MANY Stacks to Rule them all? Why Tesla FSD Beta might need a LOT OF VERSIONS Instead of One!
Vložit
- čas přidán 30. 11. 2022
- Can Tesla's Full Self Driving (FSD) Beta work with just one stack to rule them all? Perhaps not--or better yet, perhaps there is a better way to get Teslas to drive like they've lived in your neighborhood a whole lifetime!
Elon loved it on Twitter--get yours before it's gone! The "Now Pay $8" T: drknowitall.itemorder.com/sho...
Get The Elon Musk Mission (I've got two chapters in it) here:
Paperback: amzn.to/3TQXV9g
Kindle: amzn.to/3U7f7Hr!
**To become part of our Patreon team, help support the channel, and get awesome perks, check out our Patreon site here: / drknowitallknows . Thanks for your support!
**Want some awesome Dr. Know-it-all merch, including the AI STUDENT DRIVER Bumper Sticker? Check out our awesome Merch store: drknowitall.itemorder.com/sale
For a limited time, use the code "Knows2021" to get 20% off your entire order!
**Lift pucks for your Tesla: amzn.to/3vJF3iB
**Emergency tire fill and repair kit: amzn.to/3vMkL8d
**If you're interested in a CO2 Monitor, you can check it out here: amzn.to/3PsQRh2
**Check out Artimatic: www.artimatic.io
**For 30% off the amazing OneCarStereo instrument display for Tesla Model 3/Y, use this link: bit.ly/3vd7NRv
**Get 20% off an awesome Ekster wallet and/or wallet locator(!), check out my affiliate link here: shop.ekster.com/DrKnowitall
**Want to get in on WeBull's awesome Crypto and stock fun? Check out this link, and get started trading stock and Crypto!
a.webull.com/i/DrKnow-it-allK...
**If you are looking to purchase a new Tesla solar roof, solar tiles or PowerWall, just click this link to get up to $500 off! www.tesla.com/referral/john11286. Thank you!
**You can help support this channel with one click! We have an Amazon Affiliate link in several countries. If you click the link for your country, anything you buy from Amazon in the next several hours gives us a small commission, and costs you nothing. Thank you!
* USA: amzn.to/39n5mPH
* Germany: amzn.to/2XbdxJi
* United Kingdom: amzn.to/3hGlzTR
* France: amzn.to/2KRAwXh
* Spain: amzn.to/3hJYYFV
**What do we use to shoot our videos?
-Sony alpha a7 III: amzn.to/3czV2XJ
--and lens: amzn.to/3aujOqE
-Feelworld portable field monitor: amzn.to/38yf2ah
-Neewer compact desk tripod: amzn.to/3l8yrUk
-Glidegear teleprompter: amzn.to/3rJeFkP
-Neewer dimmable LED lights: amzn.to/3qAg3oF
-Rode Wireless Go II Lavalier microphones: amzn.to/3eC9jUZ
-Rode NTK Studio Microphone: amzn.to/3vjMomW
-Focusrite Scarlette 2i2 audio interface: amzn.to/3l8vqDu
-Studio soundproofing tiles: amzn.to/3rFUtQU
-Sony MDR-7506 Professional Headphones: amzn.to/2OoDdBd
-Apple M1 MacBook Pro: amzn.to/3wPYV1D
-Docking Station for MacBook: amzn.to/3yIhc1S
-Philips Brilliance 4K Docking Monitor: amzn.to/3xwSKAb
-Sabrent 8TB SSD drive: amzn.to/3rhSxQM
-DJI Mavic Mini Drone: amzn.to/2OnHCEw
-GoPro Hero 9 Black action camera: amzn.to/3vgVMrH
-Suction car mount for camera: amzn.to/3tcUfRK
-Extender Rod for car mount camera: amzn.to/3wHQXsw
**Here are a few products we've found really fun and/or useful:
* NeoCharge Dryer/EV charger splitter: amzn.to/39UcKWx
* A totally fun Tesla coil toy--awesome for the whole family: amzn.to/30rvzbI
* The Humane Interface, by Jef Raskin: amzn.to/3ajM7YT
* AI Superpowers, by Kai-Fu Lee: amzn.to/3dgRSs3
* A camping mattress for your Tesla model S/3/X/Y: amzn.to/3chSRrQ
**Music by Zenlee. Check out his amazing music on instagram -@zenlee_music
or CZcams - / @zenlee_music
**New intro by Anthony Soto: sotoe907@gmail.com
Tesla Stock: TSLA
**JOWUA:
*For an awesome wireless charging phone holder mount, check out: bit.ly/3tRMSQR
use the Promotion Code: DRKNOW5
*Tesla USB Hub: bit.ly/2PXHPit
use the Promotion Code: DRKNOW5
*Powerful wireless iPhone/Android charger for your Tesla Model 3 or Model Y, check this link: bit.ly/3b5fHSW
use the Promotion Code: DRKNOW5
**EVANNEX
Check out the Evannex web site: evannex.com/
If you use my discount code, KnowsEVs, you get $10 off any order over $100!
**For business inquiries, please email me here: DrKnowItAllKnows@gmail.com
Twitter: / drknowitall16
Also on Twitter: @Tesla_UnPR: / tesla_un
Instagram: @drknowitallknows
**Want some outdoorsy videos? Check out Whole Nuts and Donuts: / @wholenutsanddonuts5741 - Věda a technologie
Very interesting video. I am really looking forward to seeing how FSD copes when UK Beta testing starts and which brave person will be the first to tackle some of our Single Track Roads - these aren't wide enough for two vehicles to pass, relying on one driver reversing (can FSD reverse yet?) to the last passing place (a slightly wider spot on the road) to enable passing to take place. You are also obliged to honk your horn prior to corners or humped backed bridges (does FSD use it's horn at all?) to warn oncoming traffic that you are there! This all becomes very tricky if multiple vehicles are all close together travelling in both directions as you need multiple drivers to reverse and sometimes the passing places are only long enough for a single vehicle! I suspect Tesla would plump for the easy option of routing around these roads, though that may not always be possible!
I think it's better to make FSD so good in general that it can easily adapt to whatever laws, weather and driving culture the current region have.
I believe Tesla is already heading in this direction. During AI day 2 it looked like the AI team was showing a lot of detail work being done in San Francisco area and my first thought was SF will be the first Robo Taxi market. The approach will not be gated markets but HD FSD if you will for Robo taxi to operate in, with every intersection carefully detailed. In addition I have been hearing of perfect FSD drives coming from the SF market for several months now.
What means SF, please?
@@tobiasmuller5435 San Francisco, California
Brilliant. I am driving FSD beta in the Tampa Fl area and I am jealous of the LA/SanFran experiences of hours of driving with ZERO interventions. FSD beta is great for me, but I still have to intervene almost every drive. Would love overfit weights for my specific local.
Interesting idea. I had to do a 1 hour drive this morning after a fresh snowfall (first of the season) and the roads were icy. But coincidentally I had no connectivity for about 2 hours so would not have been able to download an 'icy road' stack.
I have said this before, too, and actually suggested even finer fine tuning - for specific intersections for example (the ones that are particularly crazy and wonky and confusing). I also think it could learn different people's driving styles and modify what it expects the cars around it to do accordingly - aggressive drivers, cautious drivers, drunk or distracted drivers, a-hole drivers, etc. Like if it sees someone veering through traffic aggressively it could act super cautious and give them space. If a car isn't staying in its lane very well it might choose not to pass it. If a car is driven super timidly then it may decide to pass that car asap. That kind of thing. This understanding of different drivers could then be used to not only how the car reacts to that one car, but if enough driving styles fall into a category it could change the global settings for how the car drives. So if it starts seeing a lot of aggressive drivers it can act more assertively when changing lanes for example. Or if almost everyone is driving 5mph over the speed limit then it might start doing the same.
This seems like a really good suggestion, since there are so many driving habits/rules that vary by location
Your videos have gotten so good. Focused, articulate, on point.
I could see adding GPS to the training data for location characterization and using not fine grained GPS for large location areas (like larger cities or regions)
I would make the analogy to different languages and dialects- the cars just need to be aware of the dialect of signage/behaviors in its immediate area-- however, I think weather should be completely contained within each particular stack, and same with city/highway driving… one should not rely on WiFi to download the right weights for rain vs snow vs sunny etc.
Interesting idea, but I'm not sure I would be comfortable if my car significantly changes behavior from time to time and place to place. Learning FSD has involved significant training of drivers (like me). Would a different driver need a different stack?
I imagine it is a shared subsystem with extra layers on top I think it makes sense since different drivers in different regions/biomes end up driving differently. Even an individual driving out on the open highway will drive differently from city street driving (not to mention all the extra obstacles and intersections). I imagine all of what currently exists would still be basically shared too for general driving but there would just be "extra tuning" that could be laid on top or perhaps extra network layers to adjust for various "driving styles" (more or less aggressive, where just a speed/acceleration governor isn't enough to keep it safe)
Capsule networks are not a brand new idea but having multiple neural networks that can communicate or higher level ones that rely on smaller separately trained networks outputs can be used as the inputs for (basically can be used for the smoothing or "style adjustment")
Sounds good, but here in So California I'd have to change stacks all of the time. If I went to visit my son about 45 min away, there is snow and ice on the ground and a snowstorm on the way. They sometimes have "white-outs". If I was to visit on of my daughters about 50 minutes away, there is a pretty heavy fog through the morning right now along with some rain. The issues here is that we have sub tropical, to desert (hot and sometimes freezing), heavy fog, various road types and moderate weather inland all within 50 miles of where I live. Mountains, deserts, sub tropical, coastal offshore and inland moderate valleys. There are many days out of the year that one can go surfing in the morning and snow skiing in the afternoon.
The weather differences and road conditions can vary from one extreme to another in a distance of 40 to 80 mile span. In Denver, I frequently visited Colorado Springs and sometimes Idaho Springs which included totally different weather systems as well as road conditions. When I lived in the mid-west your thesis sounds a bit more reasonable. I'm curious what FSD does when a Tornado is in the vicinity like in MO or a dust storm like in Yuma AZ or a sudden "Snow Storm" forming from the great lakes in Northern Wisconsin or other great lake areas...no clouds, just snow forming off the great lakes creating sudden almost blizzard-like conditions.
I don't think your idea is actually bad, but I do think FSD is going to have to be fluent in the broad spectrum of driving conditions in one stack if it is to become truly viable. Loosely defined, you are talking about geo-fencing at some level. I don't think Tesla will go in that direction.
I can testify to some huge differences in how people drive in different cities. In Louisville, I learned to wait for a few moments after the lights changed because so many people would come barreling through an intersection when the lights had just changed. Ignore that and you were likely to get broadsided. In Pittsburgh we have the "Pittsburgh left" when a light changes and turning vehicles are allowed to turn before those going straight proceed. This is often accompanied by the flashing of headlights, and there is always a hesitation before turning to confirm the opposing vehicle is practicing this courtesy, but it is common. In Dallas, when the lights change it is the signal to floor it and go like hell. Imagine a Dallas driver in Louisville and you get an ugly picture. then there is New Jersey where freeway traffic runs at around 85 mph despite the posted 65 mph limit. Driving at 65 on those freeways can actually cause wrecks as trailing vehicles have to slam their brakes and swerve into an adjacent lane - always accompanied by gesticulations implying violent intent. So, yes, making adjustments by locale is essential.
Great ideas. I often think that it drives better in other areas then it does in mine when watching video then comparing to my own reality and how the car performs.
Lets hope it starts to recognize snow and slippery conditions and adjust to that. So far it doesn't seem to modify its driving behaviour during poor weather conditions other then letting you know poor visibility the weather is bad or breaking/battery capability is limited.
just a thought but when travelling and the car is being fine tuned to the local areas - this could make some people anxious as we tend to know what to expect get comfortable with how the car performs have certain expectations and if its regularly changing this could be challenge for some as to not knowing what to expect. We could call it a car with multiple personalities.
The problem with this is that when you go on a trip you might go through many locations and also I don't think for $10/month they can be allow to download new neural nets to the car on demand. Also this is more of a cheat sheet type thing to leap ahead to a new maximum instead of slow progress on one massive amazing neural net that can be dropped anywhere and work. Its like how they removed the lidar because it was working but it ended up being a crunch.
Michigan here. Can have all four types of weather, rain, sleet, snow, and sunshine while waiting at a stoplight, LOL.
The same question applies to language models.. you can communicate pretty well in/with English almost in any English speaking country .. then some "minor changes" might be needed to better suit the location (slang, different naming conventions, etc). The "international English will drive fine" but slower.. but with the updated , more adapted version.. speed will improve, etc..
if it's raining cats and dogs, you'll need software that can do animal identification. oh, john, could you go over to the Taco Stand and get me 6 tacos and mail them to the czech republic?
This introduces an ongoing data cost for FSD where the owner may not have paid for premium connectivity. How would that work?
The car should be delivered with "standard tuning". The car should then tune itself by watching the primary driver for 100 miles or so when not on FSD. The primary driver knows how people drive in the area and what it takes to get things done. Maybe some of the fine tuning can be stored in the driver profile. My Missouri driving style is very different from my Illinois driving style and John's wife's idea of comfortable driving is VERY different from John's 😄.
My first thought would be how Musk, the engineer version NOT the Twitter version, would evaluate this idea using First Principles thinking. He is famous for saying “the best part is no part” and this seems like extra parts. Maybe you could do a video on this idea.
Yes. Customize for the location, season, road types. As you replied to my Patreon message about the narrow winding roads in New England. FSD has room for improvement. This is right on. Also, customize for the driver's preferences.
Correction, most people in the PNW don’t know how to drive in the snow even though we experience snow pretty much every year. People in the PNW also struggle to drive in the rain. I’m only saying most people here struggle in these conditions, not everybody.
I visited England for 3 days once in my life. Rented a car at the airport and had no problem driving on the left side of the road. There are different rules for different countries/states like no right turn on read light or motorcycles being allowed to drive on lane markers between cars. FSD would have to be aware of different rules for the area, but I don't think that should require a unique stack.
I think over fitting is a bad idea, but having a winter version that goes out every Halloween to the northern hemisphere and Easter Sunday as a good b3nchmark for the southern hemisphere. That means they can work on the stack all year, improving it. Even go as far as making voluntary for people in Texas, fl, Caribbean Islands, and such.
I'm Michigan you just don't get pulled over for speeding on the expressway. We love crossing the border back to Michigan because the stress of watching speed just goes away.
We have monsoon season here in Arizona too, some areas get sandstorms from them here....
I agree that more localized data would be helpful. It doesn’t need to be centimeter level data, but it would certainly help especially in bad weather / snow where you can’t see the lane markings well. I have to think that Tesla will do it but not before it gets the generic AI driving to a certain level. If they do it too early it might act as a crutch similar to what happened with radar for teaching the neural network.
I thought this was all very insightful conjecture. I suggest you talk to your contacts in the FSD team and ask them to get back to you to discuss any of your ideas that they are confused about, after they review this video.
This is such great information, and they may already be thinking along these lines already, but if they are NOT, you might just open up their minds to new and exciting concepts that will add significantly to the improvements to FSD going forward. And that would be very important!
Regions would be a good start. Like driving in SoCal would be way different than driving in Alaska.
In the early to mid 2000s my late Wife and I bought a Garmin GPS system that required downloading of individual regions.
An ethos stack probably would have issues with NHTSA (rolling stops, X mph above speed limit, etc). A climate stack could be simply stored in FSD which would know how to drive in snow, rain etc and activated at the start of your drive based on realtime weather along your route. A more important difference is how to drive on roads laid out in the 1890s vs 1960s, etc. The older the road, the more likely a noobie will have problems (narrow rds, no building setbacks, many non-90° intersections, unpaved, etc).
There only one scenario that I don't think FSD has tried yet and despite me asking lots of testers, no one has said one way or another or to put it another way, no one seems to know! Perhaps Doctor Know it all knows,"Is FSD trained to handle Skids? Can FSD drift safely? Could FSD safely correct a skid after the skid has already started?
FSD doesn’t need to be quirkily ‘local’ it just needs to be safe, and get you there. When you get a taxi you don’t quiz the driver about where he’s from.
I always thought that you should have the agression/mellowness be GPS related. Big cities: don't be polite. Small towns: be polite. It's now a user setting. Agression/politeness should be at least partly be influenced by how people drive where you are.
Interesting ideas ... although I would think having Localization and Weather Overlays as behavioral modifiers to a standardized foundational universal stack might potentially be less data intensive... And regarding the difference between miles and kilometers - I would think that this would not necessitate a new stack of its own, as it is a relatively simple matter of unit definition / selection - which I believe is already handled according to user preference...
Over fit FSD for the worst environment (let’s say Montréal where too many drivers consider road regulations as options) and it should be good almost every where. Personally I think that my experience over there made me feel comfortable driving in cities like Toronto, New York or San Francisco. It was very different in Europe where road signs and regulations are very different than North America. And miles or kilometres are no different, just a simple arithmetic calculation (there is CYBRLFT, a taxi driver in Southern California who make FSD videos with his car set in km and it adjusts to speed posted in miles).
No need to load from a remote source. One thing we've learned in this business in the last 50 years is that designing to finite storage considerations is mostly a waste of time. Store every scenario locally in the car and reload the same way system updates are done. Secondly, I would assume there would be learnings from developing these overlays that allow differential updating rather than replacement.
Didn't Tesla buy a small company that specialized in neural network compression? Is that being accounted for here or could they send that large NN model in a highly compressed form over cellular?
It seems inevitable that the scope of FSD for everywhere will be found to be intractable. So something like this constraining of the local domain seems inevitable.
You idea makes real sense
Makes sense.
It’s more than weather. It’s also people’s habits which are influenced by things like insurance rules and normal levels of congestion and how frequently road rage becomes violent.
As an example-why are there no FSD videos for Boston?
ARK Invest said this week they expect EVs to capture 90% market share in 2027 - 5 years from now, but you want TSLA to spend enormous resources adapting to the human drivers which are being rapidly replaced. Very much like focusing on the horse and buggy drivers while the automobile replaced them. Better to focus on solving problems that do not just go away by themselves with time. And if you think on it, you're actually suggesting to move in the direction of Geo fencing and automatically switching between Geo fenced areas.
Another thought: Certainly driving conditions and "road rules" differ. In America, if it seems not unlawful (or not forbidden or there appear to be no police cars around), drivers seem to do it. Driving in Europe, for instance, there seem to be fewer "laws" but many well-respected unwritten "rules of the road." Police are there for good reasons but are not so intimidating for ordinary driving and drivers. A typical example is "passing on the right." Another is less worry about speed limits (at least that used to be the case 🙂). However, if FSD is in charge, one would expect the car always to do both the "right thing" and the "lawful thing" and invoke no worries about police presence, etc.
I understand that you think it will require more "stacks" but I'm thinking it just needs "factors based on current conditions" regardless of geographic location, especially if you are on a long trip traversing many weather and landscape conditions. I can't imagine that FSD will ever be able to deal with cultural differences that affect driving behavior.
You might be right.
They currently create a binary for the different subscriptions: autopilot, enhanced autopilot (might actually be same binary right now), FSD non-beta, FSD Beta.
It seems like they will have a core FSD full stack that does all of the safety stuff and lane-keeping, then add on neural nets for changing lanes, taking exits, then add in everything for full FSD, then add in robotaxi (which will also probably need UI changes, not just NNs). It seems like they would continue to produce unique binaries for installation for the foreseeable future, but adding in dynamic loading would be interesting.
They may not need to download while driving, just load from local harddrive when something detects that it's needed. The map-like download seems like an unnecessary comlication.
Thanks. I’ve been watching a number of commentators gush about how great self driving is and my experience doesn’t match. I live in Salt Lake Valley which is flat but is framed by mountains. I also take trips though the desert SW and to the Midwest,Canada, SoCal… I can’t tell you how pissed I get sometimes on trips with the behavior of the automation. I’ll adjust my expectations in the future. For awhile lol. Again thanks.
It’s like a situation localization time map of the real world. Environment (snow, rain) location (long latitude) Time (Year, Month,day) Daylight (Day Night hours)
Like a human standing in the kitchen and remembers to buy some butter. But leaving into the living room forgetting to buy butter.
I believe this will be the holy grail of GAI. When the GAI can sens / explore new worlds locations.
It seems to me that your idea could be captured by adding some rows to the inputs for training that would be lat, long, month, temp, country, etc. Those would then rescale all other other values, and you still only to maintain one code base. A base to rule them all.
FSD already needs to be able to respond to emergency conditions for a whole lot of situations that don't occur normally. Continual modification of how FSD responds to the initial detection and the vehicle handling characteristics changes determines if a stored personality swap is needed to insure the safety of the passengers.
I don't know what the ratio of flash to RAM is in HW3, but my idea was to have the alternative modules in flash for fast swapping, or even sitting in ram but not processed unless they're part of the active set. They'd be updated by the normal OTA process. You could conceivably have paid DLC with names like "Ice Road Truckers" and "Verstappen's Race Day" where the training is sponsored (but probably still done at Tesla until Dojo itself is a product) by third parties.
Why would snow be a problem ? Add an occlusion filter to the input of existing video footage and feed back into the model… No different than transform sequences for training NNs…
I don't think the neural nets are that much data, if HW 3.0 can switch between FSD and NoA on the fly, in under a second. Cars can store all that data locally, IMO.
Exactlty what I ve been thinking. I live in south korea FSD needs to be trained differently
Overfitting
Robots Taxi will only need regional training and information. Cities need much more regional information than rural states that are more open road and more standard infrastructure.
Rather than administering all these location specific neural networks or fsd stacks it would be great if the cars could communicate vehicle to vehicle perhaps while parked in an area or even in transit. If you enter a new area the car senses from the surrounding fleet in the area that biases or settings can be adjusted to fit in. Probably easiest at a super charging location you get a notice....do you want to update to local area autonomy settings...?
just to add another to your list - New Jersey Jug Handles.
I think your argument is pretty far-fetched. Interesting speculation, but I don't agree.
Re Snow driving: It requires a lot of memory, road markings, lane marking, turn lanes can be covered with snow and you know the road from driving in the summer, which lane you should be in. Even things like stop signs can get covered by snow, but you know you should stop anyway. FSD only works on what it sees now, not last summer. Until it has this kind of memory, it will never reach Level 5.
I think this is a GREAT idea so long as it can be turned on/off as desired.
As a corollary, the specific driving pattern of the different driver. The local roads usually traveled, ie to and from work. Can the FSD learn that? Or are there any plan for Tesla do do this? Think robotaxi think local roadwork......
The many stack theory!
Hi. I am a human being. I have one brain that grasps and functions from my degree in biochemistry to driving my car, to riding my bike on one wheel, to a variety of construction tasks, and any other tasks I train myself to do. One Brain. 🤯🤯🤯🤯🤯🤯🤯🤯
Atlanta is an hour from Atlanta.
You can fine tune by 'the best qualified tesla driver' of each area (the most adapted, etc) .. as an update (or temporal update) while the car moves from one area to another.. We are recreating a human brain that drives.. or better , the software ("map") is what ultimately must be loaded given the geographical location it represents.. This answer will answer itself once Teslas are everywhere in the world .. I don't think the same way of driving will work for every Country.. The same happen with language models..
How large are the stacks ? Is a 1TB SSD drive enough ?
Snow and mud can be sort of similar to drive. Just my guess.
you'd think environmental inference was built in, it seems too obvious. eg infer from camera data the in picture atmospheric scattering levels to infer fog, rain, dust qualitatively measure the road conditions for slickness, roughness, snow, heat, gradients, read trees for windiness and then compensate to match. I mean doesn't driving in the dark naturally result in a different behavioural response envelope?
So you'll get the "Dukes of Hazzard" package 😂
I suppose it comes down to what data affects same functions according to conditions or area. Yes I think over fitting according to area and summer/winter would work a lot better because of the experience with so many Tesla FSD cars in San Francisco, over fitting the training there. It still may not be perfect but how much perfection do we expect to qualify for better than human? Does that mean if another area besides San Francisco has a lot of Tesla FSD cars, is the data over fitted there as well? If so how does the neural network training know what neural weights in the network to use, since ATM there only seems like one training to fit everywhere. Does it apply the San Francisco training in both areas or does it separate the training in its neural network into different layers then apply what ever layer is activated by the conditions according to area? If so this would be a bit like what the video here is getting at. It could be a problem though if only one area/layer in the neural network dominated the function in other area/layers, even when there was a lot of training by number of FSD cars there but less than the first area. Another way might be instead of layers per area in the neural net, have partitions per area in similar functions of same layer. That way a higher level layer would not be able to dominate a lower level. Having the weights of the whole neural network per area downloaded or adjusted might be better at least with present technology.
OMG Dr Know it All is advocating "geo-fencing"!
Just kidding, sort of.
In econometrics we just use dummy variables (or categorical variables or in R, factors) to allow slopes and intercepts to vary.
Personally, im more of a one stack to rule them all type of person. What i've beem thinking is what if theres that one event that never happens but suddenly happens once in a while (like snow in florida). The car needs to know how to handle it. Having overfitting stacks by regions would make it maybe (in my opinion) less capable when such event would occur. Also, Im a true believer that in 20-30 or so years, humans wont be driving cars manually at all. having different stacks to perform better by region will become non relevant as the cars will be the only ones making the driving and it will all become uniform driving everywhere. I do agree that there will need to be different stacks per countries to fit different road rules but I think that at some point, Governments might change some of the rules and make it easier for autonomous cars to deal with. I mean what if Canada's KM system gets converted in miles like the US. No biggie cuz by then , no one will be using the signs anymore apart from the cars. It can be confusing to humans to change the whole system but for the car, it would be instantly perfect with the new system/rules. same goes wich for example Europe's left side of the road driving.
Yeah those people in Nepal have some very strange habits you see cows people motorcycles cars buses everything going down the road together and it's a chaos
Thanks John seems like a lot of downloading.
In Dallas it snowed for days are going to have a fleet stuck and gridlocked? It doesn’t make sense!
Could local weather forecast somehow be an input at runtime?
I have no idea how Tesla uses NN but I wonder if it could upload the data it has seen per car, which could be aggregated to per locale and use that for the NN approximators, do a reinforced learning and download it back to the car. Instead of something like a supervised learning procedure and once in 2 weeks updates
What if roads are snow covered with no markings and ditches are filled with snow.
i do see one problem with this is that on a long road trip it might freak out the driver useing fsd and then the car crosses a state line and now does stuff the driver does not know about that the car will now just do
.
Maybe a kluge necessary to make HW3 work instead of replacing with HW4 which presumably will host bigger models. Personally, I want my car to know how to drive anywhere on Earth, without reference to an external network so my Cybertruck will work in the apocalypse, after satellite constellations have been blown up.
FSD in its current state will put you in the ditch in seconds in bad weather. Slippery roads are a whole new challenge. A passing truck can cover you and your cameras in freezing slush in seconds. With no means of self cleaning for the cameras, the existing hardware will never be FSD or Robo taxi ready. While the heat pump is a low energy way of generating heat it is dangerously inept at window defogging in high humidity conditions. You have to refrigerate the air to dry it and then heat it to defog which good winter cars do with simultaneous heat and cool. Volvos have done it this way for decades. Heat pumps can't do both at the same time. You need a PTC heater too. The front air dam (model S) is remarkably adept at moving large volumes of snow from the road to the windshield blinding your view and the cameras. Raising the suspension is remarkably effective at preventing this and increasing safety but it self lowers beyond your control. It's a California car not a winter car.
Please provide your evidence that FSD puts you in a ditch in bad weather
Maybe just leave it to Tesla to work out what to do. They’ll figure it all out in time.
Multiple stacks only makes sense if single stack is not capable enough.. That remains to be seen.
I am not sure I agree. Could all the stacks be encoded in one large neural network that understands its environment and which part of itself to use where? That's what humans do when they regularly drive in different areas.
If the differences between 'stacks' are relatively small, the car could have hundreds of them already downloaded and stored. It could load the most appropriate weights at any time.
How would it know which one to load?
@@fredbloggs5902 That would be the challenge, but the car knows where it is and the weather and temperature, so it has all the knowledge needed to decide which parameters to load.
I don’t necessarily believe all these human experiences will translate effectively to machine driving. I expect a 10% larger network could add these specialized behaviors to a single network with logic to switch over based on inputs.
How many gigs would a neural network be?
That’s the wrong metric.
So, how good FSD Beta is will be dependent on how many charging stops you make. You could completely drive across a state without charging in a Model S so you would get no benefit from this kind of system. Or you could be at a low elevation in Colorado with sun and drive up a mountain where it is snowing and the car would not know how to handle the snow because you made no charging stops?
Lol not bad for a loon just kidding. Got a question about your added display on your Y how do you like it and would you recommend that one or another display
First, how many times (with worries and troubles) have you "updated", say, Windows or OS, etc.? How long does it take (over spotty wi-fi), with reboots, etc? Updating my car on the fly seems flaky. Then consider millions of cars on the road, all trying to do this - what a pain and expense and bandwidth that would be for Tesla or any company. IOT generally involving non-critical, "small" chucks of data passing back and forth through the Internet to and from the clouds. If you want many stacks for trips and weathers, file a car "flight plan" ahead of time, download all the stacks at once to the car, and let the ego determine which to use when and under what conditions.
Zip-codes / Postal codes.
I thought FSD did this already in San Francisco? It acts as a noob because Tesla wants the general stack to learn the ropes, but 'local memory' has been turned on in San Francisco which is why it over-fits there.
Please provide your evidence that ’local memory’ is turned on in SF (or even exists).
@@fredbloggs5902 It was mentioned on the video "Tesla FSD Beta Wide Release with Chuck Cook" on the Brighter with Herbert channel.
@@James_Ryan Thanks for the reference, I’ll take a look.
If you paid for the whole thing why not get it? If you ever take vacation to the North and get into snow, what then? ask for the download?
Note that Black Tesla CZcams channel has said many times that Elon has mentioned FSD Beta is optimized for use around the San Francisco area, not his area which is near New York city in the North East.
I used to watch all his videos but haven't seen one in a long time. That's... Just subscribed to Black Tesla so I'll not miss them again.
Relying on realtime updates does not sound realistic and reliable. Maybe some of the transformers could be plug-in related and all in the firmware but swapped in and out of in use on the fly. Maybe just weights files to swap on the fly according to conditions. OTA is too unreliable for on the move adjustment.
I agree with weather and regulations....not sure I want my vehicle driving like the idiots in the next county... they tailgate and are very aggressive....we have names for them...
I see plenty of beta testers not reporting small events on their CZcams channels. In my view, if they don't, it will never be as good as a human because they keep overlooking the small things that would have been an opportunity for learning to happen. In my view, FSD will be held back by the testers. Then there is the other problem. You have to pay for hardware, for someone else to collect your information (that you are willing to hand over for free) only on the promise that FSD might happen one day. How is that even fair? The machine learning, excelled fast, because it was exposed to 80% of the conditions. And that is where it will stay, for a very long, long long time.
This is a great video. But question; wouldn't the training modules train the AI to be excellent at all conditions? The AI would learn all of the difference without having to load different driving modules. Wouldn't they be aiming for a single AI brain that understands all nuances? Once stable diffusion trained on all of the art it's now available without having to load new models.
I was thinking something like this too. But it occurred to me that there are physical limits to consider like hard drive space, ram, processing speed. The Tesla is pretty limited in these so while it might be possible to design a "great at everything" stack, it might be too large for the Tesla to handle. Breaking the larger stack into smaller, specialized stakes may be the way to go.
The Michigan Left. Well known, ubiquitous and studied by other states.
(it only needs to be better than human -and humans crash all the time )
I think it would be too complicated to administer this. Unless you are memory constrained a single solution is simpler, and Elon always favours simplicity…