-
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
For those not used to FSD (and sometimes for those who are), the way FSD drives can be a bit uncomfortable at times even if it really isn't going to do anything bad. I just installed the latest version a couple of days ago (not good enough WiFi over the weekend to download) and it's vastly improved. Because of it's desire to keep exactly in the middle (as Elon said yesterday), it comes uncomfortably close to vehicles in the next lane, and on streets with parked cars it weaves in and out. My guess is that Elon drives almost all highway and few surface streets.
Think it's a combination of many things. The car drives unhumanlike which some people feel is unsafe but it might actually be safer. Elon has a lots of history of trying the technology and understands it very well, making him trust the system more. Some areas with a high FSD density tend to work better, like SF, Austin where Elon drives a lot. Elon probably has the latest beta version. Elon probably is willing to stomach the cost of any minor accident and feels that if anyone should die in an accident it should be him, making him willing to take more risks.
bkp_duke
Well-Known Member
In my experience, FWIW, FSD behaves considerably different on different settings. I have a lot more interventions if I tell it to be aggressive and lane change vs. the default settings.
I keep it on Chill and it pretty much behaves similar to Gilles Villeneuve.
EVNow
Well-Known Member
They should be required to disclose these numbers as well.
BTW, how does Waymo compare .... my guess would be they are btter.
EVNow
Well-Known Member
I've been saying this on the AP forum for years.
Waymo/Cruise etc have scaling up (and commercialization) challenges.
Tesla has actually getting FSD to work challenge.
I think you guys are saying the same thing. Tesla's solution is general - once it works in one market, it works for all markets. Competitors' solutions are local - they are expanded to one market at a time. It's a matter of which one gets to large scale adoption first. I think Tesla is going to get there first. Even though their solution is binary - whether it works for all or it doesn't work at all, the immediate scalability is an attractive value proposition.
Vision-only solution paired with rapid advancement in artificial intelligence along with a robust data collection infrastructure will get Tesla there. My conjecture for Tesla nearing complete FSD by the end of the year got derided in the roundtable thread, but my hunch tells me that it's coming in the next year or two.
bkp_duke
Well-Known Member
Had a safety disengagement this AM. Wife got a nail in a tire on the 2020 Y P. Coming back from the tire shop, wide road in an industrial area. FSD didn't want to split the large lane, I think it assumed it was 2 lanes, and it made a MAD DASH for the right. I'm not sure it would have hit a parked car that was there, but I yanked the steering wheel and sent a report. First safety disengagement I can recall in a long time. Most are "you aren't driving fast enough for me" disengagements.
EVNow
Well-Known Member
Reposting this from a while back. Waymo & Tesla are difficult to compare because they are better in different dimensions. And as any student of science / math can tell you - vectors can't be compared, as if they are scalars.
We actually have 3 dimensions
- Scenario (or feature)
- Location
- Individual (i.e. people)
Reliability is not a dimension - its a "measure". For a given combination of the above 3 dimensions we get a particular reliability (or error probability) number. Ofcourse the reliability can be calculated for a group of individuals/scenarios/locations or any combinations of them.
Waymo has high reliability for a lot of features - but only in small # of locations and only for their own cars.
ps : Elevators work in a lot of locations but in extremely limited number of scenarios 
We actually have 3 dimensions
- Scenario (or feature)
- Location
- Individual (i.e. people)
Reliability is not a dimension - its a "measure". For a given combination of the above 3 dimensions we get a particular reliability (or error probability) number. Ofcourse the reliability can be calculated for a group of individuals/scenarios/locations or any combinations of them.
Waymo has high reliability for a lot of features - but only in small # of locations and only for their own cars.
ps : Elevators work in a lot of locations but in extremely limited number of scenarios
EVNow
Well-Known Member
adiggs
Well-Known Member
Completely agree with the idea that Tesla's solution is general, and that's the right approach. So strongly in fact that I think all the other efforts that aren't going down this path are wasting time and money. They will eventually hit a wall that they can't climb - the only way forward will be effectively starting over, with the possibility of using whatever they've developed under very stringent requirements. Like daylight, no rain, <30 mph, specific set of streets with a set list of starting and ending points - stuff like that.
FSD complete though is probably a matter of definition. If its defined as "FSD works and is no longer beta", then I can't disagree more. Partly because although I see it improving rapidly right now, with every release, and believe it will continue to do so.. autopilot has been out for like 6 years or something (maybe its just 6 years since I first got it and started using it), and that is still officially beta.
I hadn't really put that together until this last week, and I think I've realized why that is the case, and why its going to be a big problem until somebody tackles this from the legislative side.
To get deployable non-beta FSD, much less autonomous vehicles, we need a legal framework in which FSD and equivalent solutions aren't something that can be sued just on their own. You need a lot more - malice, indifferent regard to how well it works - a pretty high bar to clear.
I get here mostly by analogy. There is a pretty easy argument that the internet exists, as we know it today, because of Section 230. Whether you agree with it or not (I agree), simplistically you can't sue TMC because I post something for which I can be sued; you're stuck with me and my pockets to go after. Without that protection the big entities we know today wouldn't exist. There might not be anywhere people can post content without an editor in the middle to vet content - more of a book publishing business model than social media.
Or proving defamation against a new organization - the bar is intentionally set high so that news can be reported on without worry about a blizzard of law suits. Even frivolous law suits cost a lot of money to defend, and the risk of those law suits will chill speech.
I've realized in the last few days that we're going to need something similar for FSD.
We have something similar for cars already. I don't know if its by law or not, but if I'm driving along and get in an accident while driving my Ford F150, the idea that me or my heirs can sue Ford for my dying while driving a truck they built is mostly laughable. In order for that lawsuit to go anywhere I'll need to prove there is something about how Ford designed that truck, or how it was built, or something along those lines. It's a high bar to clear, with the default assumption being that the car crash isn't the fault of the manufacturer.
Maybe this will mean there is a need for an "FSD Certification" type process. Have your vehicle(s) pass a test administered by some organization and you are certified as having safe software that can operate with this legal protection - maybe that would work, though I suspect not. I don't know how we get there, but we might be living with FSD Beta for a long time after most of us consider it beta.
I know autopilot became a tool for me that makes my life better, and that I rely on for significant quantity of driving, years ago. FSD Beta isn't there yet for me, but I can see signs of life that it will get there. Maybe even this year, but that won't be enough for Tesla to remove the Beta tag (that's my belief, and what I invest on).
My sense of AI advancement and the initial success of using AI-based decision making. I don’t think 100% unsupervised will be achievable in that timeframe because there will be unique situations that need human intervention, such as road closures without a clear detour and navigation map being outdated, that will need to be addressed one by one. Unfortunately reinforcement learning doesn’t work for AV because the reward isn’t so clear, and RL is the only type of AI decision making I’m familiar with. I don’t have a good understanding of what methodology is being used by FSD, but it seems to be working. If the architecture has already been figured out, training the algorithms on data should not take very long. It took Open AI 4 years to go from the first GPT paper to the ChatGPT version that took the world by storm. I believe Tesla started AI-based decision making in 2021, and 4 years from that is 2025.
EVNow
Well-Known Member
I think this kind of "extrapolation" is the reason Elon got FSD so wrong over the years.
First, you need to "prove" that the extrapolation is correct in this situation or that FSD is similar enough to Chat GPT. Neither, IMO, is correct.
Industry experience is that it takes a very long time to handle the long tail. It was for Elon to prove the industry wrong ... IMO, it has been the other way round.
This is why this is just my hunch. If it were scientific, it'd be an actual argument, akin to what I did with Tesla starting to advertise
EVNow
Well-Known Member
IMO - what I described would not be "scientific". Anyway that's how I form my "hunches" ...This is why this is just my hunch. If it were scientific, it'd be an actual argument, akin to what I did with Tesla starting to advertise
I think that's a pretty big leap.
I don't believe that current fleet will have the capability for Robotaxi...most who have FSD Beta do not, hence why there has been talks from Elon about designing a specific robotaxi vehicle. Also, they are selling everything they get in and making a profit, that's more likely the reason they aren't changing the policy.
If I understood the situation, it’s not that the “current fleet” is not capable of robotaxi, it’s that Tesla doesn’t want to retrofit customers to HW4. Now, when it comes to their OWN fleet, they may find a way.
I don't think HW4 makes the difference. The camera placement is the issue. The car simply cannot safely see cars without creeping out dangerously into intersections with the current or new HW4 cameras. This is ignoring the rumors of additional cameras coming, because there's 0 confirmation they will come to the SEXY lineup. Also, they lack rear-cross traffic alerting ability.
All signs point to them using all of the data from FSD Beta to use in their purpose built robotaxi vehicle.
Also, most who follow Tesla inventory of used cars will disagree that Tesla was ever saving cars for Robotaxi, they've been sold ASAP from day 1.
Similar threads
