A 100x Investment (2019)
blog.comma.ai
blog.comma.ai
Comma is taking an incremental approach which I like. Most driving is done on long monotonous roads going straight. If you can eliminate that portion, you'd add huge value.
The other cool thing is that its completely open-source (except for code used to train the models). You can install it on your own device and run it, or fork it. I read a high percentage of users use custom forks of the software. They also have a discord where you can talk to the engineers and people working there if you have any issues or questions. Imagine calling someone from Toyota and just being connected to a auto-engineer?
There's a lot to like about Comma. There are few businesses out there with this mindset. I hope other companies take their lead and think outside the box and stop worrying about the next round and just get some hackers together and build something people want.
Much like we already mandate things like car lights cannot be on a moving piece of body work in the USA (to ensure lights visible even when boot or trunk open), I wonder if we should ensure that the critical sensors for self driving are hard mounted to the car securely?
There is inevitably going to be occasions this falls off mid-drive - these suction cups are rarely perfect in my experience.
I've always wondered if CarPlay/AndroidAuto one day might somehow just reuse existing sensor suites on vehicles to provide self driving - this would let the cellphone (which increasingly has ML acceleration hardware anyway) still be responsible for the driving with cameras that won't fall off.
I don't think that problem is even in the top 5 risks of the system
If I buy a $40k car and 3 years go by, I'd be happy to buy a system that gives me access to state-of-the-art driving intelligence, using my car's existing sensors.
Comma uses a very strong adhesive for the mount, and then the Comma housing slides into that mount very securely. I don't think it is any more likely to fall off than your rear view mirror is (which also uses an adhesive).
Comma isn't a purely optical system, it also uses the car's forward radar information in addition to its cameras.
Comma does not pretend to be a fully self-driving system (yet), and in fact it has a camera facing the driver and will warn you if you are not paying attention or disengage if you are not paying attention long enough. It's a tool to make driving easier and safer.
I'm quite confident if you sat in my car while Comma was driving you would be shocked at how good it is.
What actually might be interesting would be a "self driving" addon that is actually just simply powered by a human remotely driving your car. Obviously you would need to add cameras and whatnot and it might be priced higher than most would want but with properly fast and reliable internet a random person could be driving your vehicle from their home while you sleep.
I don't think this is actually true at all. I mean, it's a nice little plus for long road trips I suppose, but how often do you do those?
If you can't automate the whole trip, there's little value. I can't go to the bar and have a few drinks and have my car drive me home. I can't avoid parking entirely by having my car park itself. I can't just stop owning a car because there is a circulating supply of autonomous ones.
Those are the real "huge" value adds. Not being able to take a nap on a road trip, and those things are still quite far off.
The typical HN reader lives in a dense metro area where commute distances are much shorter than average. The average commute in the US is 16 miles each way. Millions of Americans spend hours on interstates every day.
This is Tesla, and Comma argument. I think it is a fallacy.
Yes, we can drive with just a pair of eyeballs. But you have to realize that the entire road infrastructure is made for humans. Signs are made to be easily recognizable by humans, human reaction time is taken into account, we have a good idea on what is easy to understand and interpret by humans, but we don't give a fuck about bots when building roads. It means that we start with an unfair advantage. And we are asking bots to be better than we are in that environment.
Humans have a remarkable ability to contextualize a situation. For example, if we see someone walking on a side lane, marked for pedestrians. We know that that person is not a threat, he is unlikely to leave his line and therefore we can pass at a normal speed. That's unless we see that the person is visibly drunk, or maybe he is a small child, in these cases, we will slow down, because he may suddenly jump in front of the car. Bots probably won't know that, and usually, for safety, they will always slow down. Similar situations happen all the time, and that would make for a terrible ride, so what should we do? Make the system ignore the "drunk man" or "small child" case and possibly kill people?
Also one shouldn't forget the communication aspect of driving. Drivers have all sorts of way of showing their intentions: turn signals obviously, but also subtle movement, you may go forward a bit to signal you intend to merge into traffic for instance, wait a bit to let someone know he can pass, etc... New drivers often struggle with that, which may cause more experienced drivers to misread their intention and potentially cause an accident, unless these drivers notice that he is indeed a newbie... or a bot.
One answer is for bots to compensate their weakness with things they are good at. Lightning fast reflexes is a given, but sensors are another aspect. Unlike humans, bots are not limited to a narrow field of vision and visible light. It is an advantage they should use. We humans are insanely good at reconstructing a 3D image from our 2D eyes, a good part of our brain is dedicated to that, computers can't compete, unless the computer is assisted by something like a lidar.
Having been programming for upwards of 40 years now, there is one fundamental truth in programming that all developers need to understand and remember: untested code is untrusted code. Take all the gazillions of corner cases that occur when driving: how many of them do you think are actually tested by the various self driving "solutions" on the market right now? Do they have a test case for a tire coming off the trailer of a transport bouncing down the road? What if a tire is bouncing down a road perpendicular to the one your driving? What if it's on a sidewalk?
It will take decades before self driving cars are well tested. The current approach of falling back to the driver for the "hard" cases doesn't work. Spend some time watching shows like Mayday that delve into root cause analysis of airplane crashes. Accidents happen not because of any one single event, but a chain of events interacting that combine to result in something horrific. This is also the case in nuclear accidents, and in software driven tragedies like the Therac 25.
Any software developer working on self driving cars that hasn't spent a good deal of time looking back at past system failures that have killed humans should bloody well not be working on self driving cars. Makings roads easier for bots is not a solution to the massive complexity of self driving vehicles.
George is bright but not a stable leader. Every month it's some new scheme or feature that will "sell 10x the units!"
No, it won't, not with the lack of support and poor attitude from Comma. If George wants to sell 10x the units it must be a consumer device and treated by Comma as one.
in his own words (from Part 2):
> I’d say that 95% of people I met in the hacker world were good people; only 5% bad. The business world has this flipped. To get ahead in the hacker world you need skills. To get ahead in the business world you need psychopathy.
maybe he needs to be more open and not let his negative feelings about business people worsen his business
Slightly re-phrasing, he said: "95% business people are bad (at helping companies achieve their goals) and 5% are good. The 95% succeed at keeping a job by being psychopaths".
I wouldn't fixate on the word "psychopath". He's not talking about serial killers, just general sleaziness and advancing careers not by being good at their jobs but, say, playing office politics well.
For a concrete example, let's consider that Zoox tweet about organizing Hispanic something event.
It took at least a few people to do it. Someone had to conceive the idea, there had to be a meeting or two discussing it, some exec had to approve it, someone had to execute it and Zoox paid for it, directly for the event and indirectly for the salaries of the people involved.
How much all this activity contributed to Zoox solving self-driving problem? Zero. Nada.
Did it help justify the continuing employment of the people doing it? Yes. Maybe they even got a promotion for going "above and beyond" their duties.
There's an argument that those people are more like parasites, draining the host company of funds, time and attention, advancing their own careers at the expense of Zoox.
For context: this happened 3 years ago.
Zoox still hasn't made a single dollar selling a product. They still hasn't solved self driving.
They got acqui-hired by Amazon for less money that they raised which means that all employees that likely worked there for less money and stock options ended up working there for just less money.
I think part of the problem is that most people in tech are monomaniacs in a multi-disciplinary world. They don't have a good general understanding of what business is, so they make odd generalisations from their own experience (which is usually very narrow) psychoanalyzing people they don't know (anyone who makes such a statement is usually telling you more about themselves than other people).
In tech, most "business people" are bad because they aren't actually running a business but a bureaucracy that has captured a business, and needs to expropriate as much value as possible. The problem is caused by the same narrowness in thought that led to the generalization about 95% of whatever.
It is fairly simple to understand in business terms: these companies are taking money, setting it on fire, and employees are making out like bandits. But this isn't what a business does. The purpose of a business is to deploy and return capital on behalf of shareholders. Again, this requires a multidisciplinary understanding of the world. The world would be a far better place (and we wouldn't be about to go through this next phase) if we had more business people in tech, sadly there are almost none (even running companies, most tech CEOs wouldn't make it past middle management in any other company...their knowledge is just too limited).
If you think this is limited to tech, you're outing yourself as not having much experience outside of tech. The whole world runs on Rules for Rulers, Ponzi schemes, and confidence games.
Lean, scrappy, profitable, and ahead? One negative comment on HN isn't exactly a crisis. A fluctuation in "profitable" wouldn't exactly be a crisis, either, given that the rest of this industry looks like a money bonfire.
> maybe he needs to be more open and not let his negative feelings about business people worsen his business
Neg harder.
I imagine some company will finally get their shit together, and start buying wholesale and installing in their cars as an option. Until that time it'll continue to grow and still be a niche product
Seems like a very strange choice.
personally, i don't think this is a "strange choice" for a company making self-driving vehicles
more like an attempt to try to expand into other markets
The comma body is the first “intermediary” for the new mission. It’s also the pathway to supporting all cars that a human can drive. The endgame of comma isn’t a box that sits on your windshield, it’s a robot person that gets in the driver’s seat. And it will be able to do a lot more than drive.
The mission is self driving, and while we believe it's solvable without the full solution to AI, having the body forces us to stay on a viable AI track and not write hacks. See http://www.incompleteideas.net/IncIdeas/BitterLesson.html
I wouldn't say it's on the level of a PR stunt giving supplies to elementary school students...
The pattern that I have commonly seen is a company getting sidetracked from mission, by something technically cool (especially dangerous with hardware where the costs are high).
> I wouldn't say it's on the level of a PR stunt giving supplies to elementary school students...
Yeah sorry: being cheeky is hard to get right in text -P
The caveats that stood out to me were that he mentions: lidar usage inflating costs, targeting ride sharing as path to profitability, and using VC money to fund non-mission goals prior to profitability. Tesla doesn't use lidar, is already profitable without having targeted ride sharing as the path to that, and is already default alive and not dependent on VC money.
- doesn't rely on lidar
- doesn't rely on pre-mapping areas
- ships intermediary products vs. big bang approach of "100% self-driving or nothing"
- makes money on those intermediary products (vs. loosing money until the big bang)
- while everyone is using Deep Learning for solving the problem, I believe Tesla is betting on DL / AI harder than most
On those big points Comma is exactly like Tesla and unlike almost everyone else.
I'm not sure which caveats you meant.
Unlike Tesla's FSD (Fools Self Driving) contraption that is getting so much hype for it not working at night and still cannot detect the driver's attention properly. Not only they finally admitted that they needed an eye tracking driver monitoring system, that still doesn't even work either.
The FSD / Autopilot delusion created and boosted by Tesla was enough for the regulators to investigate this broken system. I don't see any of this in Comma.
Hotz doesn't share your Tesla hate.
He said numerous times, publicly, that Tesla will most likely win self driving. He doesn't think FSD is Fools Self Driving. He thinks FSD is ahead of openpilot.
Hotz positions comma.ai as Android to Tesla's iOS i.e. a solution that is a little bit later and a little worse than FSD but open to all and better than 3rd or 4th best proprietary solution like cruise or waymo.
I hope you have realised that it is both autopilot and especially 'FSD' which is the problem that needs to be investigated (which it currently is) and each time it has proven not only to be unsafe, but a deceptively advertised product and at most, a dangerous scam.
So where did I say I hated all of Tesla Inc. again?
> He said numerous times, publicly, that Tesla will most likely win self driving. He doesn't think FSD is Fools Self Driving. He thinks FSD is ahead of openpilot.
He is entitled to his own opinions. I believe that unless Tesla complies with the many regulators around the problems I have already highlighted, it is still considered 'Fools Self Driving'. There is too much evidence to list here for the case for more regulations on this.
Anything that is going to 'likely win self driving' must comply with strict regulations and standards in both advertising, safety and the actual functionality. Both Tesla's autopilot and FSD with its repeated safety issues and most critically, the 'driver monitoring system' (DMS) for detecting attentiveness and eyes on the road was at first using the wheel which that has shown to be beyond unsafe and easily bypassed by its users.
Comma is a great example of a proper (DMS) with Level 3 autopilot capabilities which that has eye-tracking with night vision. Perhaps that's why Tesla had to learn from them for improving their DMS after woefully defending using the wheel for DMS and denying that they needed eye-tracking.
So one of them detects if you are using the product and requiring the driver to have their eyes on the road at all times, whereas the other will not work at all, especially at the dangerous time to drive; which is at night.
is there a reason why you don't want offer that?
Pretty much every car maker (GM, Ford, Toyota, Honda, VW) has already invested in at least one self-driving company.
All those self-driving companies have the same technical plan and business strategy: lidar, high-precision maps, it has to be level 4 or nothing at all. They're also decent at selling that plan via powerpoint, tech demos and test deployments.
I think at least 2 years ago Zoox gave a demo to a tech journalist (who produced a slick YouTube video out of it) and you would think that the technology is almost ready to go.
More than a year ago MobileEye CEO gave a talk at a mobile conference and showed a video of a car impressively navigating busy streets of Jerusalem. You would think that the technology is almost ready to go.
Comma, like Tesla, has "no lidars, no maps, ship intermediary products and use real data from real drivers to improve the technology, cheaply".
To me it's an obviously winning strategy.
To 99% or car CEOs it apparently isn't.
Therefore they are not in the market for comma and won't be until it becomes obvious that their approach is failing and comma's approach is winning.
So his position is more a sober assessment of lack of clue in major car maker execs and he decided that instead of playing the game of power point and pretend for the unlikely event of a car maker buying from them, he'll focus 100% on better engineering and winning in the long term by having obviously better product.
The other companies have a plausible path to mind-off self-driving. That's where the value is at.
Comma is developing a product that increases driver's comfort. Waymo is developing a technology that will remove the need for a driver.
What they are selling is an ARM board + cameras + CAN bus adapters for various vehicles. The stock software engages lane centering and longitudinal (gas/brake IF supported) and will keep the car lane centered.
For highway driving, it's a huge quality of life improvement.
Based on the meagre data publicly available, <$5M revenue, $16M funding rounds, 30 employees. It's not very 100x, is it?
I don't debate they're doing interesting things, but the model of "overpromise and underdeliver" is pretty much exactly the same other startups follow, just with larger funding rounds.
But it's a good peak into the mind of GH - not the cosiest of people to hang out with perhaps, considering how intent he is on bashing competitors, but also a really good hacker. I'm kinda shocked he didn't raise more than 8.1m, way more has been spent on less promising minds, especially during the heydays of the late 2010's.
How many is it now? Are you hiring?
Yes!: https://comma.ai/jobs
And one more: https://twitter.com/comma_ai/status/1548030295340769280
I just want to talk shop.
But to give your comment the most charitable interpretation:
Why would they not save up an 8 figure sum after a few years and invest in your company, or start a competitor?
isn't that the case for like 90% of SV companies?
these "mission statements" are so opaque you can't tell what actual product is
often this is used as a marketing catch-phrase to attract hires, eg. "our company has a purpose (other companies don't)"
the way they get around regulators, is they are selling the hardware and the users are responsible for downloading and compiling the software. they support alot of cars but they are utilizing vehicle sensors off the can bus and adas vehicle commands. if the manufacturer decides to encrypt or restrict this then the approach wouldnt work.
For that matter, at a certain point your car insurance might give you a big discount to have one installed and subsidize the installation and hardware costs. There's a profit motive for the first company that does so, being able to offer much cheaper rates while still paying out far less for at-fault collisions.
1. Older cars have a lockout on steering after 10 seconds baked into the EPS firmware.
2. While newer cars don't have this lockout, adaptive cruise control (ACC) is only on high trims. Therefore low trims lack radar and (possibly) continuous duty brake controllers.
It should be better in a year or two when ACC is standard on Ford. ACC is standard on Toyota, Honda, Hyundai, Stellantis, and Subaru, hence that's where the effort goes.
And people say comma is confusing.
Why do they stop using it?
> Nobody will have cars without a safety driver in cities anytime soon.
Did not take that long.
So in the middle of a turn it may just not be able to steer enough and you have to suddenly take over.
I know you have to be aware and be ready to take over in any case (and when I've driven a Tesla it's tried to kill me at least a couple of times every day), but for some reason I really want anything I buy for driver assistance to be able to apply enough force for any turn.
Am I mistaken about that or what's your experience?