the exercise sadly moot.
Even hard mode isn't that hard because GPT-2 tends to ramble on while saying nothing substantive. If I can't figure out what a paper is supposed to be talking about, it's fake. 4/4 on hard. Never read a Nature paper before.
I'm sure GPT2 abstracts would fly through many conferences screening processes. I've seen talks and posters that were utter non-sense but everybody was too polite to say anything to the person or advisors. I've reviewed articles that were completely made up and the other reviewer didnt even detect that. Nor did the editor. I've contacted editors about utterly wrong papers, criticized the article on pubpeer, and the a…
There is a long tail of weak journals in just about any field. When you think about it, this is inevitable in any society that has freedom of the press and where there exist incentives (evaluated by non-experts) for publishing. You have to evaluate journals the same way that you would evaluate products purchased in a flea market.
The side-by-side display makes it pretty easy to distinguish the one from the other, simply compare them at a level where the one that makes the least sense is the one that is nonsense. Like that I score 10/11. But when looking at just the left side one suddenly the problem is much harder, and I'm happy to get better than even. Bits that don't help: not an English native writer. Seen too many real life papers with cr…
Having read your comment first (ooh, horribile dictu on HN) I decided to try playing by only looking at the left paper and deciding if it was fake. Luckily the model seems to have picked up that "last names can be units" too strongly and the 2nd fake paper was discussing a frequency of "10 Jones".
I kept seeing certain types of grammatical error, such as constructs like "... and foo, despite foo, so..." or "with foo, but not foo..." where foo is the exact same word or phase appearing twice in a sentence.
I also kept seeing sentences with two clauses that should have agreed in number or tense but did not.
Pretty easy, even in hard mode, and not due to any knowledge of the subject matter. I'm 15 - 0 so far. I kept seeing certain types of grammatical error, such as constructs like "... and foo, despite foo, so..." or "with foo, but not foo..." where foo is the exact same word or phase appearing twice in a sentence. I also kept seeing sentences with two clauses that should have agreed in number or tense but did not.
It really does like to repeat itself.