No hype, just works: How Comma reached 100M miles in autonomous driving | E2011

[@ax7011]

Kudos to Harald for speaking the truth like an engineer.

[@abdulr2084]

great interview. Nice to have a CTO (and team) who is excited as the founder

[@tirthb]

This guy must be the greatest to be the CTO with George Hotz

[@johntrotter8678]

I like the Comma goal. My Tesla has FSD, but I have ordered an Aptera with Comma ADAS (if they will indeed exist, both the car and its Comma adaption). FSD is currently too ambitious and fails enough that it makes driving for me more stressful, not less. I expect Comma's system will reverse that.

[@danialothman]

Harald is so chill. No wonder comma makes driving chill 😂

[@gvdseype]

I used to play bfme2 in LAN against this dude. Crazy what he has become.

[@Malius1000]

Fieu!

[@vincentvega4390]

he looks like that one youtuber that goes around in new york and speaks perfect chinese

[@Player-oz2nk]

Love the way he clearly dissected Waymo's current potential demise

[@MelonHusk7]

Alternative to Tesla. It'd be like iPhone vs Android.

[@varshneydevansh]

Indeed it's

[@ehlava7331]

this was excellent

[@zinovate1]

I love my Landroid robot. It’s beautifully simple and just works

[@verynice5574]

So many facts in the first 40 seconds lol. I just need to see a single humanoid robot replace a single person at any job in an economic way. Not holding my breath lol.

[@mattyk35nyc]

🔥🔥🔥

[@thomasbuckler3977]

glasses remind me of dwight schrute a bit

[@siyuanma2323]

Copy from reddit comment: If you need one intervention every 10 rides… that means you only need like 1 operator per 100 cars. Waymo did about 20 rides per car and each intervention takes about 5 mins to solve. One car need 10 mins per day which mean one remote operator can solve 48 cars. The number is not correct.

[@Only4hands]

Math works out, but I believe you’re applying the logic that a human will work at at an efficient pace to resolve these issues. Waymo is more cobot than robot

[@siyuanma2323]

@@Only4hands Human ratio for Waymo is about 1:10 or less. They could make it lower but not the largest cost.

[@jeffsteyn7174]

Waymo is so crap. But the people saying it 1. Cant do what waymo is doing. 2. Doesnt take responsibility for the accident if something goes wrong. Clowns

[@lechavs]

5 years plus for unsupervised FSD! Taking that we're still far, I'd be surprised if takes more than 2 years It doesn't need to be perfect, just safe.

[@siyuanma2323]

Revenue are off by 1000 from the CEO: Waymo is doing 100k per week which means 5million per year, at 20 buck per ride, you got 100million-150million. Another 10x you got 1B~1.5B annual revenue. CEO was talking about 150k per year revenue. It's off by 1000.

[@Malius1000]

I was talking about revenue per car.

[@siyuanma2323]

@@Malius1000 per year or per car in lifetime? If a 150k car generates 150k per year, that's pretty good.

[@Malius1000]

@@siyuanma2323 Per car per year. That might be a little high, but should be in the right ballpark at least.

[@siyuanma2323]

@@Malius1000 Assuming car last 5 years, that means the car investment is 400% ROI. I would say it's a pretty good investment. Most of 1B spend by waymo is either building the cars with great ROI or on engineering. I think the unit economics works pretty well. They just spend more to build new cars.

[@ovidiuc4]

The fidelity of the cameras better than human eye? Come on. They're not putting 100MP Hasselblads in there. Plus our eyes see depth and self-clean. They (car manufacturers in general) should put some eyelids on those cameras.😊

[@Closer-than-it-appears]

It's even worse. Human vision operates at almost 600MP. And relying on blind spot detection for lane changing will not prevent a collision with a speeding vehicle that hasn't yet entered the detection zone when the lane change is initiated.

[@StefanNorberg]

Sigh. Waymo's stack is 95% machine learning these days. It hasn't been rule based for 7-8 years. And FFS there are no HUMANS in the loop in the DDT. They only step in if the car is stuck.

[@Malius1000]

Rule based and machine learning are not mutually exclusive. How frequently does human step in? I think my estimate of 1/10 rides might even be too optimistic.

[@StefanNorberg]

@@Malius1000 I think its less, but I am not sold on the business case either. Re ML/ rules, the latter can work as a safety net or to implement traffic law. I am aware of zero e2e unsupervised safety critical applications, including non-time critical.

[@lsh3rd]

Sadly, not my 2017 Toyota...

[@howtoactuallyinvest]

I rarely have to intervene w FSD v12.5. It's incredible and getting better faster. He's also underestimating the coming ramp up of humanoid manufacturing too. Agility is starting mass production soon w Digit, Tesla will have thousands over the next couple years then millions. What makes it "easier" than autonomous cars is the liability factor.. you're much less likely to die if a humanoid bot makes a mistake vs a self driving car. There's 27 or so humanoid companies and most are in China. The world is about to change in a very big way