AI ‘inbreeding’ among chatbots like ChatGPT could break them

AI ‘inbreeding’ among chatbots like ChatGPT could break them

Artificial intelligence is here already, it is tipped to own the future, but a new theory puts forward the case that AI “inbreeding” could lead to self-inflicted demise.

From Google to Microsoft, OpenAI, and many smaller startups at the other end of the scale, LLMs and the AI tools they train have changed technology as we know it. AI is here to stay, it is already making a significant impact on our online world and everything that runs through it. 

Further evolution appears inevitable but at what cost, and to what extent? We know already it is extremely expensive to provide the engine room that powers AI, with the server-heavy system just about coping. The advances in AI are remarkable but so is the depth of resources required for the big tech companies to underpin the system.

Irreversible defects in the resulting models

“Model collapse” considers that we might already have witnessed the peak of AI models, presenting the theory that as more and more web content becomes AI-generated, the tech will eventually “inbreed” on existing trained data leading to flaws as human-led data becomes increasingly scarce.

A new paper published in Nature delves further into this concept, which includes the following explanation:

“We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear,” states the abstract. “We refer to this effect as ‘model collapse’ and show that it can occur in [Large Language Models] as well as in variational autoencoders (VAEs) and Gaussian mixture models (GMMs).”

Big tech firms are drawn to the bright, shiny advances of AI, and the spike in shares that it propels. That is not the same as underlying profit which eventually will take its toll, and what of the climate cause? AI is power-intensive, requiring data center expansion and further demands on resources.

The dial is unlikely to be shifted anytime soon by the concept of “model collapse” but it is another interesting question that will be answered further down the line.

Image credit: Via Ideogram

The post “AI ‘inbreeding’ among chatbots like ChatGPT could break them” by Graeme Hanna was published on 07/26/2024 by readwrite.com