Dylan Patel

I almost think it's practically impossible. Because you effectively have to remove them from the internet.

I almost think it's practically impossible. Because you effectively have to remove them from the internet.

It gets filtered out. So you have quality filters, which are small language models that look at a document and tell you how good is this text? Is it close to a Wikipedia article, which is a good... that we want language models to be able to imitate.

It gets filtered out. So you have quality filters, which are small language models that look at a document and tell you how good is this text? Is it close to a Wikipedia article, which is a good... that we want language models to be able to imitate.

It gets filtered out. So you have quality filters, which are small language models that look at a document and tell you how good is this text? Is it close to a Wikipedia article, which is a good... that we want language models to be able to imitate.

Yes, but is it going to catch wordplay or encoded language?

Yes, but is it going to catch wordplay or encoded language?

Yes, but is it going to catch wordplay or encoded language?

It'll have the ability to express it.

It'll have the ability to express it.

It'll have the ability to express it.

This is what happens. A lot of what is called post-training is a series of techniques to get the model... on rails of a really specific behavior.

This is what happens. A lot of what is called post-training is a series of techniques to get the model... on rails of a really specific behavior.

This is what happens. A lot of what is called post-training is a series of techniques to get the model... on rails of a really specific behavior.

And there's a lot of history here. So we can go through multiple examples and what happened. Llama 2 was a launch that the phrase like too much RLHF or like too much safety was a lot. It's just, that was the whole narrative after Lama2's chat models released. And the examples are sorts of things like you would ask Lama2 chat, how do you kill a Python process?

And there's a lot of history here. So we can go through multiple examples and what happened. Llama 2 was a launch that the phrase like too much RLHF or like too much safety was a lot. It's just, that was the whole narrative after Lama2's chat models released. And the examples are sorts of things like you would ask Lama2 chat, how do you kill a Python process?

And there's a lot of history here. So we can go through multiple examples and what happened. Llama 2 was a launch that the phrase like too much RLHF or like too much safety was a lot. It's just, that was the whole narrative after Lama2's chat models released. And the examples are sorts of things like you would ask Lama2 chat, how do you kill a Python process?

And it would say, I can't talk about killing because that's a bad thing. And anyone that is trying to design an AI model will probably agree that that's just like, eh, model, you messed up a bit on the training there. I don't think they meant to do this, but this was in the model weight. So this is not, you know,

And it would say, I can't talk about killing because that's a bad thing. And anyone that is trying to design an AI model will probably agree that that's just like, eh, model, you messed up a bit on the training there. I don't think they meant to do this, but this was in the model weight. So this is not, you know,

And it would say, I can't talk about killing because that's a bad thing. And anyone that is trying to design an AI model will probably agree that that's just like, eh, model, you messed up a bit on the training there. I don't think they meant to do this, but this was in the model weight. So this is not, you know,