The going rate for self-driving talent is $10M per person
recode.net
recode.net
While there is a separate department for ML at CMU, many schools have machine learning and/or robotics departments. Suggesting it's the only place to become educated in this stuff is simply false.
Due to the current climate of race-diversity and gender-equality, there is an incredibly high demand for female employees in fields and jobs that females generally don't go into.
This creates a demand for ANY-level talent of that variety - no matter the qualifications or abilities (companies will even start a bidding war over the person).
There are even examples of companies openly bragging about not hiring new employees (or at least making it very difficult to do so) that are male and/or white (e.g., GitHub); or not using suppliers that are male and/or white (e.g., Sam's Club CEO).
It's common to see successful tech startups selling for $1M per head. The extremes go for a lot more money. Nothing new. Way too much clickbait in the title.
Edit: This also brings the other aspect of Western educational system in regards to monopolization resource, where they strictly control the education of AI engrs to control the price per head and market.
The other problem is also the US social environment, social conflicts and 'wars', where it is far from conducive to support the level of education necessary for highly skilled STEM grads at the cutting edge. The environment itself is conducive to a restricted output of qualified engineers, researchers, scientists, etc needed for the pace and expansion. Mind you, a significant amount of these grads are competing for a negligible amount of resource to succeed,and this is the system and environment the US prefers to culture their grads.
It is cheaper in the short term just to steal the skill and discoveries from other countries with promises of wealth, or other questionable means.
10m per head is far, far cheaper, even if CM is the only output 'available immediately'.
+ Particle filters?
+ SLAM?
+ Sensor fusion?
+ Bayesian methods?
+ VC dimension?
+ Sparsity?
+ Nonconvexity?
+ Deep learning?
I think it's gonna be a challenge to do useful machine learning for self-driving cars if you're not familiar with quite a range of math-heavy topics.
It's a difficult, 9 month program where we cover the following (some of which you mentioned):
- Computer Vision and OpenCV
- Deep Learning
- Sensor Fusion (Radar, Lidar)
- State Tracking with Filters (Kalman, Particle) and Localization
- Controllers
- Vehicle Dynamics
We have come up with this curriculum after talking to the heads of Engineering at Mercedes Benz, Otto, and NVidia at length. In addition, we have an open ended section where students can dive deeper into an area of their choosing.
Like you mentioned, a lot of this is Math heavy, which is why we have applications to enroll. Hope that answered your question!
I'm not sure if Udacity is even interested in teaching these types of courses, fwiw... though, i know Khan has lots of math courses.
We do have a few math courses such as our Intro To Statistics class (https://www.udacity.com/course/intro-to-statistics--st101) and our Linear Algebra class (https://www.udacity.com/course/linear-algebra-refresher-cour...) but I'd still recommend Khan Academy or MIT OCW's math offerings. The 18.06 class on OCW by Professor Strang is really awesome in particular.
I would be grateful if you provide good (math) fundamentals. It's much harder to understand how things work if you have to search for it yourself. Suppose you've never heard about a spark (https://en.wikipedia.org/wiki/Spark_(mathematics)), a Hilbert space (https://en.wikipedia.org/wiki/Hilbert_space), Metropolis-Hastings (https://en.wikipedia.org/wiki/Metropolis%E2%80%93Hastings_al...), or a Dirichlet Process (https://en.wikipedia.org/wiki/Dirichlet_process), then it's not so easy to find such concepts and understand their significance.
(1) If you need some help, you can find me here. I'm currently playing with MCMCs in nonparametric Bayesian methods that adapt to the structure in the real world. It is a waste to sample everything. For examples aisles in a supermarket have structure to them. In the "visual grammar" of the supermarket, they are aligned with each other. MCMC that can encapsulate this type of grammar will mix much faster.
(2) If you find a pupil interested in the combination of transfer learning and deep learning, feel free to refer to me. I'm not interested from the viewpoint of domain adaptation, but from the viewpoint of robotic communication.
- Hard real-time
- Fault tolerance
- Safety engineering
- Human factors analysis
The stuff that kills people. An automatic driving project should have someone from avionics design on board.
Also, what will the workload be like? Can this be done on evenings/weekends while holding a full time job?
When Sebastian sent me the new nanodegree heads up I was mildly excited and now as you filled in the actual content I can't wait. With MIT's Underactuated Robotics at edX (which was fun) this is probably the most exciting set of courses I can see on the Internet these days!
Great questions. We limit the first batch for two reasons:
1. This is the first batch of the program ever and we want to make sure it's great before opening it up. We want to use this batch to learn and improve and then focus on larger class sizes.
2. One unique thing about Udacity is we provide real human services throughout the program such as a mentor, a code reviewer, and a career support rep. Many of these individuals are actually current or former students of our program. Since this program has never been done before, we are limited in how many such people we can find. Once we get more students in the program, that pool will naturally expand.
Best of luck and thanks for applying!
Doesn't every CS program everywhere have courses like that? My brother is finishing up a CS at an Ivy, and he's got a bunch of those types of courses.
Now if there's a bunch of courses, there's presumably a bunch of people qualified to teach them. Surely 10M is too high a number?
Alternatively: Being qualified to teach the theory of these courses is not completely equivalent to working on products in this space.
- practical knowledge of libraries (such as OpenCV), and best practices for implementing a robust Computer Vision system.
- awareness of vehicle dynamics and the engineering behind cars.
- sensor fusion and the engineering behind collecting and processing the data a car needs.
- practical knowledge on how to implement a controller and all the required software on a car.
The general point is that these courses give the theoretical background you need but you still need the real practical skills that come with actually implementing these ideas on a real car. I think that's what makes the folks at CMU's robotics lab and Otto so valuable.
Full disclosure, I work at Udacity on the Self-Driving Car Nanodegree program and my knowledge around what skills are needed to be a self-driving car engineer come from talking to Sebastian and the heads of engineering at Otto, Mercedes Benz, and NVidia.
$10M for a truly skilled self driving car engineer doesn't seem that weird in context, since training (well, more like nurturing) one up takes much more time than money.
Also right now they seem to be aiming for "full stack" self-driving engineers, when the future will probably be commoditized libraries and hardware.
How many web developers know the details of TCP/IP?
Sure, we'd design RST controllers, on pages of paper, but I wanted to actually apply that knowledge.
I remember discovering OCW and the first image they showed was magnetic levitation and they had a lab where they had fun. In our lab, we'd crunch the numbers with pen and paper, then see how good our kung-fu was looking at how the system behaved on MATLAB.
Damn you, people with gear in their labs touching things and having fun!
Fun fact: our programming exams were with pen and paper where you'd write programs (x86-PIC ASM, C, Pascal) (you'd better debug it on another sheet before you turned it in).
LOLno, not unless you think all school are Ivies. I go to NC State, all of our CV classes (believe me, I've tried to get into some..) are limited to our ECE department, and even then they are on the graduate level. We do have some AI and data mining classes available to undergrads but none that would be sufficient for the sorts of techniques modern autonomous vehicles use. No robotics/mechatronics available for CS as well.
It's honestly kind of disheartening if you're interested in these sorts of things.
When we had a CV internship listed we got multiple applications per day with minimal promotion. A reasonable percentage of them were decent (experience, working towards post grad, etc). CV is cool shit so tons of people want to do it.
http://www.theatlantic.com/technology/archive/2015/09/self-d...
This being said, having to balance business issues and safety of a system with as long a lever-arm as this seems blindingly difficult. I wonder hos much of the real work here is ultimately more like insurance than engineering ( as if there were any real difference to start with ).
Sounds good to me.
An analogous process is lineman safety in the electric grid; at one point the probability of fatal accident for electric linemen was quite high, and it declined to next to nothing as safety procedures improved.
I don't think it'll really be safe until autonomous cars have transponders and can negotiate space in real time, but then you still have all the legacy vehicles out there that don't have transponders.
It's not like all military contractor engineers live under a constant cloud of self hatred and suicidal thoughts, either, and they build things that are about trading "many" "enemy" casualties for "few" "friendly" ones. By contrast, self driving is doing almost entirely good along any moral spectrum.
Behind the house, the second most expensive purchase in a typical person's life is their car, and yet in a car's lifetime, only 2.56% is spent driving on roads with the remainder spent parked or in traffic.
https://medium.com/self-driving-cars/how-to-land-an-autonomo...
The author is also involved in Udacity's new self-driving car nano degree: https://news.ycombinator.com/item?id=12521832
Do you know how to find the actual article on the subject?
*edit:
Found it! Google is a much better way to search medium than medium's worthless navigation/search.
Here are the three articles:
https://medium.com/self-driving-cars/how-to-land-an-autonomo...
https://medium.com/self-driving-cars/how-to-land-an-autonomo...
https://medium.com/self-driving-cars/how-to-land-an-autonomo...
At the very least I'm guessing that fresh employees have more leverage than they think. If you think 1kk is too much maybe 100k? Or maybe ask for them to pay your rent while you work there?
I have been self-teaching myself neural networks, genetic programming and algorithms and AI since the 80s. I remember the 'Decade of the Brain' the 90s and reading Patricia Churchland and Terence Sejnowski's book 'The Computational Brain'. I was also a welder building motorized and pneumatic and hydraulic animiatronics in the 90s. I started to go deep on the engineering, and it helped a lot, but there was a guy I worked with who commanded the time and space and materials in front of him, and had a gut feeling on how to put it all together.
Systems integrations is important, but interative and incremental design, also familiar as a design methodology in coding is the way to achieve results. This is because the individual engineering of subsystems, and the subsequent computer modeling fall short of the emergent behaviors of a real physical prototype.
If I were hiring, I would not be scouring Udacity or the Unis, but lone wolf garage engineers and tinkerers with the math aptitude too. Two of the successful companies I worked at started in somebody's garage, and both were not college educated. Find people who have managed to somehow put together 60% of what you're looking for and then fund them and set them loose.
Too many of the engineers I've worked with were great with churning out the stuff they were taught, but in the one-offs, or bespoke, which seems to be the fashionable word nowadays, they failed miserably with 'paralysis through analysis' too much analysis.
This is why when I had my own business in the early 2000s, I lamented the death of the trade school in the U.S. It was very difficult finding young people who could actually build stuff. The maker movement is welcoming, but a lot of it tends to be mechatronic, and high tech. You don't see too many 'makers' nowadays that are capable of fabricating without a 3D printer, or making heavy-duty iron mechanical monstrosities like some 'Junkyard Wars' aficionados were building for a while.
This is also why it is hard to find people to work on the BIG projects like tunnel-boring machines (well this has picked up somewhat), or other big equipment. Now imagine finding someone who can also design compliant controls and mechanics for all of this. Certainly worth $10M per person!
I am not calling for a wild west mentality, but the opposite of trying to legislate intelligence is a sure way to stagnation and oppression of creativity, passion and possible profit.
I was making blackpowder from my hobby store-bought kit in the 70s, and playing with all sorts of chemicals. My mom & dad did not discourage me. I had to setup my darkroom after they went to bed, since we lived in a railroad apartment. If my Dad got up to go the shared hallway bathroom, there went my negatives or prints!
I went on to melt and cast aluminum, solar ovens, tesla coils, lab mice, the whole shebang.
Anyway, I am optimistic by some of the things I have seen kids do on YouTube in the US, but some of them seem like they have been funded by a huge corporation (VIDEO: https://www.youtube.com/watch?v=92M5qcjDkaU). I give my kids as much support as I can, but even they will not accumulate the materials present in that video even with a singular-purposed feat such as that.
My son had told he might not have been able to bring his science project to school, a home-made cathode ray tube, for safety reasons even though the teacher knew it didn't emit harmful Xrays at the power and configuration of the setup. He finally proved it, and was allowed, but there were other 'mysterious' concerns about such devices and the local police.
Now I have lived in SE Asia for almost 8 years. Chinese farmers/tinkerers are building home made submarines to harvest ocean bottom sea life that are so much a part of Chinese cuisine. That and trike planes, basically tricycles with a hang glider and a huge propeller on back - sort of a real hacked ultralight you would see in the US. The scene is surprisingly large in the light of such a heavy-handed government.
Even now where I am in East Java, people hack together some weird stuff, and the police don't stop them or pull them over - imagine a train of dollies being pulled by a small 125cc scooter on a secondary road with passengers! Scary, but I am glad to be around creativity and people using their wits to solve problems with constrained resources.
Uber is valued at $60bn, has 6000 employees -> $10M per person
Google is valued at $500bn, has 50000 employees -> $10M per person
Can we get a fact check, YC?
from the perspective of the buyer, this is all part of the cost of the transaction
But that's not likely.
From the point of view of a buyer who spent $100M and got 10 engineers out of the deal, the cost was $10M/engineer, whether the engineers each got $10M injected into their pockets or not.
That's not even so: he didn't buy the engineers, they are not his property and they are free to walk away the day after, so he hardly got them. For a day, OK :-)
In acquihire-land, that's the short-hand terminology that is often used. I'm sure there are more precise (and lengthy) ways to say the same thing.