Good catch, not intentional, that's a bug. The sampler picks indices evenly spaced across the account's whole comment history, but it walks newest-to-oldest and bails out as soon as it fills a 90k-char budget. For any reasonably active account, that budget gets used up before the walk reaches old comments, so the model ends up looking almost entirely at recent activity even though the code intends an even sample across time.
Neat tool. I always presumed this existed which is why I really try and speak in generalities and minimize details about myself if I can. What model are you using? How many tokens are we spending running a search?
That's kinda neat, actually.
The summary appears to have a pronounced recency bias. Is that intentional?
Good catch, not intentional, that's a bug. The sampler picks indices evenly spaced across the account's whole comment history, but it walks newest-to-oldest and bails out as soon as it fills a 90k-char budget. For any reasonably active account, that budget gets used up before the walk reaches old comments, so the model ends up looking almost entirely at recent activity even though the code intends an even sample across time.
Neat tool. I always presumed this existed which is why I really try and speak in generalities and minimize details about myself if I can. What model are you using? How many tokens are we spending running a search?
GPT-5-mini, around ~0,0072$ per call
I sort of agree with its assesment. Recency bias is a bit strong.
It did suddenly worsen though. I was checking if it's assesment was deterministic or not.
Imagine what the NSA has on us.
Oops. Just just added to my "Frequent criticism of US military, surveillance, and big tech (many comments)." score.