Nathan Lambert
speaker
1,814 appearances
3 recordings
2 series
first heard Feb 2025
last heard 1 Feb
Nathan Lambert’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 — 2 in all, peaking in Feb 2026 with 1.
Appearances
And I think that...
my writing, I tried to do this as the writing, which makes it come across as raw, but also high information in a way that it's like some people will get it and some won't.
And that's kind of the nature of research.
And I think this is something that language models don't do well.
Particularly, they're all trained with this reinforcement learning from human feedback, which is designed to take feedback from a lot of people and in a way average how the model behaves from this.
And I think that it's going to be hard for a model to be very incisive when there's that sort of filter in it.
And I think this is kind of a
wonderful fundamental problem for researchers in RLHF is like this provides so much utility in making the models better but also the problem formulation is kind of like there's this knot in it that you can't get past so that's what I think of as like
these language models don't have this prior and their deep expression that they're trying to get at.
I don't think it's impossible to do.
I think there's stories of models that really shock people.
Like I think of like, I would love to have tried being Sydney and does like, does that have more voice?
Cause it would so often go off the rails on people.
And what is historically, obviously a scary way, like telling a reporter to leave his wife is a crazy model to potentially put in general, in general,
But that's kind of like a tradeoff, like is this RLHF process like in some ways adding limitations?
There was a lot of backlash last year with the GPT-4.0 getting removed.
And I personally never used the model.
But I've talked to people at OpenAI where they're to the point where they get emails from users that might be detecting subtle differences in the deployments in the middle of the night.
And they email them and they're like, my friend is different.
And they find these employees' emails and send them things because they're so attached to this model.
Showing 401–420 of 1,814 · page 21 of 91
← Previous
Next →