ZenRockGarden
Active Member
The dismissal of the human labeling team would almost certainly point towards them having auto-labeling working...
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Karpathy leaves - what's next for AP/FSD?
- Thread starter sids911
- Start date
The dismissal of the human labeling team would almost certainly point towards them having auto-labeling working...
black knight 3
Member
The straw that broke the camel’s back came when he was told he had to spend at least 40 hours a week logged into his assigned work site.
black knight 3
Member
I see that same way. Tesla needs CEO who don’t spend time on Twitter giving false information. Musk did that was needed. Made EVs popular.
Perhaps they are not finding any value in the labeling process at all. They might just be taking a step back to redevelop the hardware and reformulate their strategy for FSD. Keep a small trickle of work on FSD Beta so it appears they are making progress, but reduce their financial commitment. Or, we'll see, maybe 10.13 and V11 really will be amazing. It sure would be nice to hear some honest truth from an insider on all this.
"one stack to rule them all" is one of the biggest scam to showcase how gullible tesla fans.
Now they go around saying: "no one else has one stack".
Not understanding that others have had one stack from the very beginning.
That they didn't have to create and then nurse adas systems with lane keeping, acc and other features and maintain it to keep thousands of customers happy.
That their focus from the start has always been making fully autonomous system from day one.
Just like vision only and version 9...on and on. Elon each time proves how gullible people are because he comes up with a phrase/word to deceive his fans.
Oh gawd not again
joe.smith
Member
And, I was just about to say that I haven't seen any significant improvement (updates) to FSD in a while. Now, I have a new perspective; I'll cut them some slack. I'm dealing with the SAME environment here at my job (Defense Industry).
diplomat33
Average guy who loves autonomous vehicles
Karpathy would set project goals and deadlines for his team. But Elon would set the big goals and overall timeline though. Remember, it's Elon who is setting the big goal of "solve L5". It's Elon who is making promises that robotaxis will happen by X or now he is saying "safety greater than human" will be achieved by X. And when your boss is tweeting to his 100M followers on Twitter that the next FSD Beta update will be released next week, that put a lot of pressure on the FSD team to deliver.
I said back in 2020 (once I had lost all faith in Elon understanding vehicle autonomy and the challenges in achieving it) that I still had hope that Tesla had a vision for at least L3 autonomy as long as Karpathy was still around. I believe I reiterated that here when he went on his "sabbatical."
I think Elon and Tesla have understood for some time that "autonomy" in their consumer level vehicles was going to need to go a different direction. Achieving a "safer than humans" Advanced Driver Assist system seems to be the new target of "Full Self-Driving Capability" - not L4 or L5 autonomy or "robotaxis." I just think the last few months has been Elon trying to deal with the direction change without tanking the stock price and/or looking like an idiot.
His recent statements about FSD Beta pretty much already being at v11 is also interesting for me. I was waiting for v11 to give FSD Beta another shot, but sounds like I already have it. Not looking forward to having this code rolled into my Navigate on Autopilot, which I use quite frequently.
I think Elon and Tesla have understood for some time that "autonomy" in their consumer level vehicles was going to need to go a different direction. Achieving a "safer than humans" Advanced Driver Assist system seems to be the new target of "Full Self-Driving Capability" - not L4 or L5 autonomy or "robotaxis." I just think the last few months has been Elon trying to deal with the direction change without tanking the stock price and/or looking like an idiot.
His recent statements about FSD Beta pretty much already being at v11 is also interesting for me. I was waiting for v11 to give FSD Beta another shot, but sounds like I already have it. Not looking forward to having this code rolled into my Navigate on Autopilot, which I use quite frequently.
[Devil's Advocate/Counter point debate]That was said about Steve Jobs also and.........It could be good time for Tesla to move beyond Musk. Just like Jobs, Musk has laid the foundation and launched the compony on a near vertical axes. Also just like Jobs Musk is likely insufferable to work for. Would be great if Tesla had a "Tim Cook" to now professionally and fiducially lead Tesla into the stratosphere like Cook did with Apple. Drama and seat of your pants/pivot on a dime autocratic decisions may be great strategies for a startup/growth compony but it is starting to become an anchor holding Tesla back.
The L5 vs L3 thing is unfortunate in that Musk very obviously on a moonshot strategy, at the detriment of smaller incremental goals being achieved.
The improvements in AP/EAP/NoA from 2018-2020 were noticeable. The product has now stagnated for 2 years while the team focusses on the (likely near term impossible, especially with current vehicles) FSD city streets ability.
For the vast majority of people, if AP/EAP/NoA became truly bullet proof on the highway, above&beyond all competitors (many of whom have caught up).. we'd be pretty damn happy. If people could rely on it for their 30min highway commutes, dealing with the stop&go, and long road trips .. especially if hands free, it would be a huge accomplishment and selling point. Right now its good except where its terrifying, and never as smooth as even a novice human driver.
Most highway routes I drive have particular stretches that I always take over either in anticipation or once the car starts being unreliable. Not to mention that I've given up on any of the on-ramp/exit handling since they seem to have added it and never refined it.
Maybe it should be two stacks and they should have a ringfenced team continuing to work on the thing the vast majority of users have/want, rather than the stock pumping vaporware...
The improvements in AP/EAP/NoA from 2018-2020 were noticeable. The product has now stagnated for 2 years while the team focusses on the (likely near term impossible, especially with current vehicles) FSD city streets ability.
For the vast majority of people, if AP/EAP/NoA became truly bullet proof on the highway, above&beyond all competitors (many of whom have caught up).. we'd be pretty damn happy. If people could rely on it for their 30min highway commutes, dealing with the stop&go, and long road trips .. especially if hands free, it would be a huge accomplishment and selling point. Right now its good except where its terrifying, and never as smooth as even a novice human driver.
Most highway routes I drive have particular stretches that I always take over either in anticipation or once the car starts being unreliable. Not to mention that I've given up on any of the on-ramp/exit handling since they seem to have added it and never refined it.
Maybe it should be two stacks and they should have a ringfenced team continuing to work on the thing the vast majority of users have/want, rather than the stock pumping vaporware...
that is my issue too. Normal Autopilot in 11.12.2 has lost all of its smoothness and drives like a teenager again.
It's bad enough that I now use AP/NoA half as much as I used to, it's just not comfortable to use any longer. Too many pointless lane changes and no smooth corners.
If all they're doing is merging AP/NoA into FSD as it stands then FSD will be a hard pass for me.
I don't really get this. I've been on the normal firmware train since Xmas, and now in the June rollout of FSDb 10.12.2, and for me NoA was already very useable and useful 6 months ago, and I watched it progressively get significant better for me in the ~6 months since (and then no real change with my switch to FSDb firmware, other than onramp/offramp getting better with the stack transition. Without FSDb, it was confusing and inconsistent about what state it left me in after an offramp). When I say it got "better", an example is its handling of curved interchange roads from one metro highway to another. In January it often didn't slow enough in time for my comfort level and I'd disengage just before the turn-in point. It now handles those much better and progressively steps down its speed limiter as it approaches the curve.
Usually there's only two common reasons I ever disengage NoA on Houston-area highways (or roadtrips):
1. There's a couple of toll booths on Beltway 8 I've found where NoA just kinda freaks out. I think the booths must have been under reconstruction or changed layout since its last map update of them, or something. It more or less refuses to go through them on NoA and freaks out with a red takeover warning, etc. Many other similar toll lanes around the area work fine, though.
2. When the 85mph limit is an issue. There are some stretches (e.g. the northeast section of 99, from about I-45 to I-59) where it's just 2 lanes and it's common at some times of day for the fast lane to run mostly at ~95mph* in a 70 or 75. The car's 85 limit will causes issues with this no matter what (because it will try to pass slow-lane cars but not get up to the real speed of the fast lane either, so you're always being rude or stuck). In those cases I just go ahead and disengage so I can flow with the fast lane.
Without disengaging, I do "manage" its execution, though. I usually have to manually pick lane changes and/or canel lane changes it chooses to make a few times each drive. I sometimes give it a little gas when it's being under-confident speeding up for a pass, too. Still, I find it highly useful and it's almost always engaged on the highway.
* - Why almost exactly 95? Because we have lots of trucks and SUVs in our traffic, and it's common for them to be factory speed limited at 95mph, so this kinda becomes the common limiter for traffic as a whole unless the road is relatively-empty.
I also have not noticed any difference with regular AP/NoA as part of FSDb 10.12. The only gripe I have is aggressive accel/decel sometimes in bumper-to-bumper traffic, but that has been an issue for a long while and not specific to FSDb. I can't even remember the last time I had a phantom braking incident on the highway. I definitely empathize with people with degraded experiences though. There's too many anecdotes of that for it to be subjective.
How many times in the past has Elon done what he has now done with v11?
I can think of some striking examples off the top of my head, pretty sure v10 even. Elon was going on and on about how v10 would revolutionize the system, then they released 3-4 smaller patches that at best brought minor incremental improvements and some steps back, people asked when v10 was happening, and Elon said v10 was basically already here through the smaller patches.
From what I understand of the technology, labelling is there only to build out the OEDR. It’s literally just labelling objects so they’re recognized by the system, and that in itself feels like just one slice of the self-driving pie. So even if auto-labelling were happening, that doesn’t mean self-driving is solved by any stretch of the imagination. People are acting like object labelling is THE key to full autonomy…
In reality, I think recognizing objects is just the most basic foundation upon which all the more complex stuff must be built. FSD already does quite well in open empty roads, aside from objects it doesn’t recognize, but it struggles most when interacting with other traffic and I think that’s entirely unrelated to labelling and the OEDR.
I can think of some striking examples off the top of my head, pretty sure v10 even. Elon was going on and on about how v10 would revolutionize the system, then they released 3-4 smaller patches that at best brought minor incremental improvements and some steps back, people asked when v10 was happening, and Elon said v10 was basically already here through the smaller patches.
From what I understand of the technology, labelling is there only to build out the OEDR. It’s literally just labelling objects so they’re recognized by the system, and that in itself feels like just one slice of the self-driving pie. So even if auto-labelling were happening, that doesn’t mean self-driving is solved by any stretch of the imagination. People are acting like object labelling is THE key to full autonomy…
In reality, I think recognizing objects is just the most basic foundation upon which all the more complex stuff must be built. FSD already does quite well in open empty roads, aside from objects it doesn’t recognize, but it struggles most when interacting with other traffic and I think that’s entirely unrelated to labelling and the OEDR.
Yeah, for sure there's a long way to go in the big picture. 10.12 at least seems to have an extremely short-term view of the world in general. It doesn't think/plan very far ahead at all in general, even based on map data. It also doesn't seem to really maintain perception of cars or people that become temporarily occluded from view. It's like it's only reacting to what it perceives in the moment, rather than building a persistent mental model of all the actors around it and their projected future locations, or even its own future intentions. FSD never wonders "where did that car I was tracking go? I bet it's about to re-appear over there in a bit, in that lane I'm wanting to change into within the next half mile". It seems to inhabit a reality that's only based on instantaneous information. Things magically vanish and appear and this causes it no real concern. Plans last mere seconds, not minutes.
My guess is they did find value in the human labelling, but as ML models have become more powerful and the human labelers created a large databased of labelled images, the need for human labelers has dramatically decreased.
But I do agree that Tesla may revisit their strategy to something simpler. Perhaps they will look at something from outside vendors.
smogne41
Active Member
The removal of radars last year should have pointed to vision-only being in a working state as well...
It's hard to blame EM for setting and missing the deadlines, but it's for sure easy to blame him for hostile DEV environment. The 40 hours in the office BS is lack of forward thinking from him. It's obvious that some people would do 200% better at home, if they can focus and not worry about time of the day, dressing up, public washrooms, etc, etc. Other would do better in the office because at home they have kids, wives, pets, loud neighbors, etc, etc. Treating your talented workforce, especially DEVs, as slaves that you happen to feed and give money to does not make sense. They will get poached.
Andrej should've been above those rules, imho. And if Tesla is on the final stretch and all they need is more fleet, DOJO and possibly V4 hardware then Andrej is simply wasting his time there and should focus on new horizons. Possibly join Hotz, those two together can change more things than EM ever did, imho.
Andrej should've been above those rules, imho. And if Tesla is on the final stretch and all they need is more fleet, DOJO and possibly V4 hardware then Andrej is simply wasting his time there and should focus on new horizons. Possibly join Hotz, those two together can change more things than EM ever did, imho.
Similar threads
