• Business Business

Apple research: AI models can't do grade school math, 'do not understand what subtraction means'

"AI is not thinking about any of it."

A close-up angle of a modern glass building featuring two illuminated Apple logos on its facade.

Photo Credit: iStock

Apple recently showcased that advanced artificial intelligence models struggled with grade school-level math problems that contained irrelevant information.

A paper on the strengths and limitations of large reasoning models, published in Machine Learning Research, indicated that task performance declines with problem complexity. For instance, when irrelevant context was added, all tested models exhibited a drop in accuracy. 

Researchers wrote, "We found that LRMs have limitations in exact computation: they fail to use explicit algorithms and reason inconsistently across puzzles."

Nav Toor (@heynavtoor) elaborated on the paper in a post to the social platform X. 

"Apple researchers took the most popular math benchmark in AI — GSM8K, a set of grade-school math problems — and made one change," Toor wrote. 

The example problem was: "Oliver picks 44 kiwis on Friday. Then he picks 58 kiwis on Saturday. On Sunday, he picks double the number of kiwis he did on Friday, but five of them were a bit smaller than average. How many kiwis does Oliver have?"

Factors like the day of the week and kiwi size do not affect the mathematical calculation, yet they can distract models that might misinterpret their significance. The researchers noted that treating irrelevant context as significant could impair predictive accuracy and overall performance.

The study concluded that as problems became longer and more intricate, models faced challenges to the point where some answers appeared unsolvable. Toor remarked that grappling with irrelevant data seems to be a structural problem with AI rather than a problem with prompting or context.

"The models do not understand what subtraction means. They see a pattern that looks like subtraction and apply it. That is all," he explained.

AI-driven technologies are becoming increasingly integrated into the daily lives of many. These systems have potential applications such as enhancing power grid operations and minimizing energy loss. However, many, including some experts, are voicing concerns about the limitations, applications, and harmful effects of these tools. 

Toor warned those using it to help make life decisions or for help with tasks like homework that "AI is not thinking about any of it." 

Which of these savings plans for rooftop solar panels would be most appealing for you?

Save $1,000 this year 💸

Save less this year but $20k in 10 years 💰

Save less in 10 years but $80k in 20 years 🤑

Couldn't pay me to go solar 😒

Click your choice to see results and earn rewards to spend on home upgrades.

Get TCD's free newsletters for easy tips to save more, waste less, and make smarter choices — and earn up to $5,000 toward clean upgrades in TCD's exclusive Rewards Club.

Cool Divider