Jeff Dean

I mean, well, at small scale, you do lots of experiments.

So, I mean, there's, I think, one part of the research that involves, okay, I want to, like –

invent these improvements or breakthroughs kind of in isolation, in which case you want a nice, simple code base that you can fork and hack and some baselines.

And my dream is I wake up in the morning, come up with an idea, hack it up in a day, run some experiments, get some initial results in a day, like, OK, this looks promising.

These things worked and didn't work.

And I think that is very achievable because, OK.

At small scale.

At small scale, as long as you keep a nice experimental code base.

It's great.

So there's that part of the research.

And then there's some amount of scaling up.

And then you have the part which is like integrating, where you want to stack all the improvements on top of each other and see if they work at large scale and see if they work all in conjunction with each other.

And then there are trade-offs because you want to keep your, like, integrated system, you know, as clean as you can because, you know, complexity.

Code base.

Yeah, code base and algorithmically complexity, you know, complexity hurts.

Complexity makes things slower, introduces more risk.

And then, you know, at the same time, you want it to be as good as possible.

And, of course, every individual researcher wants his inventions to go into it.

So there are definitely challenges there, but we've been working together quite well.

I've stopped cleaning my garage because I'm waiting for the robots.