3 ms·
> Language models can and are useful if you specifically want to avoid marginality. To reduce noise, to remove errors. But this makes me think of Shannon Infor
by AnimalMuppet 4y ago
> Language models can and are useful if you specifically want to avoid marginality. To reduce noise, to remove errors.
But this makes me think of Shannon Information Theory. What you're describing is something that has no actual information (in the Shannon sense) at all.
And maybe that's why so much GPT output reads so blandly. Even the ones that are not glaringly wrong still read like... like food with no seasoning.