Prediction that Tesla Will Solve FSD by Mid-2022

Mr Know It All has a projection that Tesla should solve FSD (Full self-driving) to a six sigma level within 275 days. This is based upon the improvement of the FSD system matching to the amount of data from the number of miles driven used by Tesla.

Tesla currently has about 35 billion kilometers of driving data. They are adding about 2000 cars every day and this will increase to 5000-6000 cars every day by mid-2022 and 7000-8000 by the end of 2022. Each car will average 50 kilometers of driving data per day.

Mr Know It All’s projection does not include the increase in new cars per day. This could reach 2500 to 3000 cars per day by the end of 2021.

The projection also has a very conservative starting point in terms of interventions per day.

The main assumption is the improvement is an exponential function that relates to miles of driving data used.

Others Have Projected Early 2022 for Tesla FSD Solution Based Upon the Dojo Training Supercomputer

The Tesla Dojo training supercomputer lets human labellers, label 600 frames in seconds versus days.

Tesla is working towards fully automated self supervised data labelling of driving video data.

Cleaner watt compares the different software driving levels offered by Tesla.

18 months ago, Tesla was spending 70,000 GPU hours or three days to do a training run on FSD. Tesla is training limited for its improvement of FSD. The Dojo system is accelerating the number of FSD training runs to multiple times per day. This is described at 1:34 in the Warren Redlich interview with James Douma. The acceleration of training will help unlock the exponential improvement that could be possible with full analysis of the Tesla driving data.

Tesla Plaid 9.24 second Quarter Mile Average Almost 1 G of Acceleration

Mr Know It All also estimated the acceleration of the Plaid in its record quarter-mile and the use of over half a megawatt of power.

SOURCE- Mr Know it All, Lex Fridman, David Lee Investing, The Future Economy, Warren Redlich, Cleaner watt
Written by Brian Wang, (Brian owns shares of Tesla)

31 thoughts on “Prediction that Tesla Will Solve FSD by Mid-2022”

  1. Waymo and Cruze are more likely to have thousands of robot axis on the road before Tesla has one.

  2. I have mentioned that problem repeated in replies to other articles on this subject. One of the things that must be done is prevent exactly that sort of activity by lawyers. It probably will be difficult to devise laws that can be used to hold manufacturers and owners of self driving cars responsible for actual negligence while still protecting them from lawyers looking to turn true accidents into paydays.

    There needs to be a strong and continued public education effort to pound home that self driving cars won't be perfect, but will be, overall, much safer than human drivers. If enough of the public buys into that, then lobbying for laws mentioned in my previous paragraph would have a chance to succeed. (Lawyers, of course, will work hard to block such laws, but pressure from the public guided by a well designed public education effort might overcome the lawyers.)

    The education effort has to show that self driving cars will make mistakes, sometimes even silly-looking ones, but will be far safer, overall. And they will constantly get even safer as more driving experience is accumulated and used to improve their training.

    It won't be easy, but might be possible.

  3. I mentioned this before: a self driving car that fails (at a very low rate) in the same way that a human would fail is going to be much more acceptable than a self driving car that fails in a way that is seen as ridiculous.

    It's not when the automated billing system forgets to send the bill on time that people regard it as a stupid joke, it's when it sends out a bill for $0.00 owing and threatens legal action when the bill isn't paid.

  4. Oh, yeah. I had a Camry once. Well a Lexus ES300 which is a camry in a pretty dress.
    I didn't want to keep it for very long either.

  5. No, sarcasm.
    No amount of compute can compensate for ML architectures that are simply not up to the task of delivering lvl5. No one seriously believes the other players are only lvl4 because of insufficient data/compute.

  6. Considering that Brian has predicted the victory of McCain over Obama in 2008, that Maduro will be toppled, that Iran would give in to their nuclear ambitions and that D-Wave would be a dominant force in the QC world, let alone promoting the Rossi scam, I would not put too much trust on his predictions

    Just sayin`


  7. He is wrong many more times that simply "sometimes".
    And he is trying to ban me, I am just smarter

    Just sayin`


  8. Did original FSD data and majority of those miles come from vehicles equipped with something other than cameras?

  9. FSD will, or already has, become safer than human driving, especially worst FSD over worst Full Human Driving (FHD: DUI, sleepy, drug-impaired, etc.). But the problem will be that FSD will be prone to unpredictable behaviors, from a typical human POV – e.g. driving blindly into a truck, hitting a pedestrian walking her bike across a street at night, etc. Humans grow to expect certain behaviors from other humans, though when they are wrong, it can have fatal consequences. But with FSD, it won't be clear when it could fail. It could be due to anything, though certainly not EVERYthing. But, we won't know. Was it a unique trick of the light? A lack of lane markers? Or something so unique, it was just never encountered in the billions of samples provided by the incomprehensibly long database that, nevertheless, move the car's AI no closer to generalizable human-like intelligence?

  10. That's all forgivable, everyone is sometimes wrong with something.

    Only that he promised to ban you and didn't comply makes me really doubt of his capability.

  11. Yes, probably.

    I suspect that Tesla will have to completely refactor their code at least once more before they will reach a sufficient level for robotaxis..

  12. Tesla will not solve FSD, ever. A simulated neural network will solve full self driving.

  13. I agree that it will be asymptotically approaching the final state, but we are most likely to find that the final state of the computer optimisation is not the same as the legally and socially acceptable level of driving capability that will enable Tesla to declare "mission accomplished".

    If Tesla is lucky, they'll reach the mission accomplished state before the level that the system will be asymptotically approaching. If they are not lucky, it will be after such a state.

    I'm fairly convinced, for example, that robots will need a MUCH higher safety standard than humans before the various power groups in our modern society all decide to leave them to drive the streets unfettered by red flag laws or levels of law suites that make them unviable.

    Compare, if you will, the level of social acceptance for vaccines that have been shown to have orders of magnitude less negative effects than the diseases they prevent. And vaccines already have many generations to gain acceptance.

  14. Don't worry. Level 5 sdv are many years, of not decades away. And then they're will be a delay at market penetration. I would start worrying about robo Ubers only when highway truck drivers start getting replaced

  15. I bet there will not be true level 5 autonomy (FSD) for Tesla or any company before 2030.

  16. Imagine where we would be If only other players could have afford a "Dojo Training Supercomputer".

  17. The use of an exponential function, or any other function, must be justified by past experience. In the absence of such a justification, the Mr. Know It All makes a two-parameter fit to a single data point. The extrapolation to the future is therefore devoid of meaning. The same exercise, done a year ago, would have predicted 6-sigma FSD today.

  18. I think the biggest problem with computer driving will not be technical but will be legal. Tesla will always be the biggest pocket and the trial lawyers will therefore always go after them.

  19. Very dangerous assumption that you can jump from an automated task to a true conscious mimicking by "improvement is an exponential function that relates to miles of driving data used. Without changing the method…

  20. I don't disagree with what you are saying, but it's a completely different question. My beef is with the assumption of exponential progress…

  21. the issues I think we see here is when is it better than human drivers? and how much better does it have to be? also, given weather conditions are different depending where one lives its likely that the safety level will depend on the area where it is too. Most important and difficult issue i see is interacting with other drivers. Obviously, safe drivers are defensive drivers to a certain degree but if you are too passive you will never make it in certain dense urban areas.

  22. Right you are. You don't see a halving of the error rate when the training data doubles in all other ANN training so why the error rate should diminish in proportion to an exponential function of the miles driven is a complete mystery.

    A 1/log(N), as you mention, is a better guess than 1/exp(N) for sure…

  23. The basic assumption – that progress scales exponentially with the data – is unproven and even unlikely. You don't see exponential progress in other ANN's with respect to amount of data. And even if you would have such a relationship, Tesla would still have to supply the computer power to process all the data. Is it likely that Teslas computing power will grow exponentially? Why should it?

    Sorry, I like "Dr Know it all" as much as the next buy, but he was just simply way of here. Note that Tesla may still solve FSD within a year, but not because of any inevitable exponential growth.

  24. I actually took the timing into account when I recently replaced my used car. I'm hoping to get sufficient use of the new (used) car before the auto-taxis kill any residual sale value.

  25. Shouldn't it be a log curve, as the problem becomes more challenging approaching a limit and the majority of the experience of new cars will cover mostly the same situation or conditions as those that came before them? The challenge of FSD will be to encounter the hard cases to get over the finish line, and that may require a fleet of paid drivers (or disposable cars) who/which, with adequate safeguards, "drive" in dangerous conditions to create the necessary data.

Comments are closed.