When 30 Beats 29
1:06:13–1:07:47 · 94s
Griffiths shares a quirky failure mode: LLMs miscount more when the correct output is a lower-probability token, like preferring 30 over 29 because it's more frequent online.
1:06:13–1:07:47 · 94s
Griffiths shares a quirky failure mode: LLMs miscount more when the correct output is a lower-probability token, like preferring 30 over 29 because it's more frequent online.
We use cookies to understand how you use our platform and to improve your experience. Click "Accept All" to consent, or "Decline non-essential" to opt out of non-essential cookies. Read our Privacy Policy.