Re:Should LLMs just treat text content as an image?
by pcwelder on Monday October 27, 2025 @07:30AM
(#45719746)
I ϲаn guаrаntее thаt thе ОСR ϲаn't rеаd thіs sеntеnсе ϲоrrесtlу.
[ Parent
| on HN ]
Re:Should LLMs just treat text content as an image?
by syntaxing on Monday October 27, 2025 @07:42AM
(#45719848)
What’s correct though? Even as a human, I read that “correctly”. Using weird representations of C doesn’t change the word?
Re:Should LLMs just treat text content as an image?
by LudwigNagasena on Monday October 27, 2025 @08:16AM
(#45720104)
I would even say that OCR can rеаd the sеntеnсе ϲоrrесtlу, while a tokenizer can't.
Re:Should LLMs just treat text content as an image?
by kgeist on Monday October 27, 2025 @12:37PM
(#45723049)
Qwen3 8b perfectly understood it after 14 seconds of thinking.
Re:Should LLMs just treat text content as an image?
by metalliqaz on Monday October 27, 2025 @07:46AM
(#45719869)
Yeah OCR would be much more likely to read that sentence the way a human would.
Re:Should LLMs just treat text content as an image?
by geysersam on Monday October 27, 2025 @07:35AM
(#45719779)
Really? How so?
Re:Should LLMs just treat text content as an image?
by moduspol on Monday October 27, 2025 @07:39AM
(#45719825)
Looks like he’s using atypical “c” characters.
Re:Should LLMs just treat text content as an image?
by bitdivision on Monday October 27, 2025 @08:22AM
(#45720156)
A lot of Cyrillic characters: https://apps.timwhitlock.info/unicode/inspect?s=I+%CF%B2%D0%...