How difficult is it for BYD to clone his autopilot chip and will that stop the Chinese from exporting knockoffs because I'm certain they can close the data and software gap instantly by swapping out a harddrive or 10.
I think he's got 10 years before the chinese chip fabs catch up.
Musk's question how did Tesla build the best AI chip for driving in the world without being a chip designer was the far more interesting yet overlooked question.
Musk has a talent for framing problems appropriately so that optimum solutions emerge with a minimum deployment of capital. And he's not secretive about it, he applies his training in physics and phenomenal knowledge to make good bets and eventually identifies and acknowledges his mistakes.
Tesla doesn't have an advertising budget. Musk couldn't care less about it's shareprice but I think he will be a bit perturbed if identical technology emerges faster than he foresaw as was Apple when the Android ecosystem emerged.
As it stands with his chip and data it is game set and match for autonomous driving which is a software game and the leader takes all the data and a fully autonomous vehicle is worth significantly more than non-autonomous vehicle and if you are data scientist you probably want to work for the company with the best data-set and highest remuneration.
The data is the gold and Tesla owns it and no-one can steal it and use it for at least 10 years by which point battery vehicles will be cheaper than petrol vehicles which is the major problem with the model 3. It is currently uneconomic against a toyota corolla but gains ground with every piece of battery and energy storage research that musk does not pay for.
seems to be constantly tweeting about it though.
This is an oft-repeated trope, but one look at Tesla's reports will show it's not the case. Last year, approaching $100M (I want to say certainly greater than $80M) spent on "Advertising and marketing".
They have a highly effective, modestly funded, non-traditional advertising strategy.
2) I can't remember a CEO who has complained more about the stock price in particular those shorting the stock. And worst of all then go on to illegally manipulate the market to target them.
3) You have that completely backwards. Musk is the worst at solving problems with the least capital. Look at the way his manufacturing processes work or the decision to build out the Supercharger network/Gigafactory. There are definitely cheaper ways to do those. Better probably not. Cheaper yes.
4) I have worked with more Data Scientists than most and I assure you that they are just like everyone else. They care most about: company reputation, work environment, money, challenging problems etc. It's not just whether there is a great dataset or not.
It was about the best AI chip in general not based on some arbitrary efficiency number.
No one cares how many TOPS your 500watt non-redundant monster can run, because it’s not being put in an EV.
Because self driving cars are a multi-year journey and those platforms will likely become more efficient.
Remember we’re talking about systems made of multiple chips. It’s like asking “what’s the best Bitcoin mining rig?” The question is one of hashrate per joule. If someone asked if you should buy a 1000watt box that did 10TH/s versus a 200watt box that did 5TH/s, which is better?
They maxed out their available power envelope (~70 watts) and the resulting performance is about 20x what they were getting from their prior nVidia platform. In other words, they will always make the chip big enough to pull ~70 watts, at the highest efficiency possible.
My limited understanding of the hardware design is that once you have figured out the specific architecture you can scale up and down the power consumption almost linearly with the die size. Obviously plus or minus various hard limits (e.g. oversizing compute for the available memory bandwidth)
Tesla has already been working on their next chip for over a year. This chip they just started shipping finished design 1.5 years ago. This will not be the last AP hardware they ever ship.
I fully expect that as we delve further into “FSD” there will be functionality on the fringes that will need even more horsepower.
It’s easy to think in the present that there’s a logical “end” where development will be “complete” but that’s almost never actually the case. So there will be a Gen4/5/6, specifically what features they will unlock is very hard for me to predict. I think this is actually the type of thing Elon sees right through like it’s obvious.
Like how he sees a path to colonizing Mars which includes launching a global satellite internet company.
And the trained network doesn’t matter as much as the training pipeline, which will include hardware and software external to the cars.
A neural net that you can’t update is less useful, and I highly doubt that the cars themselves are doing any learning.
2. Musk has investors and the price of his cash financing goes up when people short his stock and there is negative sentiment. It was a major strategic weakness when the viability of tesla depended on short-term cashflow financing and he did what he had to save the company that he poured his life savings into. Not so much now that he's built a system where every additional vehicle sold adds additional value and long-term investors can see that. Musk thinks long-term the markets think short-term but if you run out of cash, there is no more planning and he hates the markets lack of vision and people profiting from negative sentiment effectively killing optimum solutions. The market has a lot of long-term faith in Amazon and Google and Microsoft because of their dominant networks and economies of scale and profit from complex software with zero marginal cost.
3. Solving problems in sub optimal ways isn't solving them it's compromising on them. He solves them optimally with the least capital but admits his mistakes, one of which was not judging the benefit of humans and teamwork on production lines versus robotic manufacturing technology correctly. If you are generating negative externalities you aren't reducing entropy in the system and Tesla learnt this painfully and won't forget it.
4. I'm completely out of my depth on the data scientists employer preference but if I had the choice of solving problems from a constantly growing ocean or from an artificial fishpond and I had a desire to contribute majorly to solving the autonomous driving problem and I had to factor in the probability of earning enough money to buy my time freedom as quickly as possible then the data is suggesting that Tesla at this point in time is a reasonable bet.
The question is whether he can compete with Toyota on mass scale while battery tech comes down in price but he's quite lucky Toyota misbet on fuel cells and don't readily admit their mistakes, it appears he's built a 10 year lead on the autonomous driving and can compete in the high end of the electric market profitably. The first question anyone would ask a driver of an electric BMW is why they didn't choose a Tesla? That's a fairly strong position, I wouldn't be shorting that stock.
The entrenched car lobby have only have two lines of defense left. The first which was to short him so he runs out of access to cash from negative sentiment is virtually broken on Tesla's current revenue and IP acceleration. The last line will be a delaying argument over Safety but he has a much larger dataset.
Geico falls when Tesla proves safe autonomy. There is a lot at stake here and Musk has focused primarily on the physics which is the right thing to focus on for humanity.
Does that make it the most advanced AI chip in the world? No! In fact, frankly, you're trying to have a technical discussion about a bullshit term. Tesla isn't producing an AI chip, they're producing a scalar product chip. It does a 96x96 Mult-add and that's basically it.
How does it do on any of the standard AI benchmarks? Oh it can't run any of them because it's not got support for any of them because it lacks many of the basic features of what any normal person would consider a chip capable of doing AI work.
It is the best chip to run the convolutions and pooling of 1 neural net that Tesla thinks it will want to run. It's the best 'AI' chip in the same way that the chip Nokia produces for it's mobile backhaul is the best 'neworking' chip.
The point is that the chip they're trying to compare their custom ASIC to is capable of running several AI frameworks none of which the ASIC can. Hell it can't even run single precision floating point operations.
Tesla has the best chip in the world for the one neural network they want to run. I bet you good money their custom ASIC sucks at running Waymo's AI workload.
"Tesla has the best chip in the world for the one neural network they want to run" is what I wanted to convey but don't actually have the technical background to do it.
"Frankly you're trying to have a technical discussion about a bullshit term"
You are right. I apologise for diluting the discussion and will go back to reading hacker news not creating it.
I love this site and don't want to dilute it.
I think whoever develops level 4 FSD tech that works for most Americans daily car commute is going to be a trillion dollar company. It can even require me to be in the driver's seat, not work in rain/snow, and be geofenced if it gives me at least a 5+ seconds heads up for taking over as long as it is enough time to switch context from working/whatever to drive. Think about how much time Americans spend driving for which we could free ourselves from if our cars had FSD tech that worked most of the time. If/when I have a decent commute that benefits from a car over public transportation, I would pay thousands of dollars extra for a car if it was able to free myself from driving 75% of the time.
Tesla now has got hundreds of thousands of cars that people use in level 3 mode which is then used to train their L4/L5 FSD tech while also retaining all of that real world data to train their models. I don't work for self driving tech, but given the massive amount of real world data they can collect, I think they are now most likely to hit the point where they can solve basic commutes in geofenced areas with L4 over everyone else and I think people will pay out the nose for that (with the added benefit of people buying electric cars too).
- sufficiency of compute;
- software architecture;
- complexity budget;
- sensor suite;
- datacenter.
Tesla is the most hyped and least proved contender. They moved the conversation around the AI chip but that doesn't matter compared to the algorithms that runs on it! Waymo and Cruise are years in front but mask managed to hype you and make you drink koolaid in such a way that you know think they are the leaders.
This is laughable.
This is to say nothing about their quixotic attitude towards self-driving in the face of competitors using better methods.
Heck, Jaguar boasted endlessly about their i-Pace being a great Tesla competitor, while in practice it disappointed in range (https://insideevs.com/news/340027/jaguar-i-pace-range-test-y...)
I certainly don't mind Elon talking up how Tesla is fighting giants and doing Good Things. I like "little" companies taking on big ones. I don't like selling things people that may kill them or others.
I agree with the GP comment, I really want Tesla to succeed but they (or, rather, Musk) take hype and overselling to a new level.
The chasm between Musk's superhero image and his Twitter meltdowns and missed deadlines, is likely what makes people be skeptical/annoyed with his communication style.
I mean, it's called "auto pilot". It is not an auto pilot. It is not nearly ready for unattended driving. And yet, people use it as such and are surprised when they crash into a divider or something.
There's a built in assumption that because the tech has progressed from point A to point B that it'll easily move to point C in the same amount of time. That's not at all guaranteed, and for all of the millions (billions?) being poured into self driving tech we don't have to show for it. It may be many, many years before we do.
Almost all commercial airlines have an autopilot, but not one of them flies without an actual pilot and co-pilot in the cockpit. The only thing an airplane autopilot does is keep the bearing and altitude constant -- Tesla's "autopilot" actually does a bit more than that. Is the outrage over calling it "autopilot" based on the public's misunderstanding of how an airplane's autopilot works?
Yes, a plane autopilot only maintains bearing and altitude, but that's also all a plane needs to do. An equivalent system in a car would be utterly useless, so if you're taking Tesla's terms literally then they're telling you they've made something that's of no use to you.
If, instead, you think of an autopilot as being something that allows the pilot to take their attention off the controls to do something else then, well, Tesla's autopilot explicitly doesn't do that. So it's not the right name when taken either literally or figuratively, so little wonder people get confused.
Quite the opposite, there are common scenarios when the plane needs to deviate from pre-programmed course or otherwise can't rely just on an autopilot. All of them are handled either by higher-level automation (eg. ILS for landing), or by pilots themselves. Unsurprisingly the hand-over is signaled similarly to what Tesla does - audible and tactile cues for the pilots. A pilot can also initiate take over at any time by the way of flight control inputs. Lastly, regardless whichever automation is active, the pilot in charge is held both morally and legally responsible for the airplane and passengers or cargo, and expected to turn off the the automation when proper.
All in all Tesla's "Autopilot" already does more than an autopilot does. And while I'm somewhat pessimistic on how quickly it can become better than humans both statistically and also in rare scenarios, there is constant progress by all major vendors.
Common scenarios for an autopilot to be disconnected, or to initiate handover:
- collisions avoidance; the air corridors are busy places, and maneuvering to maintain safe distance from other planes is expected
- take-off and landing, though those can be performed by higher level automation
- pattern holding around the air-port, where course change is frequent
- re-routing to avoid severe weather or handle emergencies
- handling major turbulences (autopilots have limited maximum rate of change for safety reasons)
- loss of instrumentation, due to icing, foreign object ingestion, malfunctions; a common setup is "2 out of 3 sensors need to be in agreement"
As you can see, the equivalents of some of those scenarios are already handled pretty well by self-driving cars, like self-parking, or collision avoidance.
Not if the name was chosen to deliberately be misleading.
>The only thing an airplane autopilot does is keep the bearing and altitude constant
If autopilot means that then we had autopilot in years, I remember students creating toy projects that were following the road.
The issue is that some(no idea of a number) people, including Tesla owners, think the car drives itself and do not pay 100% attention. I am not sure why they do not understand that there is a chance they will get killed or kill somebody and the driver has full responsibility.
That was probably true in the 1950s. Today's flight automation manages the entire flight profile including climbout and descent, fuel management, etc.
To get a Tesla all that is required is enough money to buy it.
(skip down to the section that begins "Here, let me give you a quick demonstration:")
Have you seen the progress in new features and reliability in just the last 6 months of OTA updates? Their cars are becoming more and more autonomous, and they are rolling out this autonomy in ways that is very impressive.
The fact that they can say a new lane change algorithm went live and they know they are now doing 100,000 lane changes a day without a single incident.
The fact that they can see when I take over during a lane change and that sequence is stored and anonymously included in future training sets.
I have been experiencing the massive pace of progress over the last 6 months of OTA updates, and I almost believe Musk when he says he can get to FSD in a few years.
IMO you may have cars get “stuck” in very strange situations which will be extremely annoying, but I think they will be significantly safer than humans, hugely convenient the vast majority of the time, and save a lot of lives in the process.
I wish Tesla the best of luck.
And I think it's uncharitable to interpret these as delivery commitments. From the tone of the delivery, they are clearly targets (probably aggressive targets) and not commitments. Those who bludgeon Musk for being overly optimistic with timelines seem to overemphasize the importance of specific timeliness and lose sight of the bigger picture.
If they put up autonomous driving videos going through a handful of major cities taking the most awkward routes through places where driving culture is typically more agressive and the narrowest of country lanes through the towns that were built for horse and cart and space is at a premium. Places like Rome, Barcelona, New Delhi. Towns that have built one way systems into their traffic management due to limited space and measures are taken to purposely bottleneck traffic like narrow road right of way signage. Show a car doing that and I'd understand their confidence.
E.g there is a town where roadside parking is allowed because it purposely bottlenecks and slows down main road traffic. You have to judge and wait by the parked cars for a space you can pass through even when their maybe cars coming towards you there is time to pass the parked cars and into space where two cars can pass eachother. How is an autonomous car going to handle that kind of senario where it has to make a judgement call that factors in a far greater amount of risk element?
Many Tesla automated driving sceptics say without lidar, Tesla has a major issue with FSD.
> How do you tell a machine what to do when there is no correct answer?
Why would you expect a machine to be able to handle scenarios that humans cannot handle?
Note: I am sceptical of FSD being available this year, or even in a few year.
That's exactly my point. Isn't that what Tesla is trying to do with FSD? It's trying to train a machine that will have to make ethical decisions.
How do you factor legal liability for a machine in the same way you would with humans in the same scenario e.g fatal accidents?
This is somewhat like the civilian version of AI military drones one day making automatic target selection.
If humans are driving around somewhat successfully without being able to handle all edgecases, as long as the a machine can do as good a job or better it is a viable alternative to human drivers.
Do you hold the owner, manufacturer, or insurer responsible to right the wrongs where an autonomous vehicle is at fault?
I would expect the insurers and manufacturers to use contractual terms that absolves them of all liabilities where the car's occupant isn't paying attention to the road.
Lidar sensors are around $10k and up, and the waymo cars I've seen have many of them.
Additionally, they only sort of say how far away things are, which sort of overlaps with the ultrasonics and radar.
So, maybe the roi on lidar would not be great.
That said, the engineer in me thinks lidar would be a nice add-on (sensor fusion). But it depends on someone actually shipping one of those dirt cheap solid state sensors we see in press releases (but not in real life yet).
1) internally feature complete - end of this year,
2) internally they believe reliable enough that the software can let the user not actively check in / driver does not need to look out the window - mid next year
3) able to convince regulators in some jurisdictions that drivers do not have to look out the window - end of next year.
Here's where people go wrong. Now I didn't watch the entire video, but my understanding is Elon is not saying that the car will never need a user to take over. He is not saying the car will drive anywhere, at any time, under any conditions.
He is saying that the software will be reliable enough to fully drive the car in all of the basic driving environments (highway, rural, urban, etc.). It detects stop signs, traffic lights, it navigates intersections, it makes unprotected left turns, it can merge, etc.
That's what I think Elon is trying to convey when he says "feature complete". There will be neural networks, algorithms, logical decision trees, whatever, to handle all of these common driving scenarios. He wants all those scenarios coded and demoable in some condition by the end of this year.
This is not something they roll out in a day. This is gradual feature deployment over the next year leading up until "feature complete" which is a point where some drivers are getting door-to-door in some cases under supervision with wheel check-ins but without any disengagement for the whole ride.
After that point it is 6-12 months of training, training, training the networks based on Fleet learning, billions of miles tracking every single disengagement, and aggregating all that user feedback into the algorithms.
They will know when the system needs to disengage. They will know every way that it fails. They will be tracking this data and providing it to regulators for evaulation. And Elon thinks around the end of next year, they could be in a position to make a case for removing the wheel checkins and telling users that this car can drive itself, and keep you safe, and that if it hits a threshold where it can't do that, it will fail gracefully.
So perhaps there will still be the intersections where the car gets stuck. If there is a road closure and the cop wants you to drive on the wrong side of the road, maybe the computer will refuse to do it and just sit there and you'll have to take over. But if you're cruising down the highway and an 18-wheeler pulls out in front of you, yeah, the car has to be able to see that and stop.
Musk is a brash and reckless CEO, no one should be surprised hordes of people are hedging for his downfall.
I'm pretty sceptical about having nobody in the front seat of the car while it drives me myself. Time will tell. I can be sanguine about it as I have nothing directly riding on it (as it were) either way.
IMO the chips are real the cars are real the data collected is real and the network is real
I am fine for Tesla to train the code with the fleet but the problem is with the marketing and drivers that put live in danger by thinking the problem is already solved.
Computers beating humans at chess was also an impossibility.
Then beating humans at Go was an impossibility.
Then beating humans at a strategy game like Dota was an impossibility (openai 99.4% win rate).
https://www.theverge.com/2019/4/23/18512356/dota-2-openai-ai...
Machine learning has made massive progress in the last decade, I would not bet against it.