FSD rewrite will go out on Oct 20 to limited beta

[Bladerskb]

One would think there will always be some traditional C++ "software 1.0" style code even many iterations later as that code is what passes data into the neural network to evaluate and get outputs. AlphaZero style AIs have traditional code to aggregate neural network value and policy predictions to do even better by having the network evaluate potential future positions. Similarly FSD beta uses traditional code to decide which predictions to use, e.g., pick a path towards the right for an upcoming turn, or ignore, e.g., don't drive through a traffic calming island even though the network believes driving straight is the most common behavior.

MuZero is AlphaZero's successor and reduced even more of the traditional software that would have provided truth of what can happen in potential future states (e.g., whether a move is allowed), and there the neural network became much larger and learned to predict how various games (Go, Chess, Atari) would behave when doing certain actions. But even then, there was still traditional code to decide how to aggregate the various predictions and do the decided action, e.g., move a piece or steer or accelerate. People have speculated more "end-to-end" training of Autopilot neural networks where the neural network gets sensor and other inputs and directly decides steering and acceleration, and that presumably would replace a lot of traditional software, but that is most likely quite some time away.

From watching FSD beta videos, it's pretty clear to me that there's plenty of of traditional code for current & upcoming lane selection behavior, but how it gets into or "stays" in a lane (e.g., avoiding objects) is primarily neural network predictions as suggested by the wobbly path visualization.

One would expect Autopilot to move more of the lane selection behavior into the neural network especially in situations that are misbehaving due to incorrect map data, but indeed we have no insights into when that would happen and even when it does, how much is still left as traditional code.

None of this is true. But making things up and spreading bs

 

[ZeApelido]

Does anyone else here even work in machine learning? It's a heck of a lot of work to get some fancy neural net to work in production. Just to replace any one module of functionality.

Did people suddenly expect Tesla to replace all the modules for operation in production? That's crazy. This isn't a research project.

And it's odd to bash the accuracy of the systems while also showing why its limited. You are inadvertently being optimistic, there's plenty of room for improvement if so much of the system still needs to be converted to 2.0 code.

So basically, really the only complaint is how long it's taking.

Which if there was someone else out there with a FSD system consumers can buy and use right now, would be more pressing. But we aren't there yet.

 

[Mardak]

One thing people have often gotten confused about neural networks is believing they're like a database with memorized behaviors for a given situation.

I was catching up on @jimmy_d old tweets from last month to see if there was any update after exploring FSD beta neural networks, and I noticed this same "Neural networks don't memorize their training data. We know this because they generalize to unseen cases." comment from just yesterday https://twitter.com/jamesdouma/status/1354428378812735490

Does anyone know if there has been a followup writeup or video somewhere? I would be quite interested in how the neural network structure changed during the FSD beta, e.g., more outputs from the BEV net that could take on more traditional code decisions such as right-of-way, as well as more insights into the individual network outputs especially BEV net. I wonder if the networks analyzed were from regular public releases which could perhaps confirm the wide fleet is running these networks in shadow mode for data collection.

 

[Bladerskb]

Anyone remember this? This is coming from the guy who has been saying Level 5 will be here, you will be able to cross country with no driver, etc, by the end of the year since 2017.

Elon's amazing way of creating hype. Its only amazing because his fans are so misinformed and gullible. So here he is literally calling the regular development process = "major unexpected setbacks".

 

[turnem]

Anyone remember this? This is coming from the guy who has been saying Level 5 will be here, you will be able to cross country with no driver, etc, by the end of the year since 2017.

Elon's amazing way of creating hype. Its only amazing because his fans are so misinformed and gullible. So here he is literally calling the regular development process = "major unexpected setbacks".

There's no arguing that Elon is always overly optimistic (maybe on purpose maybe not... doesn't really matter). Most people with a brain know to take his timeline projections with a grain of salt.

Having said all that - I still think that there is a good chance that we will see a broad/public release before the end of March.

 

[willow_hiller]

Anyone remember this? This is coming from the guy who has been saying Level 5 will be here, you will be able to cross country with no driver, etc, by the end of the year since 2017.

I eagerly await the Level 5 and Level 4 autonomous vehicles that BMW and Ford (partnering with Waymo) promised us would be launching this year; because corporate predictions made by companies other than Tesla are always correct:

BMW to Unveil Level-5 Self-Driving Car by 2021 | BMW of San Antonio

While it's no secret that BMW hasn't been at the forefront of conversations about autonomous technology, no one else seems to be developing a fully-functioning Level-5 self-driving car, either.


The levels of vehicular autonomy range from 1-5 and increase in their abilities to drive themselves without human interaction. Simply put, Level-1 needs complete human operation, whereas Level-5 is full autonomy, with the ability to navigate any street or situation without driver input. In Level-5 concept cars, automakers actually have the option to put the steering wheel back into the dashboard because of the lack of necessity for it. While this type of technology seems impossible in just four years, BMW claims that they will make it happen.


"We are on the way to deliver a car in 2021 with Level 3, 4 and 5," said Elmar Frickenstein, BMW's Senior Vice President for Autonomous Driving, at a panel discussion in Berlin earlier this week.


BMW is certain that by 2021 they will have different levels of autonomy for different situations and customers. For instance, a customer with a long work commute might opt to pay for a Level-5 autonomous car versus someone who travels less frequent.

Ford partners with Google and Uber veterans at Argo AI for self-driving cars

Last year, Ford announced plans to develop a mass-production autonomous vehicle for ride-hailing and other commercial applications by 2021. It’s already deep into road testing for autonomous versions of its Ford Fusion Hybrid.

The cars that hit the streets in 2021 would be capable of SAE Level 4 autonomy, which means the virtual-driver software should be able to handle typical road conditions unassisted.

 

[82bert]

Anyone remember this? This is coming from the guy who has been saying Level 5 will be here, you will be able to cross country with no driver, etc, by the end of the year since 2017.

Elon's amazing way of creating hype. Its only amazing because his fans are so misinformed and gullible. So here he is literally calling the regular development process = "major unexpected setbacks".

We get it. Your agenda is clear and has been for some time.

 

[ZeApelido]

Anyone remember this? This is coming from the guy who has been saying Level 5 will be here, you will be able to cross country with no driver, etc, by the end of the year since 2017.

Elon's amazing way of creating hype. Its only amazing because his fans are so misinformed and gullible. So here he is literally calling the regular development process = "major unexpected setbacks".

Advanced algorithms used "adaptive learning" techniques to adjust their predictions when the data is consistently biased in one direction.

I think many members here have brains sufficient of adaptively scaling Musk's timelines into more realistic ones.

 

[powertoold]

I keep parroting this, but Tesla is still the only car with a traffic light recognition feature in production. Not only that, but the Tesla can also stop for traffic lights (this is even another step up in difficulty, since the car has to perceive the stop line as well).

I don't think any other car out there has a consumer level traffic light recognition feature.

 

[mark95476]

I use this feature every single day. It recognizes when traffic ahead is going through greens and follows.

The green light chime is wonderful when I'm actually driving.

I keep parroting this, but Tesla is still the only car with a traffic light recognition feature in production. Not only that, but the Tesla can also stop for traffic lights (this is even another step up in difficulty, since the car has to perceive the stop line as well).

I don't think any other car out there has a consumer level traffic light recognition feature.

 

[EVNow]

So here he is literally calling the regular development process = "major unexpected setbacks".

You have zero information to categorically state that they have not encountered any "major unexpected setbacks" and what they have is the "regular development process".

Do you atleast recognize what is a "fact" and what is your "guess"/"speculation" ?

 

[Edgewood Dome]

Is there Any solid information on when FSDBeta will be rolled out to all NA cars? I thought it was to be mid Dec/2020? Been waiting since 1/1/2017.

 

[diplomat33]

Is there Any solid information on when FSDBeta will be rolled out to all NA cars? I thought it was to be mid Dec/2020? Been waiting since 1/1/2017.

No, there is no solid info at this point on when FSD Beta will roll out to all NA cars.

Dec 2020 was not the deadline for the full rollout. It was just a rollout to a small select group of beta testers within the Early Access group. At this point, FSD Beta will probably continue to go out to more Early Access testers as Tesla continues to collect data and improve the system.

FSD Beta needs to be more reliable and more safe before Tesla can roll it out to all NA cars. So it will probable be awhile before that happens.

 

[run-the-joules]

Is there Any solid information on when FSDBeta will be rolled out to all NA cars? I thought it was to be mid Dec/2020? Been waiting since 1/1/2017.

my guess at this point is sometime in 2022

 

[mhan00]

Is there Any solid information on when FSDBeta will be rolled out to all NA cars? I thought it was to be mid Dec/2020? Been waiting since 1/1/2017.

There is almost no chance of the beta going out to all NA cars anytime soon barring huge leaps in the next couple of updates. Musk said that on V11 or so it may go out to Early Access people, which is a much smaller subset than all NA drivers with FSD ordered.

 

[Mardak]

A good chunk of FSD beta errors right now seem to be due to incorrect lane count data from OpenStreetMap, and Lyft just released a report of their analysis of OSM accuracy:

How Lyft discovered OpenStreetMap is the Freshest Map for Rideshare

Currently the FSD beta path planner decides what lane it should be in based on map data to prepare for an upcoming turn, but it seems to fall back to assuming 1 lane in each direction if the `lanes` data is missing. This can also cause problems if the road actually has multiple lanes that FSD beta would assume all merge into a single lane across an intersection.

Perhaps Autopilot team originally thought the map data would be good enough with Silicon Valley nearing 90% accuracy, but after getting more private beta testers potentially ending up closer to the 30% map accuracy, they decided that moving the lane count task over to the neural network would be necessary before a public release.

 

[qdeathstar]

You have zero information to categorically state that they have not encountered any "major unexpected setbacks" and what they have is the "regular development process".

Do you atleast recognize what is a "fact" and what is your "guess"/"speculation" ?

Why does Tesla have so many “major unexpected setbacks” I feel like maybe they need better engineers or something, people who can foresee issues and fix them. At least, not be astonished and shocked by them.


As to the consensus that only idiots take what Elon is saying seriously.... one, think about what you are saying for a minute and whether you want to spend your money on someone you say you can’t trust... and two, all it takes is for one or two idiots to get a lawyer or two and sue for false advertising and with the way the courts are, win.

 

[tivoboy]

My question at this point would be, how many EARLY ACCESS people are there in the country (USA) and other countries, since I would imagine that the remaining EA owners would certainly be first.. historically, we had data points that 8% of Tesla buyers had purchased EAP/FSD, that was about 2.5-3 years ago. So, at the time that may have been only about 8-12K owners? If there are only 1K current in FSD, I imagine they might double then double again, maybe in 2 month increments, putting any further addition of non EA owners probably out to mid-summer at the earliest. Just a thought. Knowing Elon he could just release the hounds and open the floodgates with a tweet…

Alternatively, Tesla MIGHT odd some sort of survey, questionnaire, lottery to find the next batch of owners who could possibly add the best bang for the buck and value. Users with long commutes, city commutes, salespeople with varied destinations, etc. Maybe a ride share driver?

OR, maybe they do something like that with some incremental fee? An EA buyin for say $250, or $500? I mean, they don’t NEED the money with 17.5B in cash/CE and ~2B in bitcoin, but it would weed out the truly interested parties for sure.

 

[2101Guy]

one of the youtubers with FSD Beta, praised the progress being made, but in his estimation, I believe he said his best guess would be "months" not weeks, before any sort of broad release of FSD Beta level/version. He also went on to guess, that once it is released, most likely would be fairly loaded with confirmation requirements initially.

Again he was only guessing..but from what Im seeing, I tend to think hes correct. As we all know the ONE time someone violates the terms of service and uses the technology without remaining FULLY in control, and he/she crashes with FSD activated? Press will be ALL OVER IT fully blaming Tesla.

 

[EVNow]

Why does Tesla have so many “major unexpected setbacks” I feel like maybe they need better engineers or something, people who can foresee issues and fix them. At least, not be astonished and shocked by them.


As to the consensus that only idiots take what Elon is saying seriously.... one, think about what you are saying for a minute and whether you want to spend your money on someone you say you can’t trust... and two, all it takes is for one or two idiots to get a lawyer or two and sue for false advertising and with the way the courts are, win.

Kind of hard to follow your logic - but without the ability to be optimistic in the face of extreme risks - there would be no Tesla.

And those engineers have achieved things 100 year old companies can only dream of.