Hacker News new | ask | show | jobs
by pixel_popping 59 days ago
And how do you explain that they can do advanced mathematics?
1 comments
jqpabc123 59 days ago
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.

https://medium.com/@dojolabs.main/why-does-ai-get-math-wrong...

link
pixel_popping 59 days ago
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.
link
jqpabc123 59 days ago
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.

link
pixel_popping 59 days ago
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
link
jqpabc123 59 days ago
yes but most humans (also without tools, to do a fair comparison) also make significant error

Humans adopted the use of computers because they provided accurate answers at low cost.

At least until recently. Now, LLMs provide questionable answers at high cost.

link
pixel_popping 59 days ago
but LLMs + Tools (computers) right now beat large majority of humans (and we know this gap will keep widening), so how does that make them "not intelligent"?
link