Tesla CEO Elon Musk announced today that the company is hiring AI/computer chip team members to develop a Neural Network (NN) training computer called Dojo to process vast amounts of video data.
Tesla’s custom 144 TOPS (Trillion Operations Per Second) in-vehicle inference computer, where almost every TOP is useable & optimized for NN, far exceeds anything else in volume production, given the hardware needed to run sophisticated nets. Dojo, the company’s training supercomputer, will be able to process vast amounts of video training data, and efficiently run hyperspace arrays with an extreme number of parameters, plenty of memory & ultra-high bandwidth between cores.
Back in February, Musk urged programmers and engineers to join Tesla’s team. “Tons of C ++ / C engineers needed for vehicle control & the entire rest of car. Educational background is irrelevant, but all must pass hardcore coding test,” he wrote on Twitter.
Tesla’s NN is initially in Python for rapid iteration and then converted to C ++/C raw metal driver code for speed, which is a very important factor.
Today Musk once again called on everyone with an interest in developing Tesla’s supercomputer to join the team.
During the Earnings Call for Q2 2020, Musk said Tesla is working to move to a more sophisticated AI framework that will improve the capabilities of Autopilot. He explained that what Tesla has done so far is pretty much operating in 2.5D. But 2.5D is not well-correlated in time. “You’re thinking about the world in three dimensions and the fourth dimension being time,” said Musk. He explained that it is really difficult to convey how much better a 4D system would work.
“It’s capable of things that if you just look – looking at things as individual pictures as opposed to video – basically, like you could go from like individual pictures to surround video. So, it’s fundamental …
So, that architectural change, which has been underway for some time but has not really been rolled out to anyone in the production fleet, is what really matters for full self-driving.”
Operating in 4D will be mind-blowingly advanced. Compared to your average automaker’s driver-assist tech, 4D Autopilot isn’t a game changer: It’ll change the entire sport. “The car will seem to have a giant improvement. It will probably roll out later this year. It will be able to do traffic lights, stops, turns, everything, pretty much. And then it will be a long march of (updates). So it’s definitely way better than human, but how much better than a human does it need to be?”, said Musk.
On August 6th Tesla published the patent, ‘Generating ground truth for machine learning from time series elements,’ which explains the essence of 4D well.
Musk also wrote that in 6-10 weeks, Tesla will present a limited public release that will be a major improvement on FSD, and will instantly render a quantum leap in the development of Autopilot. It will not be an incremental tweak, but a fundamental, architectural rewrite. At the moment, the Tesla CEO is using the bleeding edge alpha build in his car, which requires almost zero interventions between home & work.