fasteddie7
Active Member
I think every stop sign in uniontown has to be leaning or placed next to a brick building or huge tree to obstruct it, up to code and within spec. Lol
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD discussion
- Thread starter Buckminster
- Start date
jabloomf1230
Minister of Silly Walks
Have you ever thought about bending the pole back so the sign is not blocked? It seems like the neighbors don't like this stop sign.
Knightshade
Well-Known Member
"what is the logic it is using to come up with 69?"
Have you MET Elon?
Also will be curious if NHTSA gets mad about a system that is self-selecting (rather than driver selected) a speed above the posted limit.
insaneoctane
Well-Known Member
Seems strange that Tesla FSD team hasn't bothered with reading signs yet. This seems like one of the easier tasks honestly. One could infer what this says about where the team is.
Are folks here choosing to ignore the possibility that once "solved", FSD might need another, beyond HW3.0, HW upgrade on existing vehicles to deliver? Is the assumption that the huge money printing press that arrives with solving FSD grossly overshadows such concerns?
Speed signs are often read (but not always because it's not uncommon for the camera to miss the first speed sign when leaving town so you end up with a 35 mph speed on a 65 mph highway). I don't think the HW3 cameras have enough resolution to effectively read the small print (e.g. school zone hours). Also they will have to determine of which signs actually apply to the street you are on. In some situations, it's not that easy to determine the speed sign vs street. Humans realize that if you are on a 75 mph highway, the 50 mph sign that's a bit off the shoulder is not intended for the highway but for the parallel side road.
Knightshade
Well-Known Member
I don't think there's much doubt you'll need more than HW3 to deliver and L4 or L5 system from anyone who isn't Elon or his biggest fans at this point.... But Elon already said upgrading HW3 cars further was cost prohibitive.
So most likely the ballpark 40-50k people who bought FSD when it was promises as more than L2 will get some kind of more-generous-than-what's-offered-now trade in deal, or some sort of refund+interest type settlement otherwise once Tesla finally admits they'll never deliver what they promised.
And all the post March-2019 buyers will just get a copy of the product description they bought making it clear they were never promised more than L2 on city streets anyway and that's all they're getting.
Knightshade
Well-Known Member
Elon Musk said:
Bold added.
Seems relevant enough to those figuring major release timelines into their financial projections for stuff like deferred revenue and take rates, based on the timelines the last several times Elon said they were having to do a re-write of FSD.
Also seems relevant to again encourage those who don't appear to have availed themselves of the much broader and realistic understanding of FSD tech, development, and progress where you'd have already known this would absolutely be a rewrite the plethora of knowledge to be found here:
AI, Autopilot, & Autonomous/FSD
Discussion about AI, Tesla Bot, Tesla Autopilot (AP), the promise of Full Self Driving (FSD), as well as other Autonomous Vehicles.
teslamotorsclub.com
As much as I subscribe to the University of All Things Musk,
I can't listen to any FSD promises without rethinking my subscription to the University.
I may pay for FSD at some point, not because it's good - but because Autopilot is that bad.
It's Friday - buy - day - but damn FSD statements.....
I can't listen to any FSD promises without rethinking my subscription to the University.
I may pay for FSD at some point, not because it's good - but because Autopilot is that bad.
It's Friday - buy - day - but damn FSD statements.....
So...2 weeks then?
Bold added.
Seems relevant enough to those figuring major release timelines into their financial projections for stuff like deferred revenue and take rates, based on the timelines the last several times Elon said they were having to do a re-write of FSD.
Also seems relevant to again encourage those who don't appear to have availed themselves of the much broader and realistic understanding of FSD tech, development, and progress where you'd have already known this would absolutely be a rewrite the plethora of knowledge to be found here:
AI, Autopilot, & Autonomous/FSD
Discussion about AI, Tesla Bot, Tesla Autopilot (AP), the promise of Full Self Driving (FSD), as well as other Autonomous Vehicles.
Just to clarify, which things should be considered rewrites? Radar -> Vision? Ultrasonics -> Vision (still in progress)? Highways? Occupancy?
In this particular case of vehicle planning/control code that has accumulated over the last 3 years of FSD Beta, almost all of it being replaced by neural networks should be similar to which type of rewrite, or it might be more extensive than any previous rewrite?
Knightshade
Well-Known Member
I think the most comparable is probably when they went from single cameras to surround video-- they're fundamentally changing a foundational part of the entire suite, not just, say, replacing speed/distance estimates from radar inputs with the same inputs from vision.
Okay thanks. For reference, Elon Musk started talking about surround video September 2020 even before initial pre-Safety-Score FSD Beta "v1" release:
I believe FSD Beta 10 from September 2021 initially released surround video then led to 10.2 going to public Beta with Safety Score in October, so about a year from first mention to wide release:
Although the beta audience was much smaller back then and maybe could be more acceptable to have regressions when moving fast. Then again, Tesla also needed to build up its training capability including revamping the data collection and labeling pipeline, and all of that infrastructure is ready to use now for the new neural network and rapidly expanding with Dojo.
I suppose looking very broadly at each year:
- 2020: FSD Beta can drive on city streets
- 2021: Perception rewritten with surround video
- 2022: Production-quality NoA replaced by FSD Beta
- 2023: Planning/Control replaced with neural networks
Was Tesla expanding FSD to other countries what finally pushed Autopilot team to replace Planning/Control with neural networks? There's been a long tail of capabilities not handled yet such as school zones and other signage, but maybe the reality of expanding from 300k lines of control code for "just" North America highlighted the need for a more general solution? Recent Tesla vehicle software updates do have Data Sharing policy prompts in more regions to Allow Autopilot Analytics & Improvements that could be used to train worldwide FSD.
This scheduling/utilization visualization from AI Day 2022 already shows existing neural networks like Moving objects and Occupancy initially processing presumably 360º camera pixel data on TRIPs 0/1 before passing distilled intermediate features to GPU then CPUs probably for control logic. Replacing 300k lines of C++ could similarly push processing of what used to be on CPUs to GPU, so that doesn't necessarily increase the Compute load on TRIPs and actually makes better usage of the GPU capacity.
This scheduling/utilization visualization from AI Day 2022 already shows existing neural networks like Moving objects and Occupancy initially processing presumably 360º camera pixel data on TRIPs 0/1 before passing distilled intermediate features to GPU then CPUs probably for control logic. Replacing 300k lines of C++ could similarly push processing of what used to be on CPUs to GPU, so that doesn't necessarily increase the Compute load on TRIPs and actually makes better usage of the GPU capacity.
Big night for FSD. This drive will almost certainly go viral given the fight. General public do not realise how good FSD is. Will Elon be using V12 alpha?
Last edited:
Knightshade
Well-Known Member
willow_hiller
Well-Known Member
I think they only ever demonstrated auto-labeling in terms of taking one manually-labeled frame, and projecting it forward into several seconds of labeled video. But that process still requires the first frame, and presumably any frames where the scene substantially changes, to be manually labeled.
Similar threads
