Elon Says the Only Way to Solve Self Driving is to Solve Real World AI

Elon Musk says the only way to solve self driving is to solve real world AI. You have to have cameras and silicon brains match human vision and biological brains.

Elon still thinks Tesla will solve full self driving this year which means Elon thinks Tesla will solve real world AI in a meaningful way.

Elon says there are now over 100,000 people with full self driving beta. This means almost all people who have bought full self driving in the US will get the beta this year and all new buyers of FSD will get the beta immediately after general release.

Tesla seem close to being able to have all the features for full self driving and greater safety than the average human driver. Being able to offer full self driving for all paying customers means that instead of recognizing 50% of full self driving revenue, Tesla will be able to recognize all of it. This would be about a $600 million increase in net income if it happens for Q2 2022 and it would mean an additional $200 million in net income and growing in each following quarter.

Solving real world AI for self-driving is and four wheel cars is what Elon believes can be generalized to walking humanoid robots. Actuators and others parts of humanoid robots have been demonstrated by many but they did not have the intelligence to move about the real world with minimal instructions and do useful and valuable things. There is also the missing aspect of mass production which is something Tesla is very good.

Written by Brian Wang, Nextbigfuture.com

53 thoughts on “Elon Says the Only Way to Solve Self Driving is to Solve Real World AI”

  1. If you watch any section of Musk talking ever its clear he is always in full fake it till you make it mode.

    People really do give this guy too much credit and then some.

    Do they have good engineers working at Tesla and SpaceX?


    Does he even match up to the worst of them?

    I strongly doubt it if he truly expects them to have FSD viable without human intervention this year.

    This cult of personality need to realise that Musk is basically just Steve Jobs with a slightly better understanding of engineering, but still has little more than surface level knowledge of what he is talking about most of the time.

    I see it all the time where people in comment sections desperately fill in the blanks with their own knowledge where he stumbles through his press events because they feel he must have meant x, y or z – when in reality he meant "please stop listening, I don't know enough about this to keep talking and not make a fool of myself or endanger the stock price".

    That's not to say he isn't a truly gifted salesman – he is certainly that and then some, all the best of them know how to fake greater depth than they have to work with and people lap it up from Musk with a soup ladle.

  2. Hhahahahha, that's a good one.

    I saw a video less than 2 months old where they barely started filming a demo video for FSD and the car immediately swerved into guy on a bike – the guy jokingly said they would cut that out, I'm glad he did not.

  3. Yes because I'm sure if an attack or solar event wiped out GPS we'd all just be jumping in our computer automated metal death traps for a joy ride……. /s

    GPS at this point is literally battle tested – the satellite system was built to be far more durable than most commercial space ventures precisely because it was designed for military use, and you know they pay a very hefty premium for long term durability.

    Also just becauase Waymo et al are currently using such an approach does not mean that they are limited to it – after all it was spun off from the same company that owns DeepMind, a team that has at least as much expertise in AI systems as anything Tesla can field and then some.

    Not all company projects are telegraphed in advance and I would expect that Google/DeepMind/Waymo have been working on this for a long while – the difference is that they are not stupid enough (as Musk is) to put it out front and center in beta form for the public to see while its failures are apparent so that public confidence in FSD technology can be eroded before it even goes to production.

    Musk's grandstanding for his investors with FSD beta (to generate hype and boost stock prices) is only serving to erode public confidence in this technology – I can't count the number of times I've seen people express doubts about putting their lives in the hands of a computer driven car for exactly this reason.

  4. You're an idiot if you don't think that robotaxi revenue from achieving FSD is worth that effort.

    Not to mention that the task of managing a humanoid robot control system is drastically more complicated than FSD – stating that robots was the reason at this point is like getting to the top of Ben Nevis and then saying you were training for that so that you could climb Everest.

    The 2 tasks are not even remotely in the same league of potential difficulty.

    Also solving FSD for cars does not automatically translate to solving it for trucks or long buses which are especially difficult to navigate through more tight and cluttered urban road plan settings.

    Where a car can easily manage such urban scenarios it becomes far more difficult for a truck to manage safely with its greater length and bulk, and the consequences of a screw up are vastly worse.

  5. I enjoy watching this unfold. I've been following OpenAI, GPT 3, Dalle 2, and the imminent release of GPT 4.

    AI has reached an inflection point: it is no longer about the size of the data set and brute force computing. Though those are still important, the learning parameters and efficiency of the modeling are becoming the main factors driving accuracy. I think that is what Elon is trying to express.

    "General, real world AI" is a total trap – what does that mean? That kind of broad, imprecise word usage will cause endless argument. The opposite of precise language. Semantics manner.

    Anyone who argues about such things are wasting their time. Elon is only focused on one thing with his AI right now – navigating the physical world of roads and highways with an electric vehicle. How that translates to how a biped robot makes its way in the physical world is what Musk is referring to. Navigational headstart.

    That's it.

  6. This will be what kills us, a self inflicted euthanasia due to robotic companionship.

    Assuming the coming incel/furry war doesn't kill us first.

  7. Actually, I think maybe in the past they did offer some training time on their oldest Dojo system, before it was retired?

  8. If you have good EO/SAR satellite coverage with very high revisit rate to update maps (something Google is starting to approach), along with high accuracy civilian GNSS signals that are being rolled out for GPS/Galileo/Beidou, that certainly would help the self driving issues from a route planning standpoint.

    LIDAR (and point cloud radar) is also a big help for self driving as well, when you can't geolocate or trust your maps though.

    The basic complaint seems to be dual camera forward vision with existing camera sensors is either sufficient for most self driving needs (including ranging/depth perception), or it isn't, with respect to sensor buildout cost per car. LIDAR still hasn't come down in cost due to lack of production. If the sensor makers could get over that hump, then the fight is on. Until then though, the pure camera approach (Tesla is apparently no longer attaching their radar units, so no point cloud either) is the cheapest, though are they achieving sufficient depth perception with only single side and rear cameras?

  9. It is true only in the concept, but since their maps are going to roll and update ever more quickly, they will be able to cover the entire global road system in the foreseeable future.

  10. Tesla at $10: It's a failure!
    Tesla at $100: Can never compete!
    Tesla at $1000: It'll be over any day now!
    Tesla at $10,000: It's a scam!

  11. Brian, I’ve loved your website for years. But it does seem like you hold stock in Tesla. If so could you please include a note on this at the bottom of each bullish Tesla article?

  12. Also more importantly, capable of operating after an attack or solar event which disables GPS. Maps are only useful if you know where you are. AI can work out where you are.

  13. (maybe You know about) Are there plans supporting other car manufacturers (or their customers) through insurance compliant hardware upgrades for their production lines?
    Would this enable public subsidy?

  14. Musk says it’ll lower rates and Tesla is looking into becoming an insurance provider to give quickly give an appropriate rate for self driving cars

  15. Solid murder muscle, covered in the right combination of fluff that makes them appear cute to their prey. Such weird animals.

  16. I also hope that Tesla will offer training as a service. Intel is talking about 10^21 calculations double precision computing in 5 years. Teslas 8 and 4 bits of computing is 64 and 256 times easier by comparison, so we could be hoping for 10^23-10^24 operations per second on Tesla HW in 5-6 years?

    Which would mean that we could all use their training HW to implement all our ANN dreams cheaply and quickly. It could be something similar to when both Android and Apple opened their devices for outside apps, where we saw an avalanche of new apps and SW. Say, a thousand bucks for 10^24 operations? Or ten bucks for 10^22 operations? I'd buy it in a heart beat..

    So, it could earn Tesla hundred of billions of dollars per year and usher in a completely new era in ANN usage and development. A more democratic use of ANNs, where google and open AI would not sit as gatekeepers of fantastic SW, but it would be open source.

  17. About the computing estimate. Granted, we do not know every detail, but we know enough to see that 3 million operations per neuron is an exaggerated upper bound.

    We know – not guess – that the upper frequency is 300 Hz. We know that the maximum connectivity is 10k synapses per neuron. We furthermore know that a synapse is either firing or not (no "analogue" firings can occur), so that would actually be one bit (not one byte) of calculation per synapse. And we also know the number of brain cells of a bee.

    Do you expect any results upending this estimate? Save the discovery that neurons may be working on "half" the input using different sides of the neuron, no counter evidence to this estimate has emerged in last 40 years. Or do correct me.

    The reason why we have not been able to control a worm depends on some very simple reasons:
    (1) Why should we? A worm doesn't really display any interesting behavior
    (2) Achieving a suitable neurological interface is hard

    So, almost no reward for succeeding in controlling the worm, but a large development "cost". It's not surprising few (if any?) research teams are working on a neurological control of a worm…

  18. Intention to ask, for overview, about how pre-autonomous driving is implemented into societies regulations and cultural habits.
    Difference is therefore mainly between situations of a car being driven and vehicles providing transport to passenger(s).
    A robot driving a car is level 5 (NHTSA's roadmap is >2025 for fully automated safety features and 'future technology' for describing autonomous transport, in terms of 'automated driving systems'. They avoid a non-industrial term 'self-driving'.).
    Not clear to me, where Real World AI is placed within that context of responsibility (drivers or passengers interaction with decisions on (Real World) AI).



  19. I'd be willing to put up real money that you are wrong about how far away FSD is, but let's not dwell on that subject.

    Kindly, stop saying that Elon may be a con man. Irrespective of the feasibility of FSD with current HW and SW-tools, there is absolutely no indication that Elon Musk is actively duping anybody regarding FSD.

    Everything, from the number of people that are working on FSD at Tesla, working at DOJO to his sincere interview where he doubles down year after years on FSD, indicate that at least he believes FSD will be possible with their current approach.

    So stop adding the "or conman" to the list of epithets that you tag onto any sentence with Elon Musk.

  20. Elon makes it progressively clearer why the shift to humanoid robots. Without a higher challenge like that the massive effort to build custom silicon and Dojo to train NN software would be plateauing now as it achieves Robotaxi FSD. It would still have to implement it in different countries but it would be a solved problem and the expensive team would start to fall apart. It would mean Tesla’s growth and ambition would also plateau something like Apple after iPhone, coasting over the top of the S curve. Humanoid robots that are drop in replacements for human labor are vastly more ambitious and transformative but are just the next challenge for the team. 

    There will be HW-5, 6, 7 and Dojo will become Cloud computing AI with iterative improvements and a steadily growing budget. The market for humanoid robots and new skills for them is essentially unlimited – and they of course can work in the GFs that make them by the tens of millions.

  21. A peculiar take on I’m confident we will solve it this year. The Tesla humanoid robot emphasis is what confidence that this is very much the right approach looks like.

  22. No, it’s not. Their approach using LIDAR and hidef maps is like creating software rails. There have been fully automated passenger transports that used dedicated tracks for decades. That’s not solving self driving. Their systems aren’t even close to being able to scale. Tesla’s is already mass produced as well as capable of operating anywhere without detailed recent scans of every route.

  23. There's another word for a cat the size of a teenage girl. Leopard.
    Keep those death nightmares far away from me.

  24. I gave up looking into self-driving automation issues, especially, the legal and standards making processes thereof, once I understood that it is a near-general-AI equivalent problem. Which Elon just admitted himself.

    Once I understood that (and that happened more than a decade ago), I stopped caring, because I know that all these self-driving level 4 and 5 standards will not be fulfilled any time soon.

    NHTSA is supposed to be looking into all that. So I'd start reading up on that, if you're interested in regulation and standards effort in the US. But so far, anything beyond level 3 is too complex.

    The only realistic way, currently, to implement "level 4+" self-driving is to limit cars to controlled environments, aka tunnels. The rest of our dreams requires revolutionary understanding of intelligence, which we don't posses now.

  25. Don't overreact. If you watch that section, it's clear he meant it narrowly. He's believes that the new way they're handling visual data is very cool and can be thought of as a sort of visual system AI. Also he thinks it will be available and much better than the current beta this year.

    It's sort of nice to hear him explain that they've moved beyond processing visual data as still images, but I think we all sort of figured that anyway. This isn't even news.

  26. "So is Elon Musk delusional for believing that their current HW + system optimization could reach FSD level?"

    For level 5, I'm willing to bet real money, yes (certainly for any year, during this decade).

    For a very limited level 4 (in highway conditions, in sunny states, limited to 45 miles/hour), I'm willing to give him the benefit of doubt that in 8-10 years he might. He will still find it challenging to find underwriters for those cars. So, chicken-and-egg issue for that as well. It might be that Waymo's approach is actually better for the limited level 4 situation. Again, *very limited* level 4 — and even that I'm not sure.

  27. We do not know yet how bee's brain works. We don't know all the degrees of freedom. How the activity is really organized. We don't know how all those things fire, because even the best MRIs produce disjointed time slices of activity — in quite a limited setting at that (ever tried to put an animal into an MRI chamber?). The simple fact is that what we have are some mathematical/statistical abstractions and very powerful hardware. We cannot even fully simulate a worm. Find me anything online that shows that we can completely replace the brain of a worm and keep it operating as if nothing changed.

    At most we've been able to replace and understand the functionality of a particular blob of neurons, and modulate their activity to some effect (Neuralink). We're a long, long way from translating these cool but small results into how we operate a practical AI that makes sense of object space and actuates based on that.

    Robin Hanson has been conjuring up brain emulation for decades now. Even building entire civilizations with it (in his imagination). Isaac Arthur has a lot of videos that assume that this will be the case, too. Many people believe that all this is just around the corner. Any maybe it will happen… Someday. But I'm willing to bet that it's not within this decade or even the next two. Not even for a worm.

    Elon is a dreamer, at best. A conman, at worst. Either way, if the stock price of Tesla is mostly dependent on the fruition of that dream soon, then it's toast.

  28. So is Elon Musk delusional for believing that their current HW + system optimization could reach FSD level?

    My guess is that they will make it with current HW, but I open for the possibility that HW4 would be required. But is so, retrofits are possible.

    Please note that it has been shown that when you train a large model – and the FSD is a "largish" model – you may reach a state when you see no improvement on you training loss function for many epocs, but if you persist in training your model, the loss function can first increase (bad) and then reaches a lower state (good) where all evaluation metrics are also improved and the model is more generalized.

    So, once they have reached an "OK" state, they should run an enormous run with millions of epocs and a lot of data to reach a better state. But it's no point as long as they can progress faster using model improvements and short trainings of parts/whole model.

    Such long training bout will be predicated in implementing DOJO effectively. And, it's also predicated on having more data, which in turn is hinging on reaching semi-automated labeling. Elon stated that they are not done implementing the latter in this very speech. (So still implementing ideas from AI-day).

    Clearly, just training longer with more data is not "a revolutionary" step in ANN. Implementing semi-automated labeling may or may not be characterized as "revolutionary".

    Time will tell if you or Elon Musk (and I) are delusional.

  29. So we're no way near solving this. Not ontly that we're not gona arive there gradualy, by bettering altgorithms, throwing exabytes of data or exascale computational power, we are on a totaly wrong aproach acording to Isabel Millar in her book "The psychoanalysis of artificial intelligence"

  30. OK, so we at least agree about what Elon has stated. Not general intelligence, but real world perception and acting in such a world.

    First off, Elon is not admitting that they will never make the FSD work, because he reiterates that it will be safer than an average human driver this year. Does that sound like an admission that they will never make it?

    And about reaching the intelligence of a bee… You may think that this is delusional, but both Elon Musk and I disagree. The typical bee has about 1 million neurons with at most 10 billion synapses (probably a lot less). If firing at 300 Hz (maximum frequency for a neuron), this would equate to at most 3 trillion neural updates. The HW3 in Tesla cars can utilize up to 140 terra-operations per second, so clearly the HW could – if well programmed – reach the level of a bee.

    Where are they now? They are still implementing the ideas that they brought up in Tesla AI day such as operating in vector space, having both a spatial and temporal memory. And of course, having a single stack.

    Even without these "low hanging fruits" they seem to have a system that goes – on average – many miles without interventions. The latest version 10.11.2 seems to be universally praised by the users. Note that this is still a dual stack version.

    Is this "far" from achieving a the level of a bee? Hard to tell, really. My guess is that just implementing all the ideas presented during the AI day would probably get them close or over the threshold.

  31. me, think, engineering team around Elon Musk is very capable in solving parts of a FSD system, that will be able to support average human drivers up to a level where human interaction is reduced to the most possible. Problem will stay for to take ( that full level of ) responsibility for average human drivers? (are there public, enhanced discussion results to that point, backed by governments and being internationally standardized, within technical advancement ) ?
    One might think of areas only supporting, for e.g. ГЛОНАСС, नाविक (navik), 北斗卫星导航系统?

  32. What's the impact on legal regulation for all insurance customers (within all situations), if this level of FSD is (not) activated through average human drivers?

  33. maybe, He should be more leery about enabling highly capable Real World AI within fake reality (show era)?
    maybe, He should give tasks for multi viewed studies to Universities also for relating own (biased) positions ( e.g. danger of Real World AI and real time manipulation of big data )?
    ( definition for boring tasks is pretty much dependent from some point of view, purpose and aims, thus not really (always) that simple? )

  34. Jan, I got your point. Now, please don't misunderstand me either. I would love for us to have a computer "intelligence" that would equate, in terms of real cognitive power, to the intelligence of a simple bee.

    However, *we are far from even that*! None of these powerful deep learning algos are even near, in terms of the ability of a dumb bee worker to process its surroundings. It is regrettable, but it is the truth.

    Hence, I insist that Elon is not merely "apologizing for FSD taking such a long time to develop." He is indirectly saying: we might as well try to replicate the simple yet marvelous intelligence of a bee worker. But that's the point: we cannot yet, and unlikely to do so for next several years, if not decades.

    We need really BIG breakthroughs in our understanding of what constitutes intelligence. I do not believe that Elon and his engineers will do this next year. So, IMO, he is either delusional or a liar or both (as one can reinforce the other).

  35. About context. When Elon is talking about "real world general intelligence" he continues to talk – in this and other interviews – about what can be achieved with this "real world general intelligence", and it's always about navigating and handling real objects, never about contemplating the meaning of existence or writing novels. Elon Musk believes that AI's will eventually be able to do those things as well, but he never links those abilities to "real world general intelligence" when he talks.

    Example: The car will be able to navigate streets, the robot will take on boring and repetitive tasks. The robot will be able to clean a house.

    So, from that context, it is clear that Elon Musk believes that "real world general intelligence" is equivalent to perceiving the external 3D geometry, identifying objects and choosing appropriate movements. People may disagree with such a definition of "real world general intelligence", but that is clearly what Elon Musk believes.

  36. Nah, not when you preface "general intelligence" with "real world". That "real world" actually limits the meaning to "perceiving and acting correctly in a real environment in real time". Elon is clearly not stating that the FSD SW will become self aware, but that it will "understand" the surrounding 3D geometry, sort of at the level of a fly that's been trained to recognize road signs. You have to understand the meaning in the context….

  37. General AI has been the goal and the dream of researchers for decades. Since, at least, the 60s. Most serious researchers think that we are pretty far from that. And Musk is well aware of it. *Should* be well-aware, if he is as knowledgeable as so many claim. By restating the FSD issue in terms of general AI, he is de facto admitting defeat in the short term. This should be obvious to anyone who has ever touched this subject academically and practically.

  38. How do you figure that? You may think that FSD is a scam, but what specifically signals that Elon Musk is admitting to such a thing?

  39. Indeed. General AI means consciousness. If they build sentient machines then self-driving cars will be the least of our worries.

  40. Genuine question, if you happen to know the answer: are they or their client(s), the car companies, underwriting it? Because even level 4 is quite ambitious.

  41. Elon's statement is akin to saying they will understand how the brain works or what consciousness is

  42. "To solve self driving is to solve real world AI" is not only an admission in the failure of Tesla's approach to autonomy, it is also a wrong statement, as other companies like Waymo and Mobileye are already rolling level 4 autonomy in select areas using an hybrid approach of AI, cameras, lidar and maps for autonomous driving.

  43. "I mean, I did promise the internet that I'd make robot catgirls. We could make a robot catgirl." – Elon Muske

  44. What's that? Am I seeing first signs of admitting that "full self-driving" is a pipe dream of our time? Nice way to spin this though. And, so, the scam continues!

Comments are closed.