diplomat33
Average guy who loves autonomous vehicles
Not yet. I am still at 98. But the safety simulator says I only need 90 more miles to get to 99. So I think I will take a little drive today. That should do it.
-
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
Not yet. I am still at 98. But the safety simulator says I only need 90 more miles to get to 99. So I think I will take a little drive today. That should do it.
EVNow
Well-Known Member
diplomat33
Average guy who loves autonomous vehicles
EVNow
Well-Known Member
diplomat33
Average guy who loves autonomous vehicles
Thanks. But at this point, I only need about 90 miles. If I reset I would need 100 miles and that's assuming I get a score of 100. If I get a bad drive, I could be stuck doing more miles to get my score back up. So it probably does not make sense to reset now.
What causes regression in something like FSD software, and why is it they can fix the regression in a day? Intuitively, at least from someone not in the tech field, I would think that if they can fix/improve regression in a day, then that means they are making minor tweaks of some sort to the software; and, if it's just a matter of tweaks, then why aren't they further ahead in smoothing out issues relating to left turns?
I'm not in IT (clearly from the nativity of my questions) but hopefully someone can explain this to me.
AlanSubie4Life
Efficiency Obsessed Member
I was going to comment on this regression testing too. I’m kind of surprised they can fix a specific regression without introducing them elsewhere. Maybe it was in the procedural code (makes more sense there)?
All seems extremely tricky anyway.
All seems extremely tricky anyway.
EVNow
Well-Known Member
I’ve a longer post 2 weeks back you can look up.
This time one possibility is that some change they made to left-turn, actually made it worse. They can revert the change and release in a day.
EVNow
Well-Known Member
I guess that’s where the experience comes in. While there is always a chance of further regression, a concentrated effort at one issue can help solve it quickly. Lots of time it’s just a simple mistake that got overlooked.
Ps : by now they would definitely know how to fix NN regression issues too, though will likely take more time.
Neural networks output probabilities that traditional code can have thresholds to allow/disallow. The network training process that includes new intersection training data will increase/decrease these probabilities for specific intersections, so if a known QA intersection needs to pass, theoretically the threshold could be adjusted to pass (and hope it doesn't cause additional regressions for other intersections not tested). Alternatively, the specific portion of the neural network could be retrained (either add or remove training data) such that the network outputs exceed the desired threshold again, so this avoids a full network retraining cycle.
Honestly, the guy runs Tesla as well as SpaceX, and has God knows how many other side projects… i’m shocked that he finds the time to respond to tweets at all. I honestly have no idea how he does it.
Remember he fired is PR team, so he has to do all the talking now...his problem...shouldn't be ours.
I expect they will declare FSD feature complete sometime soon--that is, it does all the things originally promised, of which AP on city streets is really the only outstanding feature (note: I am not making any judgement on how well they work). I think Tesla will then upgrade older cars to HW3 standards because the current SW stack is designed to work with HW3.This!
It's guaranteed this scenario will repeat with *all* Tesla vehicles with HW4 (or HW 3.5, or HW 4.5, whatever they call it). It's one thing to service 10k's of cars - quite another to service 100k's of cars (or whatever the # is - you get the point). It's going to happen again and again. I plan on keeping my '17 MS for as long as possible with free SCing - taking the risk of course nothing major breaks, or I don't go home first. Battery still has 300+ mile range. In this day and age where computers/phones/electronics are obsolete every 18 months, our cars are not excluded.
I'd expect the HW4 upgrade, if offered for free to FSD buyers or subscription buyers, will take well over a year. Lord willing, I'll get camera upgrades in the next month or so. Frustrating of course, but mostly on the lack of comm's and mis-truths. Technically speaking, it's understandable. Would have been nice to know they never intended this configuration to work, let alone playing games with your customers - but technology is what it is. And personally, I am glad Tesla is not backing off the "all hardware you need for FSD" statement.
Heck, by that time, they may decide to put a much better radar back on all cars. Even if the new cars don't have the wiring, they certainly have the input jacks. Wouldn't surprise in the slightest to see this come down the road in a couple of years, maybe post-Elon. Maybe they can do HW4 and the radar at the same time??
Based on Elon's comments at the Aug AI day, I think we will see FSD 2.0 and HW4 released, probably with the CT. He referenced newer cameras and, IIRC, HW4 during the event. Now Tesla cannot afford to keep perpetually upgrading every car they have ever sold since 2016 every time they upgrade FSD capabilities, so I think they will declare the current SW "done" and their obligations fulfilled, then introduce something new.
Wonder how many 100 people would still have their 100 score if they had to play the game as long as us 99 peasants lol.
Gwynne Shotwell really runs SpaceX, which frees Elon to be Elon.
rxlawdude
Active Member
I think there will be a problem with pre 2019 buyers. Expect refunds or credit towards FSD 2.0 on next Tesla purchase. They likely can't deliver the FEATURES advertised 2016-19 without serious upgrades to hardware.
rxlawdude
Active Member
Similar threads
