Large language models are fed on words from the Internet. The resulting statistical model of patterns in language can then be used to predict words that would most likely follow a user’s input: type “Who was the king of England in 1653?” and the AI outputs “England had no king in 1653.” However, it is not limited only to short, pub quiz answers. The answers can be long enough – or stitched together to become long enough – to be used as product descriptions, summaries, articles and other longer form contexts. And because many people are now using this capacity to produce website content, there is a rapidly growing set of words on the Internet authored by LLM’s rather than humans. Words which will be part of the training sets for future LLM’s.
And that could be a problem. Why? Some researchers have exposed an issue which they have termed “model collapse”. It can be thought of as being like Ouroboros – the mythical serpent which eats its own tail. What happens is that the ‘tail end’ of the data (rarer words and phrases) in the previous training set vanishes in the new one. This happens because the previous model has been multiplying the appearance of common responses over less likely ones and so the less likely ones become even less likely. As a result they are effectively ignored in the next model. In other words: as LLM-generated text floods the internet, the quantity of data available for training goes up rapidly but the quality goes down rapidly. Ultimately, this becomes catastrophic. Less probable responses vanish and the model’s statistical complexity deteriorates until, eventually, it just spews out nonsense. Rather than LLM’s becoming ever more ‘intelligent’, they become incomprehensible.
Link: AI models collapse when trained on recursively generated data
Thinking about this reminded me of the Tower of Babel (Genesis 11:1-9). This construction was built in the time of the first great human empire. The project united people in a common endeavour which, it was felt, would make them better than just their individual lives. By themselves, they were only human; together, they would be divine. Human intelligence would rise up to become super intelligence. Were they wrong in that notion? In some ways, no:
“And the LORD said, "Behold, they are one people, and they have all one language, and this is only the beginning of what they will do. And nothing that they propose to do will now be impossible for them.” (Genesis 11:6)
However, the pride within their ambitions was exceedingly dangerous. It was them falling, once again, for the serpent’s temptation to become as God by stealing his knowledge rather than learning from him (Genesis 3:5). They would find themselves with a knowledge of evil that they couldn’t resist following down into awful depravity. The humans in this empire would grotesquely pervert their own lives and ruin the earth. Great things would be done; terrible things would be done – the best of times; the worst of times. And God would have to intervene to end it. Yet, he had promised never to flood the earth again (Genesis 9:15). So, instead, God prevented the disaster by bringing about collapse through language disruption. The people found they could no longer communicate easily (Genesis 11:7) and so scattered across the earth into the different cultures we have today. That brought lovely richness to human life. It also restrained our ruinous pride.
Might heaven do the same to our AI empires and their super-human dreams?
Picture by Pieter Bruegel the Elder on Wikipedia
_-_Google_Art_Project_-_edited.jpg){kind=link}
All posts tagged under technology notebook
Introduction to this series of posts
Cover photo by Denley Photography on Unsplash
Scripture quotations are from the ESV® Bible (The Holy Bible, English Standard Version®), © 2001 by Crossway, a publishing ministry of Good News Publishers. Used by permission. All rights reserved. The ESV text may not be quoted in any publication made available to the public by a Creative Commons license. The ESV may not be translated in whole or in part into any other language.