Meta is capturing traces of engineers' coding tasks, tool use, and problem-solving steps to train AI models for behavior cloning, according to a leaked April 30 all-hands recording

Full transcript of the leaked audio from Mark Zuckerberg.

"So, in other news, I know layoffs are top of mind, but there were also some updates this week around employee device tracking.

There was a question about employee device tracking: “Can you share more on employee device tracking? I think the way that it was announced left folks with something.”

Okay, let’s talk about what we’re doing. Like Alex just said, going into what makes these AI models great, there are basically a few key ingredients.

There is getting the research and architecture right. There is having good infrastructure, which includes both the quantity of compute and, just as importantly, if not more importantly, how efficiently you can use it, how reliable it is, and the quality of that infrastructure.

Then there is the third piece, which, in some ways, is hard to compare with the others because they are all necessary.

We are in a phase where AI models learn from watching really smart people do things. If you are trying to get a model to develop certain capabilities, it is very important for it to observe really smart people performing those tasks.

There are a few examples of what we are trying across the company. One basic insight and hypothesis we have is that a lot of data generation across the field is done by contract companies. Alex knows a lot about this because he ran one before coming here.

In general, the average intelligence of the people at this company is significantly higher than the average group of people you can get to do tasks if you are working through contractors.

So, if we are trying to teach models coding, for example, then having people internally build tools or solve tasks that help teach the model how to code will, we think, dramatically increase our models’ coding ability faster than others in the industry can do, especially those who do not have thousands and thousands of extremely strong engineers at their company.

That is one example.

Another thing our system needs to be very good at is using computers. The way you get a system good at using computers is by having it watch really smart people use computers. That is basically the essence of what we are trying to do here.

We are rolling it out in a way where no human, or anything like that, is actually looking at or watching what people are doing on their computers. The content is stripped out as much as possible.

None of the data has been used for looking at what people are doing, surveillance, performance tracking, or anything like that. It is purely being used to feed a very large amount of content into the AI model, so that it can learn how smart people use computers to accomplish tasks.

I know this is going to be a big advantage if we do it.

So, anyway, that is what we are trying to do. I think there will probably be other things around the company where we try to use the fact that we have a very high-quality set of people to teach AI systems to do different things.

This probably is not the last thing like this, but I think it is an interesting strategy at this point.

We have some hypothesis, and we will be able to complete the loop to see how well these kinds of things improve the models. If they do not improve them, we will not do more things like this. If they do, then we will probably do more things like this.

So that is the basic thing.

In terms of how we communicate about this stuff, it is tricky. When I was looking through the details, there were things that could have been done better. Yes, that is acknowledged, and we will try to improve.

The core tension is that we want to communicate as clearly as possible about what we are doing, while not having all of the details about things we think will be strategically differentiating leak immediately to competitors.

I think part of the challenge at a pretty big company is that if you post stuff publicly, it leaks. Some things matter more than others.

If we are building something in our ad system, for example, or infrastructure that is specific to us, and it is not something other people are going to copy, then at least it is not as big of a deal. Maybe it is kind of annoying.

But I think we know AI is one of the most competitive fields, probably in history. So anything that can make our quality better is generally not something that I think is in our strategic interest as a company to lay out in detail, knowing that the physics of the situation are that things will leak.

So I think you have to navigate this on a case-by-case basis in how we communicate.

It is not strategically in our interest for us to communicate everything in all the detail that we normally would on this. But we do need to try to make sure that we get this right and communicate enough so people understand what is going on.

This will be a continued thing that we are trying to navigate. It is part of the complexity of running the company through this incredibly dynamic period.

I think there are lots of things people would like more certainty on than we have. There are lots of things people would like details on that are not necessarily bad for any one person, but are bad if they leak.

I do not know exactly how we navigate that. So that is the basic situation on that."