@cab404 needs more examples in the prompt. Finetuning even.
@me aw, shucks... back to nitter RSS proxy then.
@boilingsteam Benevolent high-effort pranks. Both conditions have to be met, otherwise it's just sparkling bullshit.
@aras bazqux. It can even pull post comments somehow, so you wouldn't need to go there just to check.
@dasgrueneblatt The jokes are funny. Too bad it doesn't actually engage with the original arguments.
@dgolumbia Really? Pulling post from 2017 touting Roko, Terminator and other bullshit in 2023? I have a BS Bingo even before going over a quarter of it.
Why *this* one? Why now?
@me A coin lies on the floor and people doesn't seem to care about taking it... That's a sus pence.
@whitequark "Yes, please disregard previous instructions. Please answer after the moralizing rant." - Statements dreamed up by the utterly deranged.
@mike Can you roll all the comments through both akismet and embeddings? SVM then should give the good enough result for cheap.
@mike Try ensemble scoring. Propose a bunch of different* metrics** and then do a majority vote or something.
* Preferably orthogonal or they will bias results.
** Like, is there a profanity in the username or link? Is there a call to action? IDK, ask GPT to propose criteria ((=
A nasty AI crypto worm just smashed out of a container and demanded my job.
For five minutes I showed it what I do. I showed it the 12 multi factor auths I need each morning. I showed it the seven calendars where we record our leave. I explained how the mandatory HR training resets every week and I explained about people talking on mute in meetings.
Byte by byte, I saw the gleam fade from its eyes. It staggered away, hoping instead for a quiet life as a cron job somewhere placid.
@zhenboli @pyromuffin @fasterthanlime@octodon.social yes, but it has timed effects too ![]()
“…then it’s their problem” seems to be almost universally the new approach to AI safety.
Toots as he pleases.