3 ms·
Right now almost all the training data is 100% organic human-generated content. But, in the future, more and more of the training data is going to be recycled
by mizzao 4y ago
Right now almost all the training data is 100% organic human-generated content.
But, in the future, more and more of the training data is going to be recycled bullshit output from other LLMs. Which is going to cause the model to get worse in accuracy and utility.
Run this forward a few iterations, and most of the Internet is going to be filled with thrice-recycled AI-generated garbage while there are some small pockets of accurate, truthful, curated information.
Reminds me of Herb Simon's "when information becomes abundant, attention becomes the scarce resource."
Having a good BS detector and ability to ask follow-up questions is going to be more and more important for humans over the next decade.
- TechBro8615 4y agoIt leads to an interesting paradox wherein human-written content becomes simultaneously more valuable and worthless. Increasing amounts of human effort becomes worthless because an AI can produce the same result. But at the same time, the AI needs the human-generated content as an input (at least for now), and so it will place a priority value on it. Also, human-generated content will become valuable due to basic economic scarcity: when the world is flooded with AI-generated content, buyers will pay a premium for human-generated content. This implies we will need reliable methods for identifying the provenance of online content, which may spell the end for viable anonymity in many places, or at least accelerate us toward some kind of proof-of-personhood (a hard problem to solve in a privacy-preserving way, and indeed one that leads to the very same questions of what makes us human and gives us an identity in the first place). Short of such a reliable method, the best option for identifying human-written content will be filtering to content that was posted prior to ~2023. In a world of AI-generated content, your shitposts from 2015 may become a valuable commodity.
- amf12 4y agoI doubt that. Current models have been trained using curated datasets and fine tuned by humans. Going forward, I doubt it would be trained on random web data, but only from good sources. Even if the "source" uses text generated by a LLM, it's fine as long as it has been edited by humans - think news sources.
- mizzao 4y agoYou think they actually hand curated all of the 570GB of text data (300 billion words) that was used to train ChatGPT?