Comma.ai: Refactoring for Growth
blog.comma.ai
blog.comma.ai
Also, from the FAQ:
> What happens when cars ship with openpilot-quality ADAS? > Have the auto manufacturers even shipped you a CarPlay-level infotainment system? How about OTA updates?
This level of snark is fun, but I don't want the hardware/software driving my car to be "fun" I want it to be serious. So much of Comma.ai feels unserious like this, it's interesting on a technical level, but the vibe is just way off. They charge $1000 to talk to their VP of business development... like they don't want to, err, develop the business? Wouldn't partnering with car manufacturers be a great idea here? Isn't it weird to be so against that sort of partnership? And they're seeking donations for getting their software to work on Toyota cars.
Everything sounds like it's just a bunch of hackers, which, again, is cool, but not something I'd drive and I consider myself somewhat in that category. Why would anyone without a GitHub account and a soldering iron in a drawer somewhere use this?
This annoys me because I think the product is a great idea, particularly for low/mid-range cars that might not have the functionality built-in. It just seems to be hampered by weird marketing.
You need a DIY attitude. And not just a we renovated our bathroom DIY attitude you have to be really willing to dig into understand how and why this works and where it’s gonna not work correctly with your particular vehicle.
It’s already an expensive device that goes in an expensive vehicle driving on public roads where you can get cited tickets or possible liability if the thing really misbehaves and getting an accident with it.
I consider myself extremely DIY/hacker and this is one of the things that I’ve really struggled to get past the mental hurdles.
I don't think it's so much the DIY nature that bothers me, at least it's not the first issue I'm hitting, it's the unseriousness of it all. Maybe if this was being sold by a major car manufacturer in its current form, I might start to think about how DIY it feels, but I think that would be easier to get over because it's visible. As it is, it's the invisible aspects, like the software that's so critical to things like this, which bothers me coming from a company with this sort of attitude.
They aren't against this at all. They think their product isn't good enough for this and would prefer to go it their own way for now.
> Why would anyone without a GitHub account and a soldering iron in a drawer somewhere use this?
They largely wouldn't, yet. geohot et al are aware of that and don't care.
Not sure what the problem is. The product isn't good enough, they're working on it, they want open source + bounties to improve the product, and yet the product is profitable. It's early.
Furthermore, nothing is stopping anyone from developing a competitor and getting into the driverless / ADAS field themselves. But no one has. So I'll take the man in the arena over... well, nothing.
I understand that the product might not be ready to go mainstream but that's not what the marketing says to me, in fact much of their marketing says they are better than the competition, and having followed the company for years they've been through cycles of saying it's amazing and saying it's not good enough.
Honestly I feel like the whole thing is just whatever geohot is feeling each day, but I don't know, I hope there's more strategy than that.
A GM at Ford won't be paying this. That's for jabronis and randoms who want to pitch nonsense.
I think the price is fine. Mostly it's a form to say "I don't want to talk to you." He could just say that, but people would be mad as well. No reason to take phone calls from random "business development" people anyway as a startup.
I realise that's probably the intention, but that's not what it says. They could just as easily not say this, and it would be better. I suspect HSBC doesn't take random meetings with dudes with Bitcoin trading strategies, but they don't say that on their website, instead they have uninteresting stock photos and taglines that say "you can trust us with your money".
This is the root of the problem, the company markets and communicates in a way that says "we're a bunch of hackers that don't care what you think", and that's not appealing to most people.
He's an incredibly smart guy and has worked out the best way to partner with these slow moving cargo ships is to be so good that they can't ignore you. Their team is nimble so will eventually outmaneuver them and then the meetings will be on their terms, not time wasting.
https://data.consumerreports.org/wp-content/uploads/2020/11/...
But I'm confused; while I've only skim-read, the 'overall ratings' on page 6 appear to show that Comma 2 rates as best, not second best?
> We’re growing the team by ~5 people to help solve autonomy, improve the product, and scale up production. You’ll hear more about each team in the following posts from the head of each team.
> Check out our new leaderboard and jobs page. It’s got some fun programming challenges that also happen to be hiring challenges.
In other words, mainstream usage would suggest the interpretation that these are layoffs but comma.ai is either ignorant of this, doesn't care, or deliberately used the language as clickbait.
When it comes to language you don't get to blame "mainstream euphemisms" for your own failure to communicate—you work with the language that people actually speak or you risk miscommunication.
That's... an amazing syllogism.
most people = people in the corporate tech bubble
That’s the issue.
one of my friends at another job with same benefit doesn't have kids so he goes to work at 11 and eats lunch and then stays for dinner too.
I'm sure it was actually intended as "2 free meals at work", and while it may just be poor written communication, it could also speak to an assumption about how much work defines the lives of the employees.
Arrive at 11am, lunch at 12, dinner at 5, that’s only 6 hours so far.
Is that really that long hours?
It is! This is run by George Hotz, aka geohot, aka the kid who cracked the iPhone SIM lock at 17yo, released the 1-click jailbreak for iOS before he was 20, and then went ahead and cracked the PS3 shortly after and released Sony’s private key (used to sign all PS3 software) for all the world to see.
He’s a beast. Now he’s doing Tinygrad and Comma. You won’t be seeing Corpo-speak from this guy or his team lol.
Cool to see him doing well and doing it his own way.
Yet not even a single a achievement is mentioned by name, nor any specific detail is given.
It's evidence for nothing.
not just 'a bunch of people'
> Disclaimer: I have personal experience with him when he took the PS3 ECDSA fail research I and the other fail0verflow folks presented at CCC, and a couple weeks later released the PS3 metldr keys obtained using our method, with zero credit or reference to us. Then Sony sued us all because they assumed we were working with him, even though we'd been careful not to actually release any crypto keys precisely to avoid giving the lawyers excuses to sue us.
> So yes, I was named as a defendant on a lawsuit with him, thanks to his antics. That wasn't a fun few months.
The symmetric half of the metldr key was obtained with a novel exploit I found. Then the asymmetric half was derived from that with the fail0verflow method.
But who really still cares about any of this, it was 13 years ago.
Thanks for the correction.
Are you aware of what website you are commenting on?
That's when I realized that there will always be people that love or hate any given thing, and feedback should be considered in aggregate.
If you're somewhat car+tech inclined, and have the right car, you should give this a shot.
Given that we're talking about a driver assistance system, I assume the quality of the maps can hugley affect the quality of the system. I may be wrong of course.
It works well on the interstate, though.
Toyota Camry 2020 LE
Nobody has "solved self-driving," not even close. Both Hotz and Musk keep repeating this same bullshit and neither of them has managed anything more than fancy lanekeep assist on the highway, and in the case of Tesla, a really unsafe city driver.
Seems they are up-front about it.
[1] https://github.com/commaai/openpilot/blob/master/docs/LIMITA...
Yes.
For clear communication, the title, introduction, and conclusion should all align with the body of the message with minimal surprises. If you intentionally subvert a trope and depend on people to read through the entire release to realize that, then a non-zero fraction won't because each component has a bounce-rate (headline > 1st line > rest of article). There's a bunch of people who believe there are layoffs at Comma - you can blame them for the communication failure if you want.
Except the first line says "unconjoined triangle of success" which does indeed flag this is trope subversion.
Like, for example they support: Ford Escape 2020, 2021, 2022
'openpilot upgrades your Ford Escape with automated lane centering at all speeds, and adaptive cruise control that automatically resumes from a stop.'
But even the base model of these cars already does this including coming to a complete stop. They do it very well and rarely disengage.
Classic mistake for software companies who go into manufacturing!
Lights-out shouldn't be the goal, the goal should be lowest cost, most flexible, scalable e.t.c...
By snubbing their nose at the business community and automotive executives, they have for sure secured a spot for themselves in the annals of obscurity.
They pioneered the idea of end to end learning for self driving, they were talking about it seriously earlier than most and yet Tesla got a hang of it and within a year Tesla has released a fantastic e2e model miles ahead of whatever comma released. Tesla’s model is that good, I can see it becoming a robo taxi.
I can’t help but think that if they weren’t fixated on bootstrapping, shippable intermediaries etc etc, they could have been what Tesla is now. What if they actually tried to scale? It would have been off the vibe they are going for, but if they went the standard route and raised say 500M+ which is reasonable in this space, bought a ton of GPU’s, hired a bunch of ML superstars, hired a bunch of drivers to collect data, they could have already been leading against FSD. Of course whether they actually could have raised enough capital to do all this, is a hypothetical but it would have given them a fighting chance.
Then they tried to reinvent hiring, create good PR’s, solve challenges, they said which sounds like a good idea only if you’ve never tried to scale a company. When you’re a small startup, this may be a good way to sift through candidates but if you want to play in the big leagues this is a naive no go.
First off, most superstar candidates probably have not heard of you or even if they did wouldn’t be bothered to do your challenges, PR’s (god forbid microinternship etc). You need recruiters and head hunter to go after them, go to their universities, actively seek them out. Then you also need a reason for them to join when Anthropic is offering 600k to new hires, no one will choose comma ai in comparison. (Which is why you really need to scale, scale, scale to be able to afford them). Even if you offered them higher salaries, it’s still hard to hire them, you need a nexus of existing talent, a few well known names initially which will convince them to join. This will inevitably lead to some bad hires, some over hiring since you won’t be literally vetting them to a micro internship etc, but you can let them go quickly once you identify them. This will also lead to some redundancy. Having just 21 engineers is not a good thing, when the competition is literally running away with your product. There is a tendency to become too big like say cruise with 800+, but aim to at least have a 100+ , your organization speed will ramp up, and the product can defend against competition much more quickly.
TLDR; Comma for some reason acted allergic to scaling, when a traditional scaling route could have done them wonders. Now it might be too late. I also find it ironic that the best car on the streets right now is a Waymo, I took it through China Town SF no issues and this was the company Comma spent so much time shitting on. Apparently you could just scale up traditional robotics and solve robotaxis (whether it will ever become a profitable business, jury is still out)
Tesla is counting on “humanoid “ robots. Which , uhh, yah, good luck undercutting minimum wage workers.
This was the changelog message lately.
https://www.teslaoracle.com/2024/07/13/tesla-fsd-v12-5-highw....
It probably drove away on its own.
Now you might ask why people talk about Tesla E2E all the time if it's not clear exactly what it is. I don't know either. Let me know if you figure out a charitable rationale.
As for Waymo...is it really the best? It won't even drive on highways! Plus, it's not like you can "buy" a Waymo car. Nor will you ever be able to, I suspect.
Tesla is still the leader, but with Cruise in a death spiral and Waymo burning through so much cash their balance sheet is on the verge of spawning a gravitational singularity, I would not be so quick to dismiss Comma's strategy. Unlike anyone but Tesla, they ship a product you can actually buy, and actually use.
They've delivered a meaningfully functional product that you can actually buy for a very reasonable price and retrofit to all manner of cars, while raising very little and now being profitable.