Bears itt: Can you give me a theoretical reason for why FSD can never work?

i think the question is incomplete bc there are lots of outcomes where the technology works great but FSD makes no money due to regulatory/political/legal shenanigans, very long development times and the worst set of competitors imaginable

tough gig

FSD makes money right now, $15k/user or $199/month. If they get it to work well, take rate will go up even if government allows robotaxi or not.

Human eye is 576Mpixels, without that resolution FSD can never work. That used to be the bear thesis here.

This statement is everything that is wrong with Tesla.

they bound themselves to their current hardware set by promissing and selling the product years ago.

which, from a technology viewpoint, was dumb.

but obviously the whole thing was just about making cash and not producing a viable technology.

chip is now 4 years old, cameras even older. waymo, cruise, others are already operating without drivers and will beat them to market - in a sense they are already doing it right now.

they are driving people around phoenix or san francisco without a safety driver and - most importantly - taking legal liability for it. until tesla does this (which they obviously can't because FSD is garbage) they are not even competing.

let me make it clear:

TESLA IS NOT EVEN IN THE FULL SELF DRIVING RACE AND HAS NOT EVER BEEN.

at some point there will be no reason to tinker around with decade old tesla hardware (if tesla even still exists in it's current form).

they can't even drive through their own las vegas loop autonomously. the whole discussion is so insanely pointless and dumb. the time where i could even laugh about fsd ended years ago. at this point it's just a sad spectacle.

tooth, as insufferable as he is, did a few really good posts on this

we have probably reached 98-99% fsd

you may have noticed recently that most captchas these days are noticing crosswalks, cats, bikes, stoplights, etc

these are all for training fsd

early captchas were done to digitize books that had obscure fonts which the algos didn't recognize and couldn't read - they'd give you two options - one which was the actual captcha and another that they figure if they give to 1,000 people and get over 900 typing the same thing then that's probably what it is

likewise, when you're picking out the crosswalks today, it doesn't actually know if you're accurate or not - it has an inkling but it's really paying more attention to your mouse movements and speed of clicking than what you are actually clicking to determine if you are a human or not because this is more about training it to learn to recognize crosswalks better

this has been going on for years, so we've had millions of clicks on crosswalks each day and yet we still need more - because there is no "good enough" in fsd, it needs to be perfect, there's no excuses where if the crosswalk is painted differently or in a new pattern etc that the fsd camera doesn't recognize it



it needs to be where a town can come up with their own brand new style that's not in a database but the fsd can still recognize it and it currently struggles to do that, hence why you still need to click on crosswalks

I worked for a social media company that shared photos, it was much like instagram where you could tag and label various parts of the photo - it was mostly done by people to highlight the brands of clothing in the photos - this made our tech at the time, one of the better image recognizing AIs in the world because not only were we getting tons of data but it was done by people using natural language

after about a year, we finally launched a tool that helped suggest tags, i'd post a photo of a basketball and it would reference all the photos that had similar imagery and suggest tags for me such as basketball, nba, RIP Kobe, hoops, one on one, etc

but it could still get confused at times and give out a curveball like "ass to mouth" so we soon shelved this - much like how many twitter AIs were shelved because they eventually went racist

this was about 8 years ago, i'm sure there's been a ton of improvements

but again, the main thing is this requires perfection and going from 99% accurate to 100% will take more resources than it did to go from 0% to 99%, the human eye can see a plastic bag getting blown across the highway - immediately process that it is not a threat and ignore it - in fact, we're so good at this that nearly all of it happens subconsciously - even to the point where our hearing gets worse and worse as we age, not because our ears degrade, but our brain recognizes that we never need to hear those frequencies so it tunes them out and ignores them. We've even shown that children growing up in conditions where hearing certain frequencies remains important will continue to hear those as they age because the brain won't start tuning them out.

so the fsd is going 75 on the highway, recognizes a plastic bag being blown across but only has 1 second to act - if it plays it safe and slows down or stops then it endangers the cars behind it not expecting a car to stop for a plastic bag and if it just ignores it, eventually it won't be a plastic bag but something that should have required the car to stop such as a toddler

likewise, if driving through a city, fsd can't tell apart people waiting to cross the road or get into their car or hailing a taxi and thus needs additional clues like recognizing crosswalks

we can probably get there - but it'll be more to do with the fact that we're going to have to accept that it won't be perfect and roll with the consequences - meaning that it'll forever be a "hand on steering wheel ready to take over" situation or "highway only" etc etc - the dream of robo cars and getting home from work and deciding to earn some side income by pressing taxi mode where your car goes out on its own to pickup and drop off passengers and then be back by morning to take you back to work is just nowhere close to fruition

we've had self driving cars since the 80s, they usually were competing in races in the desert sponsored by the military and they occasionally got stuck because the AI couldn't recognize what was ahead so it went into a hazard or just stopped - but many of those cars did complete the courses

the issue isn't a pushback against fsd, but rather things like robo taxis, which probably won't happen anytime soon barring some incredible breakthrough or mass overhaul of our transportation network built around the weaknesses of fsd AI - while we are ripe for a change given we haven't really built any new highway infrastructure since WWII, it'll be a slow and gradual process and if it does happen it'll probably be mainstream in other countries for a few decades before we adopt it

china is going full bore on this, especially for trucking, I could definitely see China building fsd friendly highway networks while we wait to see the results and learn from the mistakes they make along the way

i can fly to san francisco and order a waymo or cruise to drive me around town without driver right now.

what the hell are you talking about?

you misunderstood, perhaps that's my fault

waymo is "level 4" meaning it can only handle certain situations

for restricted areas where they currently operate this isn't that big of an issue because you can specifically train it to handle the few problem areas in that range or just avoid them

it is not level 5 where it can handle all driving conditions

this is what I was talking about with crosswalks, it's one thing to recognize crosswalks in phoenix and sf, it's another to recognize all crosswalks and if one were to be run without specific manual inputs in another city that used different formats, it may not recognize them

this was what i meant by making the entire road system fsd friendly using uniform systems etc etc, i imagine boston will be one of the last places to have fsd because most roads are former cow paths and thus they have an abundance of winding roads and 6 way intersections, frequent changing of lane rules (ie the right lane will regularly switch out between right turn only or not etc etc

that's what i'm talking about, you take a waymo and drop it in boston right now and you're going to have major problems

having said that waymo has had a number of accidents, fortunately driving in downtown locations at slow speed these haven't been a big issue

i am 100% sure we will have competing robotaxi services in every major US, most european and some asian metropolitan areas by the end of this decade.

i am also sure, that none of those competitors will be using current tesla fsd technology.

that's the biggest market opportunity for full self driving and it will be taken in the next 7-8 years.

ie waymo cars were driven extensively through every single street in those cities with teams in the car to take over driving when needed and others to analyze all the data to ensure the fsd was both recognizing everything and interpreting it correctly

waymo collected more than 20 million miles of testing in san francisco before they went fully driverless there - this is not a practical approach and barring any breakthroughs, it's going to be a slow rollout where affluent and high population regions with friendly legislators are going to be the priority - someone living in a suburb of milwaukee is probably at least a decade away from being able to use a service like this

if the investment is there then sure, but these companies will hemorrhage money left and right as they expand - uber is still not a profitable company because of all the costs they have with expansion and these companies have a far greater initiation cost than uber ever will - their profits will be higher though because they cut out the middle men of drivers so long term i agree this is huge opportunity but unless some company makes that breakthrough to true self driving, it's going to be an extremely costly expense up front for each region to train the fleet and deal with legal/regulation - and we all know just one mistake where a waymo car kills someone and they could find their license getting suspended pending an investigation

yeah... i'm sure google and ford won't invest in this just because they can cut out the driver which is like 50%+ of an uber ride...

whatever... i can't deal with the tesla cultism anymore...

where have i ever displayed that?

your whole argument is just grasping at straws. i honestly don't know why.

yes, waymo has driven 20+ million miles. but over the whole last decade to get to where they are now and not just to adept to san francisco, . if you think they will need 20 million miles for every new city you're just wrong.

why anyone would use uber as an example is completely beyond me.
the driver is 50%+ of an uber fare and the whole idea is to replace the driver. comparison makes absolutely zero sense.

and with that, i'm out.

you're a coward, toss out insults that don't even make any sense (i'm an uber bear and there's ample historical evidence of that here in the thread you numbnuts)

you suck at reading comprehension and just toss out empty calories and when someone refutes them you go to personal attacks and then go "i'm out"

real tough guy you are, pathetic

Why Tesla FSD can never work? FSD in general?

FSD seems to have a future. I personally think isolated FSD is folly though. A mix of real drivers and FSD, or even separate FSD systems doesn't seem like it solves anything. You need an interconnected fsd network where everything is communicating.

Tesla/Waymo style FSD seems perfect for shipping and frivolous for personal driving

Based on nearly every independent research I've read this is 100% facts.

I think maybe in 2017/2018ish Tesla had a top 5 product because there were like 8 competitors total. They have fallen at the bottom of the pack ever since.

Tesla fan boys are amazing marketers though

The hypothesis is that the current hardware is sufficient and with improved software can reach FSD. I don't think your post disproves this. Tesla's main advantage is not the hardware, it's the dataset used to train the software. Tesla has by far the most number of cars gathering useful video real world data to train their neural network, which Tesla thinks is the key to getting good performance.

There is some debate if there is some scenario where Tesla cannot see around an obstacle, I have yet to see any good proof that this is an issue and the current creeping and ghost objects seems to solve all scenarios human drivers can solve.

HW4 with faster neural processor, higher resolution cameras, fewer(!) cameras should be out this year. But HW3 seems plenty fast still, the competition from Mobileye I have seen seems to have either have lower performance, higher power draw or not be out in large scale yet. Not sure what Waymo is using, but google's TPUs are mainly good at batch processing while real time systems are a batch size of one problem, so doubt they will help Waymo in the cars.

there is no data

Consensus seems to be that it should be possible to make FSD work in theory, but bears mostly think that TSLA won't be the one winning this battle.

What if they manage to get ~3% market share of FSD business worldwide. This is probably a conservative estimation given the size of their current fleet and the amount of resources they put into this? Sure, you can think there's 0% chance that they'll make it, but what if. How much would this roughly be worth?

what's that amount?

somewhere between nothing and a lot, and i think anyone who gives a more specific answer than that is kidding themselves

I don't the figure will be static, it will quickly become a race to the bottom. ARK estimate it to be ~$11T:
https://medium.com/@TashaKeeney/ark-...n-6dba75d6bb49

(I am not a huge fan of ARKs investments, but they do present their research at least, which is better than most)

Teslas FSD in one picture:

Purchased more Tesla today at $158 and Shorted more GM at $36.