Tesla AI Day 2021 featured executives presenting the full range of the company’s artificial intelligence efforts, from computer vision to planning and controls to simulation to data infrastructure to automotive super computers to data center supercomputers. Elon Musk capped the event by announcing a humanoid robot that Tesla TSLA+0.9% is developing.
For example:
Tesla optimized the training nodes for neural network computation, with a focus on parallel matrix multiplication. And the team custom designed the D1 chip with 7 nanometer silicon etching technology.
Tens of thousands of D1 chips come together on a pizza box-sized “training tile”, which has 9 petaflops of computational power.
One million training tiles will make up the ExaPOD, Tesla’s data center supercomputer.
Baidu Co-founder and CEO Robin Li and CCTV host Beining Sa sit in Baidu’s newly launched robocar
Baidu World 2021 showed how central autonomous mobility is to the Chinese Internet giant. The company, which operates the most popular search engine in China, has dedicated immense resources toward next-generation transportation, first and foremost through its Apollo self-driving car program.
The three hour Baidu World presentation kicked off with CEO Robin Li unveiling the company’s new Robocar. The vehicle has two seats, no steering wheel, gull wing doors, and windshields that double as computer displays.
The vehicle has the feel of a concept car, rather than something that might ever hit real live streets. But the presentation quickly cut to a live demonstration of the car transporting Baidu employees in a suburban-style environment. The number of Robocars in existence, and their true capabilities, are unclear, but they are real enough to at least handle a demonstration route.
QCraft is building and operating autonomous shuttles in Suzhou, Shenzhen, and Wuhan. The company hinted at expansion into point-to-point or even robotaxis models.
Meituan, the lead investor in this round, is China’s largest food delivery company. Presumably they are interested in the application of this technology for that purpose.
I was just talking yesterday with a Cruiser who is currently located in Pittsburgh, because his girlfriend is in grad school there. I mentioned that Pittsburgh might have the best combination of amenities compared to cost of living in the United States.
PNC Park rivals Chicago’s Wrigley Field and San Francisco’s Oracle Park as the best stadium in baseball. Carnegie Mellon has ridden the strength of its computer science department to become of the country’s premier research institutions. The topographic location of the city, at the confluence where the Allegheny and Monongahela Rivers merge to form the Ohio, is spectacular. Plus there’s a funicular.
All of that makes Pittsburgh a first-class city, but what really sets it apart is the cost of living, which is half that of San Francisco. Cost of living in Pittsburgh is actually comparable to cost of living in Roanoke, Virginia.
And what is really striking are the employment opportunities. In addition to the old industrial behemoths like US Steel and Heinz, and alongside UPMC, Pittsburgh has built a robust technology ecosystem. The Post-Gazette just highlighted all the Pittsburgh technology companies going public: Duolingo, Aurora, Argo, Cognition Therapeutics, Stronghold Digital Mining.
Google has long had an office in Pittsburgh, along with many other companies seeking to hire Carnegie Mellon students. Waymo just opened an office, and Motional has a big presence.
I wrote last week that “driverless shuttles seem to be having a moment.” Today, I wrote about an announcement from May Mobility that seems to cut against that trend.
I confess, the economics of autonomous shuttles have never been obvious to me. On one hand, public transit generally amortizes its cost over many passengers. On the other hand, the benefit of autonomy is usually seen as the cost of removing the driver from the vehicle.
But if the cost of the driver is borne by many passengers, then per-passenger economic benefit to removing the driver would be small.
MLive actually provides some interesting data to crunch. First a disclaimer: with the exception of the MLive data I cite, I am making up all the other numbers here. I could be way off. Just a thought exercise.
Pre-COVID, ridership of the May Mobility autonomous shuttles in Grand Rapids, Michigan, was 7,000-11,000 per month. Let’s average that to 9,000 riders per month, 108,000 per year. The total cost of running the service for the year was about $900,000, paid by a combination of the city, private donors, and May Mobility itself. That’s about $8 per ride.
The new on-demand, point-to-point ridehailing program in Grand Rapids will have five May Mobility vehicles running at a time. I don’t see how many shuttles the old program had running at a time, but let’s say two. Let’s also imagine the program ran for 12 hours per day, 365 days per year, which is 4,380 total hours.
At a fully-loaded cost of $40 per hour (total guess, municipal wages tend to be low but benefits tend to be generous), that’s $175,000 to cover the cost of the drivers for the year. That’s about $2 per rider.
$2 per rider is something – in fact, it’s probably about the cost of a bus ride in many US cities. But it’s also a pretty small share of the $8 per rider cost of the autonomous shuttle program.
Probably the $8 per rider cost includes expenses for a vehicle safety operator, who presumably costs much more than $2 per passenger but also hopefully will eventually become unnecessary.
Still, even without the safety operator, the pilot program probably costs $4-$5 per ride.
But I suppose the end game is that over time the rests of the costs will tend toward zero, and you can only move the driver cost toward zero with autonomy.
May Mobility announced that it is advancing its transit program in Grand Rapids, Michigan, to an on-demand phase. Customers will be able to hail May Mobility SUVs at “more than 20” pick-up points in a 1.36 square mile area of downtown.
I have mostly associated May with low-speed shuttles services for public transit. Covid-19 has heavily reduced the demand for public transit. May initially did roll out a low-speed, shared shuttle in Grand Rapids, pre-COVID. MLivereports that ridership dropped over 70% after COVID hit.
The transition to on-demand SUVs adds convenience to riders, at the cost of complexity for May. A low-speed shuttle traversing a fixed route isn’t especially convenient, but it’s much easier to implement than an on-demand ridehailing system.
May’s move to ridehailing represents a kind of intermediate step between shuttles and robotaxis. Riders can’t hail an SUV just anywhere, they have to go to a pickup point. That also greatly simplifies the implementation of the system.
Bazel is a build system that Google created under the name “Blaze.” Google open-sourced the tool as “Bazel”, an anagram of “Blaze.” The tool is an alternative to Make, CMake, or Maven.
Bazel’s particular point of emphasis is the “action graph” which maps source inputs to intermediate artifacts and through to outputs. This allows Bazel to reuse components that haven’t been updated since previous builds. Bazel’s graph structure is particularly effective at this type of caching.
Bazel also works well as a remote-execution tool, allowing engineers to build either locally or on remote execution servers.
Combining Bazel’s memoization reliability with its effectiveness at remote execution makes Bazel a good choice for large organizations and distributed teams.
Build tools are rarely covered in software engineering courses and curricula – we only lightly covered them in most of the Udacity courses I taught.
But build tools turn out to be surprisingly important in day-to-day engineering work. Choosing the right build tool can dramatically accelerate a team’s efficiency, and learning to use that tool can make an engineer much faster and happier.
LeddarTech and Cognata partner to offer integrated sensors and simulation. This makes so much sense it seems obvious, in retrospect. Buy a suite of sensors less because of the sensing technology itself, but rather because the vendor can also provide hyper-realistic perception simulation for those sensors.
BMW invests an undisclosed amount in Kodiak. This happened in June, so old news, but it pairs up with Bridgestone’s investment of an undisclosed amount in Kodiak earlier this year. Kodiak seems adept at quietly scooping up (presumably) small amounts of funding. CEO Don Burnette says BMW’s investment is “purely financial” and there is no technical collaboration.
Quanergy announced plans to SPAC at a $1 billion-plus valuation. Remember Quanergy? They were “the future” when I first got into self-driving cars in 2015, with solid-state phased array lidar. Then things fell apart. They’re back, and the new CEO says the immediate demand comes from IoT, not automotive. The SPAC news is from June (I’m catching up on old email, sorry) and nothing seems to have happened since then, so the certainty of this SPAC merger is unclear.
Robotic Research will launch a driverless shuttle at US Marine Corps Air Station Miramar. Miramar was the setting for the original Top Gun movie, although the real-life TOPGUN program has since relocated to scenic Fallon, Nevada. This news has circulated for a while. I mention it here because (a) I’m catching up on old email, and (b) driverless shuttles seem to be having a moment, perhaps because they’re a relatively easy, low-cost pilot program to get up and running.
Apropos of “driverless shuttles seem to be having a moment,” the nation’s (world’s?) largest driverless shuttle network will launch at Colorado School of Mines, outside Denver. Nine EasyMile shuttles will traverse the campus. Unclear if a safety operator will be onboard, or if these shuttles will be “fully driverless”, although I assume we’d have heard if they were to be fully driverless.
Yet again in driverless shuttles, over in China, Apollo is upgrading its Apolong minibuses to next generation hardware. The minibuses operate in 22 environments and have traveled 120,000 kilometers (80,000 miles).
Motional, which started life as Boston-based nuTonomy, and later expanded to Singapore and then Pittsburgh, is expanding its California presence. The Santa Monica office will grow and a Silicon Valley office will open. Most significantly, though, self-driving cars will start testing in LA. I don’t write much about Motional, which tends to be relatively below-the-radar, relative to similarly-sized AV firms. But it’s worth noting that the now-ubiquitous Point Pillars algorithm for object detection with lidar came out of Motional (then named nuTonomy). One of the more significant published research accomplishments by an AV firm.
Cruise just released a video called Poppy & The City, which is a real ode to both autonomous vehicles and the city of San Francisco, Cruise’s home.
There are so many Easter eggs in this video. I wasn’t involved the production of the video, so I had so much fun watching and picking things out.
[0:02] Since I joined Cruise during the pandemic, I haven’t actually visited the various Cruise garages. This looks like maybe the Cruise facility at 1201 Bryant Street, San Francisco?
David Silver 