-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Elon: FSD Beta tweets
- Thread starter Terminator857
- Start date
Better question - “one more refinement” till what target is achieved exactly? We all know FSDb will forever need one more refinement
Presumably because with this machine learning methodology, the refinements themselves will come from analysis of data from a very wide sample a previous-rev users.
We always need to keep this in mind. Even with more conventional technologies that most of us have worked on, you learn from field experience with prior product generations. I couldn't count the number of times we went to market already knowing of issues with the present design - which is fine and often quite sensible - but gaining further valuable insight from customer experience. Then later presenting the plans for the follow-on generation, and of course fielding questions from annoyed higher ups " why didn't you think of all that the first time?"
Back to the machine learning paradigm, it's even more fundamental that you need tons of data that you can grade for many details of success and failure, in order to produce the next round. Even when the architecture isn't changing fundamentally, as it does for major releases.
I suspect Elon suffers actually less from that development-ignorance disease, else he wouldn't stay the course with his goals even though the original timelines are shot and the whole world is telling him he's wrong, a charlatan etc.. I realize it isn't fashionable to give him any credi while everyone is impatient, but I found that last tweeted comment to be completely sensible even amidst the current rollout (which I hope to get very soon).
(@Ramphex : No I didn't post this shillification to move up the queue, but neither would I refuse it.
I have my principles but that's not one of them.)I’ll have to link the link again. But I watched a new report about autonomous driving that covered the last so many year. It detailed the investments from Google, Fore, VW and discussed MobileEye and Tesla along. With others. Topic such as MASSIVE cash investment and profitability (NOT) were discussed along with long term marketability and this profitability. Many seem to be dialing back investment, sell off purchases at a great loss and collaborations to reduce overhead.
It was quite interesting. I’m sure some will scream fake news; as is the provide it to be or I don’t believe it - but there is no way to provide it to these people - as is par for course these days.
EDIT: Link found & provided.
It was quite interesting. I’m sure some will scream fake news; as is the provide it to be or I don’t believe it - but there is no way to provide it to these people - as is par for course these days.
EDIT: Link found & provided.
Considering that 11.3.3 had arguably the worst performance we have seen on chuck’s UPL in over 6 months I have to agree with Elon on this one. It also had zero improvement over 69 on his memorial park drive and was also worse on his unmarked roads test (which the car will never solve until they actually try and tackle it like they did with the UPL).
AlanSubie4Life
Efficiency Obsessed Member
We’ve seen FAR worse than that; it was 4/6 here as I counted it (quickly). It was only disturbing because it kept stopping in the middle of the road. Also probably worse per the actual scoring criteria. But baseline is 4/6.
But Chuck’s last attempt was 5/8 (which is worse). And per the “official” criteria it was 1/8. It also stopped in the middle of the road. So nothing new to see here.
The ULT/UPL has never worked. They saw failure, called it success per social media, and moved on.
Now everyone drives around not attempting unprotected left turns, except the simplest possible types (because they don’t usually work!). Success!
This is what Elon's "one more round of refinement" was referring to, not needing to wait for 11.3.4
Honestly I’m kind of confused by these recent tweets and the rollout.
It sounds like .4 is what they’re focusing on; or whatever they call it. Typically further rollout would be stopped when he makes these kinds of tweaks.
But .3 has gained significant traction alight before is post and especially after.
This seems against the grain from every other release.
It’s like they’re “okay” with .3 and will confide to push it out but are not happy with it and .4 is also going to drop soon(ish?h as well.
This one is a VERY confusing rollout.
It sounds like .4 is what they’re focusing on; or whatever they call it. Typically further rollout would be stopped when he makes these kinds of tweaks.
But .3 has gained significant traction alight before is post and especially after.
This seems against the grain from every other release.
It’s like they’re “okay” with .3 and will confide to push it out but are not happy with it and .4 is also going to drop soon(ish?h as well.
This one is a VERY confusing rollout.
Don’t think too hard about whatever Elon tweets. This “refinement” could be 11.3.4, 11.4, 12.0, or anything in between.
I don’t. Take them with less than a grain of salt: especially these days. As it was the combination of the tweaks in contrast to what’s actually happening. Drastically different than usual. Besides I said .4 or whatever it ends up being. Could be v12.
Life is one day and time.
FAS (Beta) is one year at a time.
;-)
sleepydoc
Well-Known Member
something I've noticed fairly consistently with past rollouts is that the performance improves somewhat over the first several weeks. I'm not sure if it's because the car gets new mapping data or the neural nets are updated but I've definitely seen some modest improvements.
I'm not at all surprised by this. V11 was a major re-write that merged the highway stack and also adjusted some of the programming paradigms. it's quite typical to see some regression in these cases.
Like I said above, it's a significant rewrite so it's not unexpected that they need a fair amount of testing and revision. A typical rollout needs a couple of revisions. Looks like they may be doing one more on this one.
Knightshade
Well-Known Member
NNs can't be updated without a firmware update. Maps can be though.
I suspect .3 is going wide-ish because it's the first version that both solves the FSD NHTSA recall AND is "good enough" to go wide.
I got it late last night FWIW, haven't gotten to drive on it yet.
Main thing I'm curious about is if the highway stack now being on the fsdb code means it drops back to basic AP less often in moderate rain or not... (which it did just yesterday in moderate rain, dropping from NOA to basic AP and forcing me to do my own exiting of the highway like some kind of peasant)
On local roads previously fsdb seemed better about still mostly working in moderate rain compared to how legacy NoA handles it (though heavy rain still killed it like it does noa)
DanCar
Active Member
I perceive the same. Possible reasons:
1. At first I have high expectations, then they get dropped to zero, then the car performs better than my crappy expectations. :0)
2. After a few weeks I have better understanding where it succeeds and where it fails?
3. I know how to maneuver the vehicle so it doesn't get into trouble situations or it is just not active in such situations.
Last edited:
johnmodely
Member
I read that tweet as not referring to the next point release, but as typical musk fashion of needing one more major release where all NN are on video and FSD beta is "complete" in his mind
My theory about how a specific build gets better over time is Tesla has runtime parameters that they can adjust OTA without a full update. This might control things like how assertive the car is around VRUs or minimum separations to other vehicles or other other things that make the car hesitant.
Basically Tesla is being hyper careful to minimize any risk of the car running into something (at the cost of it being too hesitant or slow). As they gain confidence in the changes they bring the assertiveness back up to where it was before.
Basically Tesla is being hyper careful to minimize any risk of the car running into something (at the cost of it being too hesitant or slow). As they gain confidence in the changes they bring the assertiveness back up to where it was before.
sleepydoc
Well-Known Member
I’ve considered options 1 and 3 but there are specific cases where it seems to improve (e.g. a turn lane that it always got wrong but starts getting right).
Knightshade
Well-Known Member
None of that stuff would survive a reboot. This has been discussed a number of times, including specific details from @verygreen about it
sleepydoc
Well-Known Member
I haven’t read the details from verygreen if the software can survive a reboot then the parameters can as well. It all depends how the system is set up.
Similar threads
