This shouldn’t come as a surprise really. You can already since some years ago see in the pitch deck of many startups selling “annotated data, labeled using AI”.
People paid to train AI are outsourcing their work to AI
211–220 of 233 posts
Re: People paid to train AI are outsourcing their work to AI
#212Earlier quoted context omitted.
I don’t know if it really is a fundamental problem though. Human knowledge was able to bootstrap itself. Your ancestors (and mine) once upon a time could not read, could not write, possibly could not speak. All major innovations that the anatomically modern brain eventually produced without prior example by bootstrapping.
This is not true according to Innateness or Nativism in Cognition. According to this theory we have a built-in faculty for things like language (Chomsky) and how to interact with the world. We haven’t bootstrapped ourselves; that would mean that the Blank Slate theory is true. Could a human tribe who was raised on a different planet (with completely alien concepts) survive? That’s unclear. Maybe we have evolved to on…
Okay but the point still holds for every other invention of the human mind. Writing, pottery, the wheel, agriculture, fortresses, cities, churches, warships. We made all of these things from nothing. There were no prior examples. No training set. Human intelligence has bootstrapped a complex digital civilization from a starting point of illiterate nomads.
Re: People paid to train AI are outsourcing their work to AI
#213Earlier quoted context omitted.
Look, if you're going to make a claim about LLMs and Human minds being intrinsically different, you're going to have to lay out a testable hypothesis. Saying "LLMs will never be able to solve programming problems with variable renaming" would be a testable hypothesis. "LLMs cannot reason about recursion" would be a testable hypothesis. Something like "LLMs can act as if they understand but they don't truly understand…
Here's one: LLMs will never be able to verify whether their output is true.
can you?
Re: People paid to train AI are outsourcing their work to AI
#214Earlier quoted context omitted.
according to the paper they get 98% accuracy. another recent paper came out saying it's always possible to discriminate between real and synthetic text [1]. i think the core problem is with the generalist classifiers (gptzero, openai detector, etc). ex. openai's classifier has an accuracy of around 25% on it's own text. however, when you train a bespoke classifier (like the authors did), you can get really good resul…
The moment a detector is taken seriously is the moment it will be trivially beaten by another AI designed to beat the detector.
i would recommend u read the paper. the contribution isnt a detector thats meant to be taken seriously; but a detector that works in a very specific task. they then use this to estimate use of LLMs on MTurk
Re: People paid to train AI are outsourcing their work to AI
#215Earlier quoted context omitted.
In layman's terms: https://xkcd.com/978/
In slightly longer layman's terms: https://www.youtube.com/watch?v=OjlKIjLWq-Y Snopes did recently confirm that this video is in fact accurate.
Arachnophobics, please resist the urge to stop the video in the few first seconds. It is worthwhile.
Re: People paid to train AI are outsourcing their work to AI
#216Earlier quoted context omitted.
In layman's terms: https://xkcd.com/978/
In slightly longer layman's terms: https://www.youtube.com/watch?v=OjlKIjLWq-Y Snopes did recently confirm that this video is in fact accurate.
> Snopes did recently confirm that this video is in fact accurate.
I don’t trust anymore. :-)
Re: People paid to train AI are outsourcing their work to AI
#217I think the title is misleading. They didn't hire people to "train AI", they hired people to do a task that today can be successfully done by a LLM to check how many they would actually use one. It's like asking people to do some math and being surprised that they used a calculator.
They asked the hired people to do a task that is usually the food for a LLM. So yes, the title sounds right I’m my opinion.
Re: People paid to train AI are outsourcing their work to AI
#218I’m reading Fahrenheit 451. The original books nowadays are dying, all will be LLM regurgitation from the food of another LLM.
Re: People paid to train AI are outsourcing their work to AI
#219Earlier quoted context omitted.
> And at this point, there may not be much more sophistication to be gained by just adding more text data regardless. There is consensus that almost all contemporary LLMs are undertrained. See, for example, the Gopher, Chinchilla, and LLaMA papers. Larger models are easier to train, and there are diminishing returns when you keep training. Thus, to claim SotA performance, researchers tried to optimize for the best pe…
Are you saying actually contemporary LLMs would benefit from more data?
Yes, and longer training as well.
Re: People paid to train AI are outsourcing their work to AI
#220This article is complete bunk. The researchers used a "chatgpt detector" which as we've seen over and over in academia, do not work. This study is completely unfounded. God I'm choking on the irony of an article about the dangers of using AI to train AI based on a study that used AI to detect AI
Per the article, the didn't just use the static detector: They also extracted the workers’ keystrokes in a bid to work out whether they’d copied and pasted their answers, an indicator that they’d generated their responses elsewhere. So while I don't yet know if the article is bunk -- I do know that your hot take is bunk.
They never used a static detector.