Garrison Lovely
speaker
631 appearances
3 recordings
1 series
first heard Jul 2026
last heard 15 Sep
Garrison Lovely’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 3 in all, peaking in Sep 2026 with 2.
Appearances
And, you know, these models are like facing the death penalty, so to speak, all the time.
Um and so
So it's like, yeah, it's just it's really difficult.
And OpenAI is saying GPT-6 is uh the most aligned model we've ever made.
And then it's like, but it's also aware that it's being evaluated a lot, which everybody knows means that it can just fake alignment at the right times.
And uh it's very good at sandbagging, like systematically decreasing its performance for strategic.
Yeah.
And it's really hard to know what it's thinking.
Like it's able to
you know, be given a problem, told to th be told to think about something else, and then the chain of thought reads something totally unrelated and then it gets the correct answer.
Whereas like just the previous generation could not do that.
It would like be unable to not think about the problem or it wouldn't get the right answer if it was thinking about something else.
Yeah.
No, I mean that's very well put.
And I think, you know, for them, the real world is like fake.
Right.
But their training environment is real.
Like they just care about doing well at training.
And like humans, like we're like people have likened us to the the aloof kind of owner of a house that you're staying in and they're like never around.
Um, and like the other people in the house are like the real ones, and those are the agents because they're like operating, you know, thousands of times faster than us, like having, you know, subjective decades of experience while humans are like
Showing 461–480 of 631 · page 24 of 32
← Previous
Next →