I don't think the video is fake, either. But I do not think it is of much value either. Basing anything just on that is ignorant IMO.
His point was implied not explicit - he believes the fact that the video is still there 12 months later is a clue that internally they are making enough progress to believe that cameras + deep learning software trained by hundreds of thousands of vehicles in the real world is still the solution. The team members have changed but at least publicly the technological approach to solving the problem has not.
With Google/Waymo, MobilEye and e.g. Audi we at least have a decade or so of their work and progress in the autonomous space in the public.
The wildcard question is how much a decade of formal research matters in this field - whether the new approach of distributed fleets feeding training data to a central computer will quickly overcome years of formal research. Maybe it won't but maybe it will. There is at least some circumstantial evidence that Daimler & GM may have come to such a conclusion.
Daimler: is a very proud organization with a deep budget and many, many years of research on autonomous driving. Yet the last 24 months have seen Teslas trounce Benzes on the road. And now for some reason Mercedes announced an Nvidia partnership in January of 2017 to produce a car together.
General Motors: also many years of internal research and tens of millions of dollars invested. Yet the exec team saw something in the 12 man start-up Cruise Automation that was so amazing they laid out $1B in stock and cash to purchase it. Cruise was launched only a year prior by an MIT drop-out and had only 12 employees.
Tesla's approach to solving this problem is brand new because it was not physically possible even 5 years ago and now that it is physically possible Tesla is the only company doing it with a fleet of this scale - it's an experiment in massive deep learning in real time that has never been attempted before. AP1's "fleet learning" may not have been what Tesla claimed (yet AP1 got better over time). But AP2 is uploading actual video.
HD lane and position mapping As for HD mapping Tesla claims to already be hard at work on it gathering data and they claim they have been for months or a year now. Mobileye? - bladerskb says they have a massive mapping project already underway but what I have been able to find says the Road Experience Mapping project will not start gathering data until 2018 - despite the fact that Mobileye signed up partners last year. If the cameras are in customer cars I did not know that and haven't seen an article saying so.
Incidentally Amnon Shashua's only academic paper I can find on driving policy was published in October 2016 - not "many years ago." This was after the highly publicized breakup with Tesla, and at the same time Tesla was implying publicly that Mobileye was basically scared of Tesla's internal development efforts and tried to stop them (Mobileye of course said Tesla was lying). Yet this paper was published
before Intel bought Mobileye for $15B and a lot of people had been saying for some time that Mobileye's curated neural network IP was rapidly becoming less valuable in this dawning era of teraflop monster mobile computers that could possibly "brute force" learning with less human curation needed and far less training time than in the past. It looks to me like a transparent effort to bolster his reputed expertise in
driving policy aka the core of the self driving problem right at the EXACT moment he was "for sale."
https://arxiv.org/pdf/1610.03295.pdf - his paper from October 2016 entitled "Safe, Multi-Agent Reinforcement Learning for Autonomous Driving."
Driving Policy and Tesla Uploads: It does not seem reasonable to believe that Tesla is limiting the use of its video uploads to mere object recognition - it seems far more likely they are working on driving policy reinforcement learning just like Amnon discusses in his paper. The difference between Tesla and Amnon is that while Amnon may have 15 billion dollars from Intel now - he cannot go back in time and build a fleet of 100,000 cars (200,000 by end of 2018 - 400,000 end of 2019) and get their video data to train his network. He doesn't have a fleet. He will soon but as far as I know there is no other fleet of hundreds of thousands of cars running monster hardware (upgradable - 11 tflops today - 20 next year if needed - or more) that continually tests these networks and feeds back video data to train and refine them on a near-continuous basis.
If mobileye has a fleet anywhere near the scale of Tesla's then correct me but I thought they did not yet - that the cameras to do the uploading are coming in 2018 to their newly signed partners.
With others we have actual evidence of their progress over a number of years. There are believable roadmaps, there is history, there is evidence of other things than just recognition etc.
You like history, tradition, stability, organization, written plans, civilized discussions with regulatory agencies lol - you are very European. I mean that in the best way possible.
If Tesla is already doing the same things, then we lack the data that they are. Hence it is more a question of faith at this stage.
Yes, and I have it lol