Sam Altman

I think no matter how neutral you try to write the thing, it will either be useless because it will just say, I can't answer that because there's politics in everything, or it will have some sort of point of view, which is why what we think we can do is write down what we intend for our default. People can debate that. If there's bugs in there, we can look at the bugs.

I think no matter how neutral you try to write the thing, it will either be useless because it will just say, I can't answer that because there's politics in everything, or it will have some sort of point of view, which is why what we think we can do is write down what we intend for our default. People can debate that. If there's bugs in there, we can look at the bugs.

If there's problems with how we defined it, we can change what the definition is and retrain the system. But, yeah, I don't think any system can be – no two people are ever going to agree that one system is perfectly unbiased. But that's another reason why personalization matters so much.

If there's problems with how we defined it, we can change what the definition is and retrain the system. But, yeah, I don't think any system can be – no two people are ever going to agree that one system is perfectly unbiased. But that's another reason why personalization matters so much.

If there's problems with how we defined it, we can change what the definition is and retrain the system. But, yeah, I don't think any system can be – no two people are ever going to agree that one system is perfectly unbiased. But that's another reason why personalization matters so much.

Yeah.

Yeah.

Yeah.

Actually, here's one thing I've been thinking about recently as a principal. Like OpenAI has not adopted this at all, but this has just been an idea that I think gets at what you're saying. Let's say we discover some new thing where it's like if you do this, people learn way better. If ChatGPT responds always with the Socratic method or whatever, students using it learn way better.

Actually, here's one thing I've been thinking about recently as a principal. Like OpenAI has not adopted this at all, but this has just been an idea that I think gets at what you're saying. Let's say we discover some new thing where it's like if you do this, people learn way better. If ChatGPT responds always with the Socratic method or whatever, students using it learn way better.

Actually, here's one thing I've been thinking about recently as a principal. Like OpenAI has not adopted this at all, but this has just been an idea that I think gets at what you're saying. Let's say we discover some new thing where it's like if you do this, people learn way better. If ChatGPT responds always with the Socratic method or whatever, students using it learn way better.

But let's say user preferences are not to get the Socratic message. Users just say, like, I just want you to answer my question. Tell me. Right. Then, like, how should we decide what to do there as the default behavior? And one idea that I have increasingly been thinking about is... What if we're always just really clear when we make a change to the spec?

But let's say user preferences are not to get the Socratic message. Users just say, like, I just want you to answer my question. Tell me. Right. Then, like, how should we decide what to do there as the default behavior? And one idea that I have increasingly been thinking about is... What if we're always just really clear when we make a change to the spec?

But let's say user preferences are not to get the Socratic message. Users just say, like, I just want you to answer my question. Tell me. Right. Then, like, how should we decide what to do there as the default behavior? And one idea that I have increasingly been thinking about is... What if we're always just really clear when we make a change to the spec?

And so you'll never have our thumb on the scale hiding behind an algorithm, which I think Twitter does all the time, for example, and all sorts of weird things there. We'll always tell you what the intended behavior is. And if we make a change to it, we'll explain why.

And so you'll never have our thumb on the scale hiding behind an algorithm, which I think Twitter does all the time, for example, and all sorts of weird things there. We'll always tell you what the intended behavior is. And if we make a change to it, we'll explain why.

And so you'll never have our thumb on the scale hiding behind an algorithm, which I think Twitter does all the time, for example, and all sorts of weird things there. We'll always tell you what the intended behavior is. And if we make a change to it, we'll explain why.

But if we do discover something like what you just said, or like what I just used as an example, and we say, okay, when people are using it for education, we are going to use the Socratic method because it does seem to have this measurable effect, and here's why we're doing it. We can debate that publicly. Maybe we change our default if you convince us otherwise.

But if we do discover something like what you just said, or like what I just used as an example, and we say, okay, when people are using it for education, we are going to use the Socratic method because it does seem to have this measurable effect, and here's why we're doing it. We can debate that publicly. Maybe we change our default if you convince us otherwise.

But if we do discover something like what you just said, or like what I just used as an example, and we say, okay, when people are using it for education, we are going to use the Socratic method because it does seem to have this measurable effect, and here's why we're doing it. We can debate that publicly. Maybe we change our default if you convince us otherwise.