My latest article in Forbes.com is about Blickfeld, a Munich-based lidar startup with expertise in mirrors and MEMS. I enjoyed interviewing Chief Experience Officer and co-founder Florian Petit.
“Blickfeld isn’t the only MEMS lidar on the market, but co-founder and Chief Experience Office Florian Petit touts that Blickfeld lidar outperform due to their mirrors. “Lidar faces a trade-off between silicon, where small is good, and optics, where large is good,” Petit explains.
Blickfeld has finessed that tension by developing lidar mirrors with unusually large aperture, which allows them to maintain a wide field of view and high signal-to-noise ratio, while packing a large number of beams into a box roughly the size of an apple.”
At the beginning of every year I predict what will happen by the end of the year, mainly with regard to autonomous vehicles. Then at the end of every year I see how I did.
This year’s predictions are different than those of years past, because this year I am avoiding self-driving car predictions. Now that I am a Cruise employee, I need to avoid any public predictions that relate to Cruise or GM. I tried drafting a set of predictions about the industry and caveating them with some version of “not counting Cruise or GM”, but it got too awkward.
Instead, this year I am predicting a grab bag of topics that are of personal interest to me.
The Chicago Cubs will not win the 2022 World Series.
No states will secede from the US.
The Chicago Cubs will not win the World Series.
Linux will remain my primary operating system for personal computing.
The Phoenix Suns will not win the NBA Championship.
Teddy Bridgewater will not start the first game of the regular season for the Denver Broncos.
The 2022 Winter Olympics will take place.
Joe Biden will remain the President of the United States.
China will not mount a land invasion of Taiwan.
My now-five year-old son will no longer wear a mask to school.
I will not be hospitalized because of COVID.
Republicans will hold a majority in either house of Congress.
I will travel internationally.
The CDC will recommend I receive a 4th COVID booster shot.
I will not use a Mac as my primary work computer.
My primary residence will remain in Burlingame, California.
US economy will not experience a recession.
US will not invade another country.
I will be able to work in my office without wearing a mask.
I will contract COVID.
Kiev will remain independent of Russian control.
I will live more than two months of the year away from my primary residence.
S&P 500 will finish the year above where it started.
Somebody in my extended family will be hospitalized because of COVID.
I will buy a car.
My personal inbox will have zero messages at some point during the year.
I will feel an earthquake.
Average US housing prices will decline nationwide.
Most years (with some exceptions) I forecast predictions in January about what the coming twelve months will bring for self-driving cars. At the end of the year, I evaluate those predictions.
My predictions from January 2020 held up very well at high confidence levels, and then fell apart at low confidence levels. In particular, I wildly overestimated progress on driverless delivery and on international autonomy.
✔ No Level 5 self-driving cars will be deployed anywhere in the world.
✔ Level 4 driverless vehicles, without a safety operator, will remain publicly available, somewhere in the world. ✔ No “self-driving-only” public road will exist in the U.S. ✔ Tesla will remain the industry leader in Advanced Driver Assistance Systems. [DS – I don’t have a good objective source of truth for this. Arguably GM Super Cruise or Comma AI have surpassed Tesla Full Self-Driving Beta. But Telsa FSD Beta still seems to me like the standard again which everyone else compares themselves.] ✔ An autonomy company will be acquired for at least $100 million. [DS – A whole bunch of lidar companies were “acquired” by SPACs in 2021. That is not what I had in mind when I made this prediction, but I guess it counts.] ✔ Level 4 autonomous vehicles, with or without a safety operator, will remain publicly available in China.
✔ C++ will remain the dominant programming language for autonomous vehicles. ✘✔ A lidar-equipped vehicle will be available for sale to the general public. [DS – Some have been announced, but I don’t think any are actually for sale yet. Maybe outside the US?] [Update: Audi A8] ✔ My parents will not ride in an autonomous vehicle (except at Voyage or anywhere else I might work). ✔ Tesla will not launch a robotaxi service. ✘ Fully driverless low-speed vehicles will transport customers (not necessarily the general public).
✔ Waymo will expand its public driverless transportation service beyond Phoenix. [DS – I count Waymo’s free transportation of pre-screened beta customers in San Francisco. Although this is not what I envisioned when I made the prediction.] ✘ A Chinese company will offer self-driving service, with or without a safety operator, to the public, outside of China. ✔ A self-driving Class 8 truck will make a fully driverless trip on a public highway. [DS – TuSimple just barely slid under the wire!] ✔ Aerial drone delivery will be available to the general public somewhere. [DS – I think the Zipline-Walmart partnership counts.] ✔ Tesla will remain the world’s most valuable automaker.
✘ Fully driverless grocery delivery will be available somewhere in the US. ✘ Tesla Full-Self Driving will offer Level 3 (driver attention not necessary until requested by the vehicle) functionality somewhere in the world. ✘ A member of the public will die in a collision involving a Level 4 autonomous vehicle (including if the autonomous vehicle is not at-fault). ✘ A company besides Waymo will offer driverless service to the general public, somewhere in the US. ✘ A company will deploy driverless vehicles for last-mile delivery.
✘ Level 4 self-driving, with or without a safety operator, will be available to the public somewhere in Europe. ✘ A Level 3 vehicle will be offered for sale to the public, by a company other than Tesla. ✘ The US requires driver-monitoring systems in new vehicles. ✘ The industry coalesces around a safety standard for driverless vehicles. ✘ Self-driving service will be available to the general public, with or without a safety operator, in India.
A colleague had to miss a work co-presentation yesterday, so I made some deep fakes and inserted him in the presentation virtually. It was kind of a riot.
I was surprised how much effort this turned out to be – I kind of thought there would be a bunch of turn-key deepfake services to use. Instead, I found a Google Colab and wound up doing a lot of manual video and audio editing, and it turned out kind of comically bad.
Deepfakes still seem to require a fair amount of artistic ability, in a sense.
Nonetheless, it was a lot of fun.
Here’s a kind of comically bad deepfake of my brother Adam spouting something about Python.
The university dropout from Halifax, West Yorkshire, learned to code as a child. As a teenager, he built an app that hit No 1 in Apple’s App Store, and another that was featured in the iPhone’s giant TV adverts. And at 21, he landed a spot at the vaunted start-up boot camp Y Combinator, which brought him to California. He never left.
The autodidact went on to create courses in flying cars and virtual reality for the billion-dollar education start-up Udacity — despite no formal training in those fields. He founded the self-driving car firm Voyage and in March sold it to Cruise, a subsidiary of motor giant GM.
I first spoke with Oliver years ago when applying for an engineering position at Udacity. I didn’t get the job, but a year later I was at Ford, working on self-driving cars, when Oliver wrote me out of the blue. He asked for advice on how to teach self-driving cars at Udacity.
I met with Oliver for a few weeks, suggesting a curriculum and projects. Eventually Oliver recruited me to join Udacity and teach the program myself, together with Sebastian Thrun, the founder of both Udacity and Waymo.
While I built the Udacity Self-Driving Car Engineer Nanodegree Program, Oliver incubated a self-driving startup called Voyage, within Udacity. Eventually, Voyage spun out as its own endeavor.
Years later, I wrapped up my own Udacity work and followed Oliver to Voyage, and now to Cruise.
I’ve had a great experience working with and learning from the lad from Halifax.
Forbes Alan Transportation editor Alan Ohnsman has a short interview with Zoox CEO Aicha Evans that is mostly about vague restructuring at Zoox.
“The company..isn’t making headcount reductions as part of the business review, though some early Zoox employees are leaving, Evans tells Forbes.”
I find the kicker of the article really interesting.
Zoox currently has about 1,500 employees and isn’t planning to reduce that number though some “early, prominent Zoox people” are leaving the company, Evans said, without identifying individual team members.
“We’re celebrating them, we’re thanking them for that first phase,” she said. For some, the company’s initial period was probably more fun than where the company finds itself now. “This phase, it’s a little less sexy, it’s more grindy. But this is what gives you the stripes to actually build a company that ships products and changes society.”
That seems correct to me. Different people thrive at different stages of a company. The folks who are interested in taking on a science project are often different than the people who want to manage the profit and loss statement of a business.
Science projects are also a lot riskier, so if one does pay off, it should pay off big. I don’t know if that happened at Zoox or not, due to their fundraising challenges. But presumably the current leaders seeing Zoox through to shipping product have a good sense of the business’s viability.
Last spring, Facebook published SEER, a new approach to self-supervised deep learning.
One of the core challenges for most deep learning efforts is securing labeled data. The neural network needs labeled data for training, so that the network can learn when it’s right and when it’s wrong, and how wrong it is, and then improve.
Unfortunately, lots of datasets don’t come with labels. The solution is often to pay a third-party vendor to ship the data to a country with low labor costs for manual human labeling. Even in very economical locations, this effort becomes very expensive. And surprisingly error-prone.
Over time, most companies have gotten smarter about how to automatically label a lot of data, but human labeling remains important.
Facebook’s SEER approach skips the labeling entirely, using a “self-supervised” approach to learn directly from the raw data. Instead of labeling different images with “cat”, “dog”, and other descriptors, SEER learns to correlate similar images together. The basic idea is to extract features from each image and then assign images with similar features to clusters.
The second contribution of SEER is an architecture for training a network at Facebook’s scale. The Facebook AI team behind this effort documents their use of RegNets (regulator networks) to trade off compute power for memory capacity, and scale the system.
Self-supervised learning seems like it might become important for robotics, and autonomous vehicles, particularly in the planning pipeline. This is an area in which it can be hard to even know what labels to assign to raw data. If we could instead design a system to let the network learn for itself, that would be a big step forward.
In Forbes, Brad Templeton has a great article on the strengths and weaknesses of radar, relative to other sensors, and why so many companies seem to be interested in it right now. He pegs the analysis on Waymo’s recent radar announcement, but Templeton does a really nice job covering lots of different aspects of radar generally, including how Tesla and Mobileye do (and do not) use this sensor.
Because radar tells you how fast a target is moving towards or away from you, those targets stand out from all the stationary things in the world. You get reflections from stationary objects (like a stalled car in front of you) but it’s hard to tell reliably from all the other stationary things — like the road, the fences, the signs and more. Early radar users had to just ignore any returns from fixed objects, which is why you saw radar-equipped Teslas plow into the side of trucks crossing the road and emergency vehicles stopped in the left lane.