George gave me a demo drive and a tour around comma offices after our podcast chat. Link to video of conversation: https://www.youtube.com/watch?v=_L3gNaAVjQ4

LINKS:
Comma.ai's Twitter: https://twitter.com/comma_ai
Comma.ai's Website: https://comma.ai/
George's Instagram: https://www.instagram.com/georgehotz
George's Twitch: https://www.twitch.tv/georgehotz
George's Twitter: https://twitter.com/realgeorgehotz
Comma.ai YouTube (unofficial): https://www.youtube.com/channel/UCwgKmJM4ZJQRJ-U5NjvR2dg

OUTLINE:
0:00 - Introduction
0:30 - Comma demo drive
8:14 - Tour of comma offices

CONNECT:
- Subscribe to this YouTube channel
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/LexFridmanPage
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman
- Support on Patreon: https://www.patreon.com/lexfridman

George Hotz gave me a demo of the Kama 2 on a ride back to the Kama AI offices in San Diego.

I found both the Kama 2 system and the entire Kama AI team and operation very impressive.

As 1 robot to another we're in this together my brother.

And now on to the ride and tour with George Hotz.

I'm here with driving with George Otts and looking at the Comma 2, trying to find a stretch of road where we can enable it.

I mean I could enable it here, but this isn't really where it shines.

so we don't have, we don't have red lights and stop signs yet.

I don't want to ship it until we have a real holistic 1.

So you would be able to then like stop at a stop sign.

So this is the, so we're gonna have to stop at this light.

We have a good, we'll go on the 163 a bit and that's a good test of whether your level 2 system is any good.

Super Cruise can't do it because they're not good enough with roll on the road.

How's San Diego in terms of quality of roads and lane markings and all that kind of stuff?

Pretty mediocre, like the rest of California.

What's your favorite reason for using Kama?

You'll do it without a disengagement, almost.

You can, you know, it's 2 hours to LA from here.

Get in your car and you just sit there and watch.

So it doesn't technically require perfect lane markings?

In fact, you can do lane changes on roads that don't have lane markings because it's all end-to-end.

I'm just going to take control here to do it, Especially since there's a cop there.

So we'll just get over 3 lanes and then this road's a decent demo.

Yeah, so I think only us in Autopilot can do this road.

And it's not just curvy, it's also banked.

And if you don't, yeah, if you don't do that well, if you don't have a model that can detect that.

So the longitudinal policy is actually just being done by the car.

Hyundai has quite a competent ACC, probably the best 1.

So we haven't really felt the urge to move to our gas and brakes.

There's a few cute things where which ours will do better and more holistically than theirs.

How hard is it to get it to be this smooth?

They're used, but they're more used as a, see, like right there?

That's the stock system being too harsh on the brakes.

If you're a human driving, you would have been much more gradual on the brakes there.

We're really going to push to switch all the longitudinals to open pilot once we have end-to-end longitudinal and the same way we have end-to-end lateral.

What's harder, end-to-end longitudinal, end-to-end lateral?

We put a lot more work into end-to-end lateral because there's a lot more subtlety in that.

The car companies have managed to build competent longitudinal.

We have competent longitudinal using old school policies.

For example, like when a lane splits out, Should you follow that split?

And making it all smooth and not freak out.

And this thing, you know, the reason I feel very confident in keeping my hands off the wheel, Our torque limit is so much lower than Tesla's.

No, we do injection tests where 1 person sits in that seat with a joystick and the other person keeps their hands off the wheel and that person, you can't look at them.

They're free to jam the joystick in any direction.

And if you don't feel safe in that maneuver, they take back over here because we're getting off here.

that stretch of road is a competent demonstration of open pilot.

Maybe when we get to the city streets, I'll engage on the city streets.

Is city streets like a compelling use case?

Or is that just kind of a intuition builder?

Without end-to-end longitudinal, it's pretty, like, it's not really useful as a product.

So it's more just, it's cool that it can do it, and like, we don't gate it off, but most of OpenPilot is highways.

Like, most of the value we'll gain is highways.

But also, well, not just highways, also like, like, 1 lane each way kind of things, like double yellow kind of roads.

I was doing all the mountains up in Colorado with this thing.

You get to look at the mountains before It drives.

Now remember, the longitudinal policy is Hyundai, and it's not ideal for cities.

But we're going to go through this intersection with no lines.

See, notice how there's no line there, yet our placement is still almost human ideal.

So, yeah, I mean, there you see we did a stretch of, you know, a minute of city and perfectly human, nothing sketchy.

So places like that are where disengage on gas is really nice.

You always know when that hand off is happening with the noise.

It didn't like keep the ACC on or anything.

God, if it kept the ACC on and I make the turn and it's like trying to accelerate at a weird time.

You know, you're just asking to have a bad time.

I like the sounds, the visuals, that clear when it's on and when it's off.

That's the intersection without thinking about it.

It's communicating the uncertainty with this movement a little bit.

Humans actually wobble a little more than you think too.

You notice it a lot more when they're this.

I think Jess Levinson mentioned this on 1 of his interviews.

He's like, so, you know, we have all these, we have the safety driver sit in the car and mark every tiny mistakes it made.

And then we also had the safety driver sit in a car that was being human driven and mark all the little tiny mistakes it made.

It doesn't light up anymore, I don't know why.

We're almost at 30 million open pilot miles.

So here, here's dailies, weeklies, and so actually, yeah, dailies are like, maybe not

So those are used daily and then used weekly.

Weeklys are almost 2.5 and monthly is a little over 3.

how many devices are on our botnet in the last hour.

In the last 30 days, we had 2,200 in the US.

So percent of miles engaged is about 50 percent.

Our disengagements are split pretty evenly between cancel, gas and break.

Most of our segments are now coming from Comma 2s.

Comma 2s is the big breakout for this company, yeah.

This is our live link to the Europe office.

George just erased a bunch of top secret things.

Look, the mission's at the top of the board.

Solve self-driving cars while delivering shippable intermediaries.

For legal reasons, it's a compute cluster.

But the Comma 2 is going to get OpenPilot 1.0.

I mean, we're getting close on a lot of these.

I'm going to start lowering the wheel touch timeout.


Speaker 0: George Hotz gave me a demo of the Kama 2 on a ride back to the Kama AI offices in San Diego. Then we took a quick tour of the office. I found both the Kama 2 system and the entire Kama AI team and operation very impressive. Plus I got a 3D printer named after me. As 1 robot to another we're in this together my brother.

And now on to the ride and tour with George Hotz.

Speaker 1: I'm here with driving with George Otts and looking at the Comma 2, trying to find a stretch of road where we can enable it.

Speaker 2: Start looping back to the office. I mean I could enable it here, but this isn't really where it shines.

Speaker 1: Yeah,

Speaker 2: so we don't have, we don't have red lights and stop signs yet. I don't want to ship it until we have a real holistic 1.

Speaker 3: Until we

Speaker 2: have an end-to-end 1.

Speaker 1: So you would be able to then like stop at a stop sign. So the longitudinal control. Did you just enable it?

Speaker 2: I just enabled it, yeah. So this is the, so we're gonna have to stop at this light. But

Speaker 1: How do you enable it?

Speaker 2: Cruise control.

Speaker 1: So it's the integrated knobs on the car.

Speaker 2: We have a good, we'll go on the 163 a bit and that's a good test of whether your level 2 system is any good. Super Cruise actually can't do it. Super Cruise can't do it because they're not good enough with roll on the road. Autopilot can do it and we can do it.

Speaker 1: How's San Diego in terms of quality of roads and lane markings and all that kind of stuff?

Speaker 2: Pretty mediocre, like the rest of California.

Speaker 1: What's your favorite reason for using Kama? Like, when do you find the most joy?

Speaker 2: Oh, 0, it's long trips. It's, you know, driving from here to LA. You'll do it without a disengagement, almost. You can, you know, it's 2 hours to LA from here. You could just go to LA.

Get in your car and you just sit there and watch. And it's a great time. So we can engage here. So, you know, watch.

Speaker 1: So it doesn't technically require perfect lane markings?

Speaker 2: No, not at all. In fact, you can do lane changes on roads that don't have lane markings because it's all end-to-end. So I have to get over 3 lanes. I'm just going to take control here to do it, Especially since there's a cop there. So we'll just get over 3 lanes and then this road's a decent demo.

Yeah, so I think only us in Autopilot can do this road.

Speaker 1: It's Pretty curvy.

Speaker 2: Yeah. And it's not just curvy, it's also banked. And if you don't, yeah, if you don't do that well, if you don't have a model that can detect that. So the longitudinal policy is actually just being done by the car. Hyundai has quite a competent ACC, probably the best 1.

So we haven't really felt the urge to move to our gas and brakes. There's a few cute things where which ours will do better and more holistically than theirs. But really it's about the turning.

Speaker 1: How hard is it to get it to be this smooth?

Speaker 2: Hard. 5 years of engineering.

Speaker 1: All of this end-to-end, huh?

Speaker 2: We weren't detecting the lane lines. They're used, but they're more used as a, see, like right there? That's the stock system being too harsh on the brakes. If you're a human driving, you would have been much more gradual on the brakes there. We're really going to push to switch all the longitudinals to open pilot once we have end-to-end longitudinal and the same way we have end-to-end lateral.

Speaker 1: What's harder, end-to-end longitudinal, end-to-end lateral?

Speaker 2: We put a lot more work into end-to-end lateral because there's a lot more subtlety in that. The car companies have managed to build competent longitudinal. We have competent longitudinal using old school policies. But lateral, there's so much subtlety. For example, like when a lane splits out, Should you follow that split?

Should you stay in the center?

Speaker 3: Should you

Speaker 2: stay to 1 side? Or so much subtlety there.

Speaker 1: And making it all smooth and not freak out.

Speaker 2: And this thing, you know, the reason I feel very confident in keeping my hands off the wheel, Our torque limit is so much lower than Tesla's.

Speaker 1: So you can't do anything crazy?

Speaker 2: No, we do injection tests where 1 person sits in that seat with a joystick and the other person keeps their hands off the wheel and that person, you can't look at them. They're free to jam the joystick in any direction. And if you don't feel safe in that maneuver, they take back over here because we're getting off here.

Speaker 4: But yeah,

Speaker 2: that stretch of road is a competent demonstration of open pilot.

Speaker 1: Yeah, that

Speaker 2: was impressive. It's not the high end. Maybe when we get to the city streets, I'll engage on the city streets.

Speaker 1: Is city streets like a compelling use case? Or is that just kind of a intuition builder?

Speaker 2: Without end-to-end longitudinal, it's pretty, like, it's not really useful as a product. So it's more just, it's cool that it can do it, and like, we don't gate it off, but most of OpenPilot is highways. Like, most of the value we'll gain is highways. But also, well, not just highways, also like, like, 1 lane each way kind of things, like double yellow kind of roads. I was doing all the mountains up in Colorado with this thing.

You get to look at the mountains before It drives. So I can engage it here. Now remember, the longitudinal policy is Hyundai, and it's not ideal for cities. But we're going to go through this intersection with no lines. Oh, yeah.

See, notice how there's no line there, yet our placement is still almost human ideal.

Speaker 1: Yeah, this is great.

Speaker 2: So, yeah, I mean, there you see we did a stretch of, you know, a minute of city and perfectly human, nothing sketchy. So places like that are where disengage on gas is really nice. Really easy to, you know, hand off. You always know when that hand off is happening with the noise. It didn't like keep the ACC on or anything.

God, if it kept the ACC on and I make the turn and it's like trying to accelerate at a weird time. You know, you're just asking to have a bad time.

Speaker 1: I like the sounds, the visuals, that clear when it's on and when it's off. Love it. Well done. Look at this, let's get on the bike. Damn.

That's impressive. That's really impressive.

Speaker 2: That's the intersection without thinking about it.

Speaker 1: Okay, that was awesome. It's communicating the uncertainty with this movement a little bit. I like it.

Speaker 2: Humans actually wobble a little more than you think too. You notice it a lot more when they're this.

Speaker 1: Yeah, that's an interesting point. Yeah. Yeah, true.

Speaker 2: I think Jess Levinson mentioned this on 1 of his interviews. He's like, so, you know, we have all these, we have the safety driver sit in the car and mark every tiny mistakes it made. And then we also had the safety driver sit in a car that was being human driven and mark all the little tiny mistakes it made.

Speaker 1: So this is the Karma office in downtown.

Speaker 2: Karma office is right here.

Speaker 1: We're at Karma AI offices with Mr. George Hartz.

Speaker 3: Can you give

Speaker 1: us a little tour?

Speaker 2: It doesn't light up anymore, I don't know why.

Speaker 4: Do you

Speaker 2: think most of these dashboards are okay? We're almost at 30 million open pilot miles.

Speaker 3: Wow.

Speaker 1: It's a hell of a number.

Speaker 2: So here, here's dailies, weeklies, and so actually, yeah, dailies are like, maybe not

Speaker 0: 2,000, 1.5.

Speaker 1: So those are used daily and then used weekly. That's really cool.

Speaker 2: Weeklys are almost 2.5 and monthly is a little over 3.

Speaker 4: This is

Speaker 2: how many devices are on our botnet in the last hour. Botnet last day, botnet last week.

Speaker 1: Is this all over the United States?

Speaker 2: This is, no, this is everywhere. So here's the maps. Here's the map of the world. So we're everywhere in the world.

Speaker 1: Wow, that's awesome, man.

Speaker 2: So let's see. In the last 30 days, we had 2,200 in the US. And then Korea, Canada, Taiwan. Japan's up there. This is our percent engaged.

So percent of miles engaged is about 50 percent. Percent time engaged is about

Speaker 0: 30.

Speaker 2: Our disengagements are split pretty evenly between cancel, gas and break. Most of our segments are now coming from Comma 2s. Comma 2s. Comma 2s is the big breakout for this company, yeah. Oh yeah, it's profitable.

This is our live link to the Europe office. We come as a multinational corporation.

Speaker 1: George just erased a bunch of top secret things.

Speaker 5: Censorship, the censors came in.

Speaker 1: Yeah, yeah, yeah.

Speaker 5: Look, the mission's at the top of the board.

Speaker 1: Solve self-driving cars while delivering shippable intermediaries. Amen. What do we do badly? What can we do better? Right.

Speaker 2: Good questions.

Speaker 1: OpenPilot 1.0, when is that coming up?

Speaker 2: When those things are done.

Speaker 1: PyTorch. Okay. I like how that's a bullet point. That's like move everything to PyTorch.

Speaker 2: Yeah. Oh, sorry. It's not a data center. It's a compute cluster. For legal reasons, it's a compute cluster.

It's a compute cluster. OK.

Speaker 1: That sounds sexier, too. Data ingestion, multi-chain. Cool. So this is

Speaker 2: like a

Speaker 1: vision board for OpenPilot

Speaker 0: 1.0.

Speaker 2: It's not gonna happen for a while. But the Comma 2 is going to get OpenPilot 1.0. It's like what I said with the hardware. So this is all gonna come to the Comma

Speaker 0: 2,

Speaker 2: just with OTAs.

Speaker 1: OTA, over the air, we're gonna get it.

Speaker 2: Yeah. I mean, we're getting close on a lot of these. I'm going to start lowering the wheel touch timeout. Alex, you want to? This is Alex.

Speaker 3: Hi, nice

Speaker 4: to meet you.

Speaker 2: She's our COO.

Speaker 1: COO, nice. Hey,

Speaker 2: good to meet you. She runs the back there stuff.

Speaker 1: What's the technical description of the back there stuff?

Speaker 4: You want to see?

Speaker 1: Yeah.

Speaker 4: I think it's more exciting than this

Speaker 1: stuff.

Speaker 4: It's better you ask.

Speaker 2: We're ahead of all the secrets, right?

Speaker 4: Nick, do we have any secrets out?

Speaker 1: I'll blur it out. I'll blur the top secret. Stress test.

Speaker 4: This is our final test. The cometoos hang out here for about 24 hours. We test them for their heat, their temperature, their light sensors, we do cosmetic checks, screen checks. So, pretty exciting. You can see they all start over there as phones.

Then they get retrofit.

Speaker 3: It's a

Speaker 4: beautiful aspect. You can see We have a

Speaker 3: class of

Speaker 0: 100

Speaker 4: phones out right now. They're getting tested. Cool.

Speaker 3: 3D printed cases. Automatues

Speaker 4: being made. Automatues and then.

Speaker 1: It's kind of exciting actually.

Speaker 2: Lex, this is Mitchell. He's the

Speaker 5: latest addition to

Speaker 2: our machine learning team. That's Chris, also known as Virtually Chris. He's a famous YouTuber.

Speaker 1: Oh, wow. How did fame change you?

Speaker 3: I introduced you as

Speaker 2: a famous YouTuber. Yes, I was the open pilot YouTuber. So they got hired. Nice communication.

Speaker 1: Don't let it get to your head. Yeah.

Speaker 2: Is that it?

Speaker 1: That's it.

Speaker 2: That's our head of hardware.

Speaker 1: Awesome. What's all the what's that stage of the of the assembly

Speaker 4: sink and the fan and then we put the cases on them.

Speaker 1: Awesome.

Speaker 4: Yeah, they're taking cases off, screens off here, they're testing screens over here, oven.

Speaker 1: Yeah, 3D printer, you want to show that?

Speaker 4: Oh, I'm afraid to take it in there. It's very hot and loud.

Speaker 2: Well, it's also the home of the compute cluster.

Speaker 4: It's not a data center. It's not

Speaker 2: a data center. That's right. It's a compute cluster.

Speaker 4: Prepare yourself.

Speaker 3: It's the

Speaker 2: sound of productivity. Yeah.

Speaker 3: Oh my god, it's really hot. Fresh bag, cases. We're running jobs.

Speaker 4: This is what you see on the outside of the Commodores. These are all of our printers. We've got a lot.

Speaker 2: Yeah,

Speaker 3: I

Speaker 1: like how they have names. Meg. Rings. Why, what's the purpose of the rings?

Speaker 4: They used to be for athletic

Speaker 3: activity, but

Speaker 4: now we have too many printers. We can't anymore. We just got about 20 printers in the last few months. They've come from all over the world. Literally, Poland, Germany, Peru, France.

Speaker 1: How long does it take to

Speaker 3: print a single 1?

Speaker 4: Yes, we'll name number 49 for you, Max.

Speaker 1: Thank you.

Speaker 4: We'll call him Max. I told him he could be number 49. I

Speaker 1: could be, I could be.

Speaker 3: So great,

Speaker 4: 49's not working right now, but I promise 49 will work. It'll be Lex number

Speaker 0: 49.

Speaker 1: I appreciate that, That would mean a lot to me.

Speaker 2: This is cool.

Speaker 3: So unfortunately because we got to keep the bugs and dust out, they're covered, but... 1UI already forgot the

Speaker 2: official name. The compute

Speaker 3: cluster. So these are our CPU machines. We have 3 racks that look like that. 2 petabytes of spinning disks.

Speaker 2: We're in

Speaker 3: our 3080s. Our 3080s are in only desktop computers for non-business purposes. These are GPU machines.

Speaker 1: What's in there? V100s! What's in there?

Speaker 3: V100s! Ha ha ha ha ha!

Speaker 4: It's like a farm, Every day we kind of go

Speaker 3: out there. Yeah, I said

Speaker 2: yesterday. 21st century farm. We got to hire a good farm bill player to come farm us the produce every day. These are the ones that are ready to ship.

Speaker 4: Wow. I think this is These are all of our growing harnesses. These are all the cars we support. MCA, Bosch, Developer, Ninebot, Subaru, Toyota, all the Hyundais. All the way to J, A9, Nissan, VW.

Speaker 1: And all of these are assembled by hand?

Speaker 2: Yeah. That's cool. That's amazing.

Speaker 1: That's legit. That's awesome.

Speaker 4: Yeah, I don't know.

Speaker 2: That's the office.

Speaker 4: Have you been for a ride among the cars?

Speaker 2: He took me

Speaker 1: for a ride. It was awesome. It was really surprising how well it worked.

Speaker 4: Where'd you take him on the highway?

Speaker 2: We did like the 163. In

Speaker 1: the city a little bit too. It held up.

Speaker 2: Definitely worked. You didn't show him

Speaker 4: any secrets though, did you?

Speaker 2: No, I didn't show him any secrets.

Speaker 5: I mean, it's like keep on driving, You know? That's why we replaced it with a heart. We're all big Elon fans.

See all Lex Fridman transcripts on Youtube

Comma.ai Drive and Tour with George Hotz and Lex Fridman