Brett Adcock

We have to produce action output out of the robot.

And it's a very different thing than next token prediction for language models.

We ended up finding that the team we had in place, my team lead, the folks we have here all from Google DeepMind or certain areas of top AI programs, and they're really good.

The team now we have is over 50 or so on our Helix team internally.

We just found that that team we had internally, we just ran circles around them.

Like, every day.

We had a hard time getting, like, you know, in robotics, you gotta, like, run the robot, see how it does.

You know, like, you have, like, want to run a new, like, AI experiment or do some ablations, like, and some evals, you need to, like, run the robot at the end of the day and see how it does.

Like, sim is one thing.

You can get certain far running simulations and looking at loss curves and stuff, but we need, at the end of the day, like, do we need to get, like, see how the robot does?

And we just had a hard time getting them in the office.

We had a hard time, like, basically, like, like, basically, you know, advancing stuff together as a team.

Ended up, like, the strategy we had internally and the team we had was just, like, complete superstars.

They're the best robot learning folks on the planet, this sort of figure.

And it got to a point where, you know, I got a call one day.

It was just like, you know, we were, like, also week to week, like, showing them how we were doing all this work.

And I got a call one day saying, like, hey, we're, like,

you know, we've been watching your progress, it's unbelievable.

And, you know, we're thinking about doing robotics work internally.

And I was just like, ah, this is over.