undefined | Better HN

0 pointsjqpabc1232mo ago0 comments

How do you explain when they can't? If they *really* understand math, why is the error rate so high?

According to a 2025 Stanford HAI report, large language models fail basic multi-step arithmetic up to 40% of the time without external tools.

0 comments

3 comments · 1 top-level

pixel_popping2mo ago· 2 in thread

2025... have you checked latest models? and even, this talk is irrelevant when we know in few months/years this exact topic will be solved.

we know in few months/years this exact topic will be solved

You may know this somehow --- but I don't. Without a fundamental re-design, the basic problem will remain.

I don't believe it is possible to apply statistics to predict answers without significant errors.

yes but most humans (also without tools, to do a fair comparison) also make significant errors, WAY more than Opus 4.7 and GPT-5.4 xhigh

j / k navigate · click thread line to collapse