linux-works
Active Member
with it being a re-write and all, maybe they should call it, uhhh, "noa's arc" ?
(sorry, I'll let myself out, now)
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Elon: "Feature complete for full self driving this year"
- Thread starter dckiwi
- Start date
with it being a re-write and all, maybe they should call it, uhhh, "noa's arc" ?
(sorry, I'll let myself out, now)
powertoold
Active Member
It's funny that we think a month between software updates is a long time (last was Jan 26). My phone gets less updates 
diplomat33
Average guy who loves autonomous vehicles
It's funny that we think a month between software updates is a long time (last was Jan 26). My phone gets less updates
That's because Tesla spoils us with so many goodies.
But I think you bring up an interesting point.
Is it good that Tesla gives us "beta" software but improves it frequently with OTA updates or would it be better if Tesla only did OTA updates 1-2 per year but delivered finished, fully validated features?
Last edited:
Without the interim releases, how could they every get enough testing to created fully vetted finished versions?
diplomat33
Average guy who loves autonomous vehicles
Tesla could still use the Early Access program for testing. Plus, Tesla says they use Shadow Mode to test features stealthily:
"Features in “shadow mode” run silently in the background without actuating any vehicle controls whatsoever, which enables Tesla to test how features will perform in real-world driving conditions before we deploy them to the customer fleet."
So, Tesla could still use Early Access and Shadow Mode to test features even if they only did releases 1-2 per year.
That would help, but EAP only includes a small subset of the full fleet, and a Shadow disagreement doesn't necessarily mean the NN action was wrong.
Not necessarily. They can release anything and call it fsd v1. The invented the term and can thus define it how they wish and evolve it over time.
DanCar
Active Member
Full self driving has meaning to everyone. Trying to change the meaning is shameful in my opinion.
Based on own experience developing leading software technology, I'd say no.
We release even more often than Tesla with about 4-8 releases every week to critical applications. There's a reason for that, it's to always keep a solid baseline where you always know that your code is working well in real life use, and being able to quickly perform corrective actions without first proceeding weeks/months in the wrong direction.
Speaking for my own project, there is no way we could release software just 2 times a year and keep the same pace as we develop now. You loose a critical part of iterative progress. And you would need a lot more management and testing/QA overhead. Rapid releases means good-bye to traditional merge-days.
A plus is that rapid releases increase the flexibility a lot. You can literally prioritize a task and turn it into a released feature/bugfix within days when a business situation or opportunity requires it.
diplomat33
Average guy who loves autonomous vehicles
Thanks. I was curious because I know some people complain that Tesla releases "beta" features that never seem to leave the "beta" stage. So I did not know if people might prefer fewer updates but updates that actually released non-beta, finished features.
I think the software would become a lot more "beta" if they tried to keep the same pace, but with less releases. Because each release would contain more changes never tested with real users.
They could however start an opt-in Beta/Preview/Fast- ring mode for people that want access to new features quickly and don't mind bugs. And only release software already tested in Beta-mode known to be good in Standard-mode (slow ring). With Tesla's fleet it should give them enough users in the Beta- channel to still keep a good baseline, while reducing the number of releases for normal users.
diplomat33
Average guy who loves autonomous vehicles
Tesla to deploy more functionality quicker after Autopilot core rewrite coming soon, says Elon Musk - Electrek
In addition to the above, Elon has previously said this new rewrite would benefit from testing in the dojo. I understand that will not be ready until the fall. Does that mean we will see two significant steps? One this year with the new platform, and another next year, when performance gets improved from training in the dojo.
Of course both steps would be gradual with improvements over time, and a major rewrite will likely have some short term regression in some areas, and new bugs to deal with.
Personally I think feature complete FSD is still a couple of years away, but it is going to be a fun ride getting there.
In addition to the above, Elon has previously said this new rewrite would benefit from testing in the dojo. I understand that will not be ready until the fall. Does that mean we will see two significant steps? One this year with the new platform, and another next year, when performance gets improved from training in the dojo.
Of course both steps would be gradual with improvements over time, and a major rewrite will likely have some short term regression in some areas, and new bugs to deal with.
Personally I think feature complete FSD is still a couple of years away, but it is going to be a fun ride getting there.
diplomat33
Average guy who loves autonomous vehicles
In the recent podcast interview, Elon mentioned that dojo will probably come online towards the end of this year.
clydeiii
Member
Releasing only 1 or 2 times a year goes against Tesla's development model, which is to push updates to the fleet and see how they perform in real-world conditions. They need to iterate that over and over again, and 1 or 2 times a year simply isn't fast enough.
So basically they are still at the stage Google was at about 5-6 years ago, trying to train their image recognition algo (not really AI) to recognize objects.
That reminds us of a fundamental weakness of the system. The algo only understands a limited number of things. If it hasn't been trained to understand say a tree branch lying in the road then as far as it's concerned it's not there. Most of what is in its field of view it doesn't recognize.
And with no lidar to tell when something is a solid obstacle it's going to get confused easily.
That reminds us of a fundamental weakness of the system. The algo only understands a limited number of things. If it hasn't been trained to understand say a tree branch lying in the road then as far as it's concerned it's not there. Most of what is in its field of view it doesn't recognize.
And with no lidar to tell when something is a solid obstacle it's going to get confused easily.
Two words: drivable surface
If you can detect a smoothish flat surface, you don't need to categorize every possible object/ obstacle/ hypothetical that could be in your path.
Counterfeit Detection (Part 1) - Tim Challies
diplomat33
Average guy who loves autonomous vehicles
I don't think it is quite that simple. The driveable space is not always going to be a smooth flat surface like a road. In some cases, the driveable space might be a dirt road or even a grassy area that are definitely not smooth flat surfaces.
The driveable space is really any space where you are supposed to drive on. Context matters. If you are driving on I70 the driveable space is the lanes, denoted by lane markings. If you are driving in a residential neighborhood, the driveable space is the asphalt area in between curbs or driveways. If you are driving in a very rural area, the driveable space might be a dirt road. If you are driving to an outdoor concert where the parking lot is full, the driveable space might just be a cordoned off grassy area.
Last edited:
Similar threads
