, 22 tweets, 8 min read Read on Twitter
professor gato here.

today's topic: why $TSLA FSD is a complete fiction, why the architecture is a fraud, why the basic methodology can never work, why their "data" is not data at all, and why advanced summon is not coming soon (or ever) because parking lots are too hard.
the $TSLA approach is a brute force attempt to solve what radiologists call the "aunt sally problem"

it's trying to use image sort and recognition. this is a complete and total dead end

it's the classic example of something humans are REALLY good (and machines really bad) at
elon's claims that "humans just use their eyes to drive so we don't need lidar" misses the key point that ABSOLUTELY NO ONE KNOWS HOW HUMANS DO THIS.

this is where aunt sally comes in

it's the key construct here on image identification and it's, frankly, a total mystery.
YOU can always recognize your aunt sally. if she changes clothes, puts on a hat, is in bad light, turns to 3/4 profile, it's no issue for you. there she is. you know her

a machine does not. theoretically, this is b/c humans are not really "seeing" her the way a camera does
you're barely seeing her at all. the res sucks, etc. most of it is extrapolation & fill in. humans miss enormous amounts of what they see yet wind up visually aware of lots of stuff they did not, in fact, register. it's weird neuro wizardry

your visual cortex is not a camera
it's not clear quite what it is. but it is VERY good at plotting the relevant data onto the physical world to allow you to navigate that world.

no computer is

so the idea that this camera input is just like human eyes but better is both wrong and irrelevant

it's the wrong data
worse, we have no idea how to process it.

how do we know? radiology. why is a lung MRI or CT not read by a machine to look for cancer?

because a machine cannot spot w/ even rudimentary reliability that which is instantly obvious to a radiology intern.

humans see the pattern
they see aunt sally. now consider what a staggeringly easy problem this is compared to a parking lot. it's one, static image. you know what it's supposed to look like and what you're looking for.

and they STILL can't do it.

a parking lot is a million times this complex.
not only are there many & unpredictable kinds of objects in unpredictable configurations & angles, but much of it is moving

when you drive in a parking lot, 80% of what you do is modeling intention

moving around parked cars is easy

guessing what moving things will do? hard.
will that car pull out? does the driver see me? is that pedestrian going to cross? is it a child or dog, prone to sudden, poorly planned moves? are they waiting for that spot? did they wave me through? are they on the phone? the ballet of shared intention is VAST.
and it's unteachable to a machine. we have NO IDEA how we do it. you know how to drive and where to be because you can model what other people are going to do. if each were just a cube in a game, you'd have no idea. all that info would be gone. that's FSD.
all the $TSLA "data" is not data at all. it's noise. it's just unlabeled, unidentified images devoid of context or useful information. nothing in that data can teach driving because it lacks the information inherent in driving.

consider watching a bruce lee movie
would this teach you kung fu? you see it, you watch the moves. does it let you do them? only in the most rudimentary sense. you say "ahh, block the punch, hit them back"

ok. but how are you to know what punch was coming? the movie is no help there. you cannot anticipate
"and then the driver slowed down & drifted left" is not useful data.

why did they do that? what input led them to anticipate this being the proper course?

video images tell us nothing on this score

knowing a fighter circled left is far less useful than knowing why
driving, like fighting, is 90% anticipation. by the time that punch is on the way and you see its target clearly, it's too late to block it. by the time that car is in your lane, it's too late to swerve. everything is a last second attempt to dodge a corner case

that's death
you need to see that punch coming from the hip, from the set of the feet, from the clenching of ab muscles and shifting of weight. just tracking the fist moving means you'll just get a good look at getting punched in the face.

you CANNOT drive that way. it's insanely unsafe
this is why even if $TSLA did collect all the data they claim to (and they don't, not even 1% of it) it WOULD NOT MATTER.

it's just more fist tracking with no ability to anticipate one.

and even that is too generous, because, of course, of dear old aunt sally.
you have to recognize a fist to track it and respond. you have to see it, ID it, and respond. but what if the fist comes from a funny angle? what it it's masked by rain or fog or shadow or sunglare?

then you drive right into a firetruck.

this whole architecture is a dead end
for it to work, you would need to solve an aunt sally problem a million times as complex as the one no one can solve in radiology.

THEN you'd be ready to START on the hard part or trying to teach context and anticipation.

this cannot be done.
this whole strategy is what a stupid, arrogant fake would dream up. "we're doing it like humans do it" is such a stupid claim as to beggar belief

humans can tell a shadow from a truck. computers cannot. this is why every sensible FSD company uses lidar, because lidar CAN tell
musk won't add it because he knows his FSD is fake and lidar would add to car costs. to keep sales up, he's pushing untenable and deadly tech. the l2 smart cruise tech is a dead end for L4-5 FSD.

$TLSA will NEVER get there. not on this architecture.
any claim to the contrary is pure lies

the company hides its data on actual AP safety then makes outlandish claims it refuses to support empirically

this "product" needs to be taken off the highways

if you wanna risk your life, go base jumping. don't risk mine w/ janky tech
Missing some Tweet in this thread?
You can try to force a refresh.

Like this thread? Get email updates or save it to PDF!

Subscribe to el gato malo
Profile picture

Get real-time email alerts when new unrolls are available from this author!

This content may be removed anytime!

Twitter may remove this content at anytime, convert it as a PDF, save and print for later use!

Try unrolling a thread yourself!

how to unroll video

1) Follow Thread Reader App on Twitter so you can easily mention us!

2) Go to a Twitter thread (series of Tweets by the same owner) and mention us with a keyword "unroll" @threadreaderapp unroll

You can practice here first or read more on our help page!

Follow Us on Twitter!

Did Thread Reader help you today?

Support us! We are indie developers!

This site is made by just three indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3.00/month or $30.00/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!