Less than a week since presenting the progress and achievements of Tesla’s Full Self-Driving (FSD) computer chip to the world at “AI Day”, Elon Musk has defended his comment that the semi-autonomous driving suite’s latest software rollout was “not that great.”
Currently under investigation by the US-based National Highway Traffic and Safety Administration (NHTSA) for alleged crashes with emergency vehicles, Tesla’s advanced driving assistance suite Autopilot and its higher-level FSD, which Tesla wants to make fully autonomous, are under intense scrutiny.
Musk apparently didn’t help matters on Tuesday when he said in response to a Tweet, “FSD Beta 9.2 is actually not great imo (in my opinion), but Autopilot/AI team is rallying to improve as fast as possible. We’re trying to have a single stack for both highway & city streets, but it requires massive NN retraining.”
FSD Beta 9.2 is actually not great imo, but Autopilot/AI team is rallying to improve as fast as possible.
We’re trying to have a single stack for both highway & city streets, but it requires massive NN retraining.
— Elon Musk (@elonmusk) August 23, 2021
This doesn’t sound particularly reassuring on the face of it, but Musk took to Twitter once again early Wednesday (Australia time) to clarify that statement.
The latest comment came in response to one of Tesla’s FSD beta-testers, a group of approved Tesla owners who have been testing the EV maker’s latest software rollout since late 2020.
“v9.2 performing “great” despite @elonmusk claims. This is a 22 min drive. No interventions. (As usual),” said the Twitter user known as “Earl of Frunkpuppy”.
“It’s amazing by most standards, but we are aiming for 1000% safer than the average human driver,” responded Musk.
It’s amazing by most standards, but we are aiming for 1000% safer than the average human driver
— Elon Musk (@elonmusk) August 24, 2021
This in itself gives away the extent to which Musk, and Tesla’s AI team, is aware of how much better self-driving cars need to be than humans for them to be accepted by the broader population.
If self-driving cars were even twice as safe as human drivers there would be less deaths as a result of car accidents.
But it would seem that we fallible humans are 1,000 times more forgiving for deaths caused as a result of human negligence, incompetence or just because someone is plain old “having a bad day”, than we are if a machine “makes a mistake”.
On top of that, a machine does not “make a mistake”. It operates within the parameters of its programming, therefore the programming must be developed to deal with an almost uncountable number of situations that can happen on the road, at any speed.
Tesla is so confident it will soon have solved the AI puzzle that it will also be able to produce a humanoid robot using the same hardware and software stack that can do everything from get the groceries to other more most dangerous, and/or boring jobs.
Watching AI Day, after getting through the initial shock of having to translate uber-geek-speak on the fly, one of the things that impressed me the most was the incredible intricacy of the human brain, and the complexity of what’s being done to match it.
Stepping through the number of processes that Tesla’s AI team have developed using algorithms and hardware so that a car can do something a human does everyday is to put it frankly, mind-blowing.
From how to read pinpoints of data about the curbs and lines on a road and join them together into a birds-eye view, to using eight cameras to combine sensor data to determine the actual location of a moving object, and then label that object as a car, a van, a child or a dog, AND then predict where that object is at any given moment if it becomes occluded (hidden by another object) is just the start.
To do this it has embarked upon a plan to auto-label objects, and use simulations to train its algorithms to handle edge cases. By the way Elon, we still need a “Roo Mode”.
Tesla has also completed building its own AI chip, that it has then packed into a mega-computer known as the “Exapod” that it will use to collect data from its fleet of cars, auto-label the data, use that to retrain and then deploy the upgraded algorithms to the fleet.
Thank goodness MIT AI expert Lex Fridman took the time to create an “AI Day highlights” video to break it down for the layperson.
In the video below he spells out the main takeaways and also gives his top three reasons he thinks what Tesla is doing is exciting.
“The entire picture that was presented on AI Day was amazing because the Tesla AI machine can go through the iterative data engine process of auto labelling plus manual labelling of edge cases so that labelling stage plus the data collection, retraining, deploying and again you go back to the data collection, the labelling, retraining and deploying,” he says.
“And you can go through this loop as many times as you want to arbitrarily improve the performance of the network. I still think nobody knows how difficult the autonomous driving problem is, but I also think this loop does not have a ceiling.
“I still think there’s a big place for driver sensing, I still think you have to solve the human-robot interaction problem to make the experience more pleasant but damn it, this loop … is incredible,” he says.
“The second reason this whole effort is amazing is that Dojo can essentially become an AI training as a service, directly taking on AWS and Google Cloud. There’s no reason it needs to be used specifically for the Autopilot computer.
“You can basically use this for every machine learning problem, especially one that requires scale,” he says.
“Finally, the third reason all this was amazing was that the neural network arch and data engine [oipelines is applicable than much more than just roads and driving, it can be used in the home, and the factory, and by robots in any form as long as it has cameras and actuator, including yes the humanoid form.”
We’ll just leave this here:
Bridie Schmidt is associate editor for The Driven, sister site of Renew Economy. She has been writing about electric vehicles since 2018, and has a keen interest in the role that zero-emissions transport has to play in sustainability. She has participated in podcasts such as Download This Show with Marc Fennell and Shirtloads of Science with Karl Kruszelnicki and is co-organiser of the Northern Rivers Electric Vehicle Forum. Bridie also owns a Tesla Model Y and has it available for hire on evee.com.au.