-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Daniel in SD
(supervised)
It was the FedEx truck, they probably don't have ABS (and if they do it's probably not as advanced as modern 4 channel systems).
We could also mention that the car puts itself into that situation by erroneously changing into a closed lane, long after a human driver would
have recognized it was closed and never attempted changing into it
In a half hour video where the car is going 15-20mph for most of it and with frequent stops, how many miles would that have racked up? 10 miles maybe?
I don’t find the drive particularly impressive. Navigating narrow roads and such is neat, but the video is mostly very low speed (I think Whole Mars Blog is manually adjusting the speed down in spots too) with mostly simple scenarios with low traffic. The few instances with actual traffic are where the car exhibits some strange behaviours.
FloridaJohn
Member
This is the biggest problem with comparisons. Without the data from Tesla, the optimistic people see Tesla as having nearly completely solved the problem (two weeks away!), where the pessimistic people see a problem every mile.
Tesla needs to start publishing actual data so a real comparison can be made with other autonomous companies.
powertoold
Active Member
Tesla has large NN models that can extract more information from cameras, but they run those models on their clusters rather than in the car.
It's possible that Tesla won't be able to achieve L4/L5 with the current hw3, and it might be forever stuck in L2 (because it'll only be 2-3x better than human safety). It's going to be difficult to get to 10x human safety with hw3, because of how large the NN models need to be to consider all the edge cases.
EVNow
Well-Known Member
Is that speculation or informed guess ?
I don't know about L5 - but I'm fairly impressed with various videos. Unless they are editing out a lot of mistakes - it looks pretty good to me and much better than the videos I watched when the fsd beta first came out. We are definitely getting close to "button" time.
Last edited:
Knightshade
Well-Known Member
It's a very informed guess that Tesla will need HW4 (or more) to offer anything above L2 because the current system, which obviously still needs improvement, is already using so much compute that the original intended dual-node redundancy is out the window.
...Until Tesla folds in the highway stack (which Elon says is currently behind the production highway code), supposedly in 10.1, and then introduces a whole new set of problems that block the release of the "button" for another 11 months.
For a very long time there will be a shiny new
I mean, what are they actually waiting for? What metric do they use? It's L2, it's always the driver's fault, right? This code makes mistakes. V9 made mistakes. What is the threshold for giving it to more people? The current beta testers are already driving around without their hands on the wheel, letting the car go very far into dangerous situations. Will the average public be worse? Will the average public be better if the code gets better, giving them *more* confidence in a system which will fail 1% of the time?
I heard V9 was really close to being ready from supporters. Now I hear v10 is close also. What needs to be fixed?
How do you know they will not need more cameras or sensors along with HW4 to achieve L4?
Daniel in SD
(supervised)
This is what I'm dying to know. The only thing that makes sense to me is that they're not sure if it's safe so they're starting with a small group of people (like the phases of a drug trial).
Yes, but you can't use this process when you change the product constantly. The FDA doesn't allow you to completely reformulate your drug and then say all the data from V9 applies to V10, and then the V10 data gives you rationale to release the reformulated V10.1.
You only do that when what you have is stable. "FSD" is not going to be stable for years. The only thing they can really find out is if humans are a good backup to very unstable systems. I keep hearing the FSD beta has no accidents, so apparently it is? But we also have all that data from the L2 Highway code- and Tesla claims this is safer than a human alone. How can an L2 system be unsafe?
I mean, the next release is supposed to completely change the highway stack- the one that has been stable for years, and that we have actual statistics on. It's getting less stable, not more stable. This is the exact moment in time where you are furthest away from knowing how safe it will be in public, which you would think means we're not about to go wider if they are safety first, yet Elon is saying we're still close.
So again, why are they waiting? Could it be a non-safety issue, like they know people won't be impressed at what they get after 3 years of waiting for $10K? That the PR is actually better letting 10 people make videos and everyone just see those, rather than having people experience it themselves? Or that the system only works in narrow geofenced areas, and the NDA doesn't allow people to discuss that? I mean, not releasing it makes all these questions valid.
This is not a company that has taken stability and safety with AP seriously in the past. Tesla has $1B+ of customer money for this feature. It's L2 and safety is not an issue, we'll blame the driver in all cases. What's the holdup?
EVNow
Well-Known Member
You are making a lot of assumptions. Optimization takes time.
Let me put it another way. Only people who are industry experts (i.e. autonomous vehicle industry) probably know what it takes to get to generic L3/L4 level. Even those may not understand what it takes to get to that level with vision only (most experts are on the lidar side). So, unless you are one of those in the industry or an academic who is an expert on A.V. ... its all speculation at this point.
There is another twist - SMEs in the industry are unfortunately biased. They either work for a competitor of Tesla or work for Tesla.
ps : No - a certain "senior software engineer" who took a course on CNN and has 5 years of IT experience is not an SME of A.V.
powertoold
Active Member
During AI Day, I remember Elon saying he's confident hw3 will get to 2-3x human safety. To me, that's not enough to enable L4+.
EVNow
Well-Known Member
Actually 2x to 3x human safety level (i.e. a reportable accident every 20 to 30 years) is good enough for me. Infact, we are all doing just fine and are completely happy to be driving around with ~ 1x human safety level.
BTW, I don't remember the quote. Need to see exact working and context.
powertoold
Active Member
Daniel in SD
(supervised)
Knightshade
Well-Known Member
We know for a fact the regular PRODUCTION code doesn't have enough compute to enable full redundancy. Let alone all the extra city streets stuff needed just to get to where the FSDBeta videos are today.
Sure, you can SPECULATE they'll magically be able to add a bunch MORE capabilities and somehow also magically reduce compute usage a TON with 'optimization'- enough to squeeze ALL of that back into the single node it overflowed out of before the first FSDBeta came out.
But it seems less likely of an outcome than "needing at least HW4" does.
Honestly a bigger question to me right now is do they run out of compute on BOTH nodes of HW3 before they solve this or not?
Because if they do they're rightly screwed for much further progress until HW4 is going into cars- which sounded like late 2022 last I heard.
Whereas if HW3 has enough compute across both to solve it they can keep moving forward to that, then just give anyone wanting >L2 an upgrade to get redundancy back.
If that were true we wouldn't be mandating things like ABS, AEB, etc.
Which part of what I wrote, in any way, suggested any claim to my knowing that? I literally didn't even mention sensors or cameras in the post- since its actual topic was on-board compute, not sensors.
Elon has been confident about a lot of things that are not true. He was confident HW2 would be good enough too. FSD was "definitely" only 6 months away in 2017. He's surprised "the button" is later than June. He talked about radar as important until he didn't. Why would we take his "confidence" as any worthwhile data point?
When Elon quotes safety targets, he usually quotes 1:1B miles (via "nines"), which is >10X as safe as a human overall, although only about 3X as safe on the highway.
And the irony of those- we're requiring features that take over when the human fails, because the real data is that accidents happen when humans get distracted, overwhelmed, or run out of skill, and these systems help in those cases.
Meanwhile, Tesla seems to be trying to improve safety by developing L2 systems which only function when humans are paying attention and make no attempt to perform backup functions to human failures. The only safety argument I ever hear for Tesla's current AP features is it makes you more relaxed/less tired? Which completely ignores the tradeoff that more relaxed people tend to be less alert.
Similar threads
