Hacker News new | past | comments | ask | show | jobs | submit login
Comma.ai: Refactoring for Growth (comma.ai)
133 points by ppsreejith 6 months ago | hide | past | favorite | 104 comments



"Refactoring for Growth" sounds like layoffs to me, as does the section of this post going into team structure, and yet there's no layoff announcement. Why not just "Growth for Growth"?

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.


> Everything sounds like it's just a bunch of hackers

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.


How many of those accomplishments were truly his, I wonder ? https://social.treehouse.systems/@marcan/109561314682149814


The only thing the link contains is a bunch of people claiming that his achievements were built on other people's.

Yet not even a single a achievement is mentioned by name, nor any specific detail is given.

It's evidence for nothing.


The next message in the thread gives specifics about the PS3 crack referenced in the parent:

> 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.


So, it sounds like the individual is complaining that he didn't mention them, but also complaining about Sony sueing them, and somehow manages to blame him for Sonys actions? Like Sony gave them credit they didn't want, but they want to complain about not getting the credit they didn't want?


It's as if you gave an academic lecture on a casino's vulnerability to being robbed, and then someone actually went and robbed the casino. You can be mad at the asshole who robbed the casino for two reasons; A: he's put you in hot water that you would have avoided without the robbery; B: he took credit for discovering the vulnerability (or at least didn't correct people who assumed he must be a real hotshot at casino security design)


Except that nobody robbed anyone.


Weird, I could swear that I haven't seen that message.

Thanks for the correction.


The worst part of this critique is that it's not even true, and marcan knows better. He clearly dislikes me so much that he is willing to lie.

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.


except marcan is https://archive.org/details/console-hacking-2010

not just 'a bunch of people'


marcan is at least 2 people


A great thread to show how Mastodon is no better than Twitter. Really obnoxious and toxic and not something I want to be near.


That's not platform specific, it's the humans


Exactly! And I would expect hacker news of all places to understand that.


Didn't geohot leave Comma like ~2 years ago


He left Coma for tinycorp the same way Elon left SpaceX for Twitter.


Didn't Geohot also leave for twitter?


He was only at Twitter for a few weeks, maybe a few months before quitting because it turns out things are A Bit More Complicated Than That TM. This is one of the reasons why I would feel weird about entrusting my life to Comma.ai. I'm not that keen on Move Fast and Break Things when the things are my bones.


He famously criticised the engineers there, joined, delivered basically nothing and then left.


You forgot to mention how he also single-handed solved search at Twitter /s


> Wouldn't partnering with car manufacturers be a great idea here? Isn't it weird to be so against that sort of partnership?

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.


Charging $1000 to talk to their business development person does not suggest taking partnerships seriously to me.

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.


He talks about this in his webcasts, in the past he said he has wasted untold amount of time with executives and middle management at so many companies, who all promise the world but their only real power is to organise the next meeting.

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.


> Charging $1000 to talk to their business development person does not suggest taking partnerships seriously to me.

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.


> A GM at Ford won't be paying this. That's for jabronis and randoms who want to pitch nonsense.

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.


A GM at Ford absolutely will be paying this if they want a phone call. Joining our Discord is free however.


I stand corrected. Still, I agree with your decision.


When Consumer Reports tested a Comma 2 against ADAS from car manufacturers, it rated second best.

https://data.consumerreports.org/wp-content/uploads/2020/11/...


Thanks for sharing the document.

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?


You're right - I misremembered; I last actually read the report when it came out. I thought GM's Super Cruise came out on top, but it seems like Comma did!


They do care, on their Twitter they say they want to get it into Best Buy someday


> Everything sounds like it's just a bunch of hackers

Are you aware of what website you are commenting on?


A side-project for a very famous early-stage startup VC fund that offers advice, networking and mentorship so that hackers aren't just winging it?


You don’t need a github account or a soldering iron to use this.


I think they’re being a bit facetious, but the underlying point stands.

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 took the leap and it's been more than I could have wished for. i don't have a ton of miles, maybe 10k miles but no issues. it's just level 2 driving which is better than anything OEMs offer.


Same, I love mine, I really wanted what basic autopilot gives in a Tesla in my Highlander and it does that incredibly well.


I also really love mine. It has done what it has advertised extremely well, in my opinion. Makes driving chill. I have a '24 Ioniq with HDA2 and it's a night and day difference between Comma's implementation of HDA/LKAS and stock. If the stock version is like a 3/10, OpenPilot is like a 8 or 9/10. Stock HDA would just turn off randomly in the middle of turns. Its LKAS was very jerky which not only scares you, but scares other people on the road as well.


how does it compare to Blue Cruise on Ford vehicles?


blue cruise isnt available on my full road trip route.


Thank you, I was being sarcastic, like the marketing material I highlighted in fact.

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.


I will refer to this comment the next time anyone complains about enterprise speak in announcements.


Eh, no matter what you do, you can't please all people. Some people like the corporate speak, some hate it. I got the reverse a lot, where people recommended movies or shows as ABSOLUTELY THE BEST THING EVER, and I watched them and hated them.

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.


You don't need a soldering iron to install it, so your entire comment comes off as rude, dismissive and ignorant.


Did you actually read the article? They are just re organising their teams and are actually recruiting 5 more people


No layoffs! You’re conditioned by mainstream tech euphemisms to think that from this post. geohot is straightforward type.

> 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.


> You’re conditioned by mainstream tech euphemisms to think that from this post.

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.


The way I see it, "the language that people actually speak" didn't include refactoring as a euphemism for firing people; it's something corporations came up with. Now we have someone who tries to eschew this euphemism. I think it's a good thing. Using words in their plain meaning makes for a clear communication.


OK, this thread has generally been shitting on comma, but this comment broke the camel's back for me. Is your line of reasoning seriously "other companies misuse words to cushion their shittiness" -> "Comma used the words as they are meant, and didn't have any shittiness to need to hide" -> "we're used to everyone misleading us with those words and Comma didn't" -> "Comma is misleading us by not misleading us"?

That's... an amazing syllogism.


I frankly don't care about Comma at all, I care about language. When most people reading a bit of writing interpret it one way, they're right and the author is wrong. That's all.


For you

most people = people in the corporate tech bubble

That’s the issue.


Yes, if most of the people who will be reading your press release will misinterpret it you have a problem. Audience matters, and if you dismiss your audience as a "bubble" you will fail to communicate with them.


Maybe consider the possibility it’s intentional. Basically by writing this way he got rid of all the annoying people! It’s a filter to only get the attention of beyond the surface thinkers.


I completely agree. Communicating in this way is definitely not being "straightforward". It's either ignorant and tone deaf or just snarky.


I’m not looking for a job, but out of curiosity, I looked at their jobs page and saw that one of their benefits was two meals a day. Umm—I’d prefer not to work somewhere where I was eating two meals a day at the office.


you couldn't see yourself eating breakfast at 8 or 9 and lunch at 12-1 every day? two meals is a perk at my job and it's great.

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 can't see trying to get to work without a breakfast first. But having two lunches isn't the worst idea.


What's the point you're trying to make here? That the role is bad because it implies needing to work from an office? Or that the daily hours are long enough to cross the boundaries of two meal times? Or you take offense to the idea that meals are eaten in the office, as opposed to somewhere else? I'm beyond confused.


I think the point is that the wording doesn't distinguish on the workplace, so it reads as the benefit being "you get to have 2 meals a day" in general life, which for most people having 3 meals a day would feel restrictive and overbearing that their employer would be dictating such details of their life.

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.


He was talking about the long hours. Normal office hours do not span two meals.


They don't? I'd love to get a free breakfast at like 9:15 and lunch at 1. That's not even remotely long hours. Others might work 10-7 and like having lunch and dinner. That's two sets of hours where 2 meals makes great sense. When I was single and living alone, I would have loved that benefit. The breakfast version, I would still do now, since I'm not really fully ready to eat until after 9.


The details are lunch and dinner. If I’m in the office at dinnertime, I’m there too late.


I eat 3 meals a day at my 9-hour office and it's great


At one point I was eating three meals at ~8:30 hours at work. Breakfast just before 10am, lunch at 1-2pm, dinner after 6pm.


> Normal office hours do not span two meals.

Arrive at 11am, lunch at 12, dinner at 5, that’s only 6 hours so far.

Is that really that long hours?


Most people have dinner at like 7pm. It does vary by country but I'm pretty sure that's what they do in the US. 5pm would be an extremely early dinner.


I'm not eating dinner at the office, unless I intend to stay a few more hours (which is only warranted by freak incidents).


But what about second breakfast?


Also not directly related, I've been using comma for 4 years and it's been really fantastic. I feel safe and confident with it, and used it on many long trips. Stop-and-go traffic is no longer bothersome, and long boring drives benefit from the safety of the system.

If you're somewhat car+tech inclined, and have the right car, you should give this a shot.


My car’s radar unfortunately doesn’t work under 28 mph, so stop and go is a nonstarter.

It works well on the interstate, though.


The comma pedal allows for it if you can source/build one


I’ve looked into it, but I think latitudinal control may be locked below a certain speed as well.

Toyota Camry 2020 LE


I guess(?) you are in the US. Is their product available anywhere else?


They ship pretty much anywhere. As long as you have a compatible car[0] with the right options, it will work.

[0]https://comma.ai/vehicles


Yes, I think they ship worldwide and they have various language-speciffic channels in their discord as well. I don't see why you would think it's US only.


Mostly because of the Maps. It's not trivial to have accurate Maps of all the world. Google Maps is, for example, quite less accurate here un Switzerland compared to UK, where I have also seen it.

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 doesn't rely on maps, it relies on its own cameras and the onboard cameras+radars.


Not directly related to the article, but it seems to be they are the way they are is because they are dancing around regulations.


I noticed their cabin mounted device does not "look at mirrors" so I dont trust it to change lanes.


> While changing lanes, openpilot is not capable of looking next to you or checking your blind spot. Only nudge the wheel to initiate a lane change after you have confirmed it's safe to do so. [1]

Seems they are up-front about it.

[1] https://github.com/commaai/openpilot/blob/master/docs/LIMITA...


Exactly. I've heard this argument before, that "humans only need their eyes so you only need one forward-pointing camera" but this makes no sense to me. Humans turn around and look at stuff. Your head moves around constantly. You wouldn't trust something like this on the highway, much less in a city.

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.


Does it change lanes?


Announcing that you're hiring with a title that would normally be a layoff announcement sure is a choice I guess? Either this was a weird attempt at being clever or they're really bad at communicating.


We make fun of companies announcing closing down with an "our incredible journey" post, now we're complaining that companies use "our incredible journey" to actually reminisce about the journey, rather than shut down? Let's be reasonable here.


> we're complaining that companies use "our incredible journey" to actually reminisce about the journey

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.


> 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).

Except the first line says "unconjoined triangle of success" which does indeed flag this is trope subversion.


That's even more niche: not only do people have to read beyond the headline, they must also be familiar with a nerdy TV show that first aired a decade ago (Silicon Valley).


Most cars have Autosteer now provided by MobilEye, I'm wondering where comma ai think they fit going forward?

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.


> Scale up production until we have a lights-out factory

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...


Looking at the dashboards they openly shared with us, I'm not seeing the reason to 10x unit production. Are they just manifesting demand?


This is never going to be pass a compliance review, and as others here have stated there's no way to bag any big cash returns with this approach. This product is only ever going to appeal to a wiry haired hacker type with a soldering iron in one hand, and a C++ textbook in the other.

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.


Haven’t seen a geohotz stream in a while. Is he ok?



thru the footer i discovered https://blog.comma.ai/autonomy/ which i have to call out for clear simple communication which i loved


I will try to be objective here and give them a bit of a reality check, they tried to reinvent the “VC”, tried to reinvent the vibe of SV company but at some point they seem to have lost the ball.

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)


Or they could have been one of the dozens upon dozens of failed self-driving startups that promised the moon and never shipped a product. Sheesh, if only they'd hired a bunch of headhunters to find money-hungry Ivy League CS grads, they could have joined the party of smoldering VC scams!

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.


Agree totally; what Comma has achieved is noteworthy precisely because they've not gone the obvious SV route.

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.


I've also been confused why they stay so small. It seems like such a good idea, why not scale faster and larger. My guess is they are so bound to the hacker ethos and culture. Also the regulatory gray area the product sits in, scaling too large would put a target on them.


Wait, what is Tesla’s “e2e” model? And you “can see robotaxis”?

Tesla is counting on “humanoid “ robots. Which , uhh, yah, good luck undercutting minimum wage workers.


It's not clear what Tesla actually means when they say "E2E". The expectation most people have is that there's a single NN between sensor data and steering wheel/pedals out. Based on the limited technical detail available from "AI day" and such, what they're actually doing is more along the lines of a bunch of ML models glued together in a pipeline.

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.


“FSD (Supervised) v12 upgrades the city-streets driving stack to a single end-to-end neural network trained on millions of video clips, replacing over 300K lines of explicit C++ code”

This was the changelog message lately.

https://www.teslaoracle.com/2024/07/13/tesla-fsd-v12-5-highw....


I have one of these. It’s not there.


> It's not there.

It probably drove away on its own.




Join us for AI Startup School this June 16-17 in San Francisco!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: