Ep. #3, Exploring Recursive Intelligence with Qizheng Zhang

Got it.

So just at a very high level, the self-improving AI, recursive AI is really top of mind in the overall industry and, and the research community.

And so as you're thinking about either continual learning, self-improving AI, at a high level, how much of this stuff do you think belongs in the harness and how much of this stuff do you think belongs in updating the model weights and biases automatically?

Yeah, I think that's a great question 'cause there have been a lot of debates recently, especially online, about whether we should just stay in a context or the harness space or we should go into the weight space.

And I think the short answer here is that these two methods are not directly competing with each other, or I wouldn't say one is an alternative for the other.

It's not that we're doing something in the context space.

It's not that it's preventing us from doing something in the weight space.

And I personally believe that they have different, uh, relative advantages.

For example, for, uh, doing things in the context and harness space, the obvious advantage is that, uh, it's enabling faster development iterations.

It's al- it's also giving you much more interpretable learning outcomes, as you can see directly what has been learned from, for example, the context space, and so on.

For things that go into weights or things you want to internalize, for example, if you have something that, uh, will be persisted for a long time, uh, some very repeated job you know will happen for, uh, many times in the future, or some more fundamental capability you want the model to have persistently, you want the model to have internally, that should go into a weight space.

So yeah, briefly speaking, I don't think these two methods are I pick A, so I, I, I would not be going for B completely.

It's more like, I believe ultimately it would be a combination of both.

And so I'll take an example.

So I mean, I was reading, let's say Cursor, for example, with the, the last model that they built themselves.

So with my understanding, so obviously Cursor has their own harness that they built and they improved, but they took a lot of the outputs of-- from that, and they did a lot of like post-training on Kimi 2.

So in your view, like let's say if you're building a harness and you have-- you, you, let's say you have the access to the model to update it, do you think for now, like do these things-- do they have to be hand in hand? Is my-- is my question.

Like, do you have to be improving your own harness and then taking some of that and feeding it down, back down into the learning and updating? Or no, for now, and then for now, I know, I mean, it's like the state of the art today.

Do you think these two are still separate from each other, so, uh, there's some kinda new architect, basically the improvements of the model are a different team, different people working on it versus p-people improving the harness automatically?

Yeah, I wouldn't say these two are two separate jobs in some sense, especially given now that, uh, model development is a lot of the effort are spend-spent on the post-training phase.

I would say a lot of the times, many of the efforts we spend on post-training are overlapping with the efforts we're spending on developing harness or even test time approaches.

Mm-hmm.

Yeah, so when we first, for example, talk about test time related methods a few years ago, we were mostly talking about chain of thought or how we enable longer thoughts in reasoning models.

So that was described as something that you spend more compute on a, a single question during test time.

That was the original test time compute.

Got it.

So just at a very high level, the self-improving AI, recursive AI is really top of mind in the overall industry and, and the research community.

And so as you're thinking about either continual learning, self-improving AI, at a high level, how much of this stuff do you think belongs in the harness and how much of this stuff do you think belongs in updating the model weights and biases automatically?

Yeah, I think that's a great question 'cause there have been a lot of debates recently, especially online, about whether we should just stay in a context or the harness space or we should go into the weight space.

And I think the short answer here is that these two methods are not directly competing with each other, or I wouldn't say one is an alternative for the other.

It's not that we're doing something in the context space.

It's not that it's preventing us from doing something in the weight space.

And I personally believe that they have different, uh, relative advantages.

For example, for, uh, doing things in the context and harness space, the obvious advantage is that, uh, it's enabling faster development iterations.

It's al- it's also giving you much more interpretable learning outcomes, as you can see directly what has been learned from, for example, the context space, and so on.

For things that go into weights or things you want to internalize, for example, if you have something that, uh, will be persisted for a long time, uh, some very repeated job you know will happen for, uh, many times in the future, or some more fundamental capability you want the model to have persistently, you want the model to have internally, that should go into a weight space.

So yeah, briefly speaking, I don't think these two methods are I pick A, so I, I, I would not be going for B completely.

It's more like, I believe ultimately it would be a combination of both.

And so I'll take an example.

So I mean, I was reading, let's say Cursor, for example, with the, the last model that they built themselves.

So with my understanding, so obviously Cursor has their own harness that they built and they improved, but they took a lot of the outputs of-- from that, and they did a lot of like post-training on Kimi 2.

So in your view, like let's say if you're building a harness and you have-- you, you, let's say you have the access to the model to update it, do you think for now, like do these things-- do they have to be hand in hand? Is my-- is my question.

Like, do you have to be improving your own harness and then taking some of that and feeding it down, back down into the learning and updating? Or no, for now, and then for now, I know, I mean, it's like the state of the art today.

Do you think these two are still separate from each other, so, uh, there's some kinda new architect, basically the improvements of the model are a different team, different people working on it versus p-people improving the harness automatically?

Yeah, I wouldn't say these two are two separate jobs in some sense, especially given now that, uh, model development is a lot of the effort are spend-spent on the post-training phase.

I would say a lot of the times, many of the efforts we spend on post-training are overlapping with the efforts we're spending on developing harness or even test time approaches.

Mm-hmm.

Yeah, so when we first, for example, talk about test time related methods a few years ago, we were mostly talking about chain of thought or how we enable longer thoughts in reasoning models.

So that was described as something that you spend more compute on a, a single question during test time.

That was the original test time compute.