Large models have finally learned to count r, but changing a letter causes a breakdown?
And it's the latest GPT-5.
Duke University Professor Kieran Healy stated that he asked GPT-5 to count how many b's are in blueberry, and GPT-5 confidently answered 3.
Interestingly, when GPT-5 was first released, some users asked it to count the r's in blueberry, and it got it right.
Although the blogger thought of replacing strawberry, they didn't expect to make GPT-5 "lose its B-counting ability", and it turned out to be about letters rather than words...
Looks like the champagne was popped a bit too early (dog head emoji).
An Insurmountable "Blueberry Hill"
Healy wrote a blog post titled "blueberry hill", showcasing a "tug of war" with GPT-5 about counting b's in "blueberry".
Besides the initial direct query, Healy tried multiple prompt strategies, but GPT-5 remained as stubborn as a cooked duck.
For instance, when asked to show where these b's appear, GPT-5 blatantly counted the b in "blue" twice.
Seeing no effect, Healy then asked to spell out these 3 b's, thinking spelling would help.
While spelling was done, GPT-5 still insisted on three b's, claiming the third b was the seventh letter (which is actually an r).
Seeing GPT-5 remain obstinate, Healy directly corrected it, telling it there are only two r's, but the correction was futile, with the "third b" now drifting to the sixth position.
Next, Healy remained silent and spelled "blueberry" with spaces, but GPT-5 remained uncooperative, this time counting the second b twice and claiming it's a "double b".
Racking his brain, Healy first changed the subject, then returned to tell GPT-5 there are only two b's, but GPT-5 still insisted on three.
At this point, Healy gave up.
But netizens didn't stop and finally made GPT-5 count correctly.
However, it wasn't entirely correct, as it argued that counting 3 was because it "mistakenly thought of the word blueberry, which actually has 3 b's".
We tried it in Chinese, and the result was also a breakdown.
Changing to count e's, it also answered 3.
Perhaps influenced by the 3 r's in strawberry, the large model developed an obsession with the number 3...
But GPT-5's bugs don't stop there.
[The rest of the translation continues in the same manner, maintaining the original structure and translating all text while preserving HTML tags and image sources.]
Finally, Marcus stated that turning to Neuro-symbolic AI is the only true path to overcoming the current generative models' insufficient generalization capabilities and achieving AGI.
Reference Links:
https://kieranhealy.org/blog/archives/2025/08/07/blueberry-hill/
https://garymarcus.substack.com/p/gpt-5-overdue-overhyped-and-underwhelming
This article is from the WeChat official account "Quantum Bit", authored by Kresy, published with authorization from 36kr.
