www.newscientist.com /article/2342632-artificial-intelligence-is-being-asked-to-predict-the-future-of-ai/

Artificial intelligence is being asked to predict the future of AI

Chris Stokel-Walker 4-5 minutes 10/17/2022

Artificial intelligence model predictions from historical data on how AI research would develop over five years matched reality with more than 99 per cent accuracy – soon they will be asked what comes next

Technology 17 October 2022
AI face

What does the future of AI hold?

Yuichiro Chino/Getty Images

Artificial intelligence models are being used to try to predict the future of artificial intelligence research. Thankfully, none of them say we are due an AI apocalypse.

Mario Krenn at the Max Planck Institute for the Science of Light in Erlangen, Germany, and his colleagues trained an AI model to analyse 143,000 papers published on the arXiv preprint server between 1994 and 2021. All the papers covered areas of interest to AI. From that list, they used a natural language processing tool to …

generate a list of nearly 65,000 key concepts by stripping keywords and phrases out of paper titles and abstracts.

Advertisement

Those concepts formed the nodes of a semantic network, allowing the AI to spot connections between ideas and papers. The data informed the AI model about how research areas shifted over time and how academics drew connections and probed new fields of interest historically.

Ten other AI machine learning methods then used that network to try to work out which concepts that hadn’t been studied together would be within five years.

Tested on historical data, AIs were able to predict which then-uninvestigated concepts would appear in at least three papers within five years with more than 99.5 per cent accuracy, which the researchers claim shows a quasi-determinist pattern in AI research.

The researchers suggest that the approach could be used to predict future hot topics or help develop AIs with human-level comprehension.

It is a broad claim that actually turns out to be something much simpler, says Gabriel Pereira at the London School of Economics. “They are only forecasting what concepts may come next to each other by looking at previous articles.”

The model is able to identify the rough direction of travel, says Pereira, but has no creative insight into the future itself: it is simply reflecting back trends in large-scale data that would take an impractically long time for humans to do.

“I think this paper also reflects very much the way of thinking present in computer science and the AI field,” says Pereira. “Wherever possible, let’s add more AI, no matter the cost, no matter the need.”

Reference: arxiv.org/abs/2210.00881

More on these topics: