Actually.. So, have this 2018 M3 LR RWD. I've had the FSDb for roughly two weeks now. And before that, had been running around (first) with EAP, then with the pre-Beta, FSD package. This last is basically EAP with a couple of extra features. As my spouse puts it, this is my hobby. It's cheaper than having a boat or an airplane
.
I'm also an engineer who, for a living, figures out what died on complex equipment, with the intent of making sure that
that never happens again. And I have a kind of eclectic background for a EE: Lots of down-to-the-silicon hardware design, circuit boards, and lots and lots of software. I've never coded actual neural nets, but, what with a digital signal processing background, knowing about Markov processes and such, I very definitely get the idea.
So, FSDb is all about the testing, not the features. If one wants to chug a couple hundred miles or longer down the road and feel relaxed when one gets to one's destination, FSD is very much the way to go. Not that I use the auto lane change feature on interstates that much (feels a little too risky for me), but the rest is just fine, including going between two interstates via the ramps. And the current FSD (not beta) can do lane keeping and some stop-at-stop-lights stuff, but won't do turns on local roads.
FSDb is not for the weak of heart. The release notes state that the car can and will do the wrong thing at the wrong time; they're not kidding. On average on a 20 mile local road trip I hit that "record" icon anywhere from five to twenty times. Each one of those hits is a "safety" hit. Going down the middle of an unstriped, 35 mph road going up the brow of a hill and not being able to see over it for oncoming traffic. (Locals know to hug the right side of the road - but not FSDb.) Jerkily going through intersections, scaring the bejeezus out of nearby cyclists (that was today). Classifying the danger zones as mild, serious, and critical, I'd guess about 50% mild, 35% serious, and 15% you-gotta-be-kidding-me where, without manual intervention, we're talking about bending metal. Believe me, I look at people pushing baby strollers or with toddlers in tow in a whole new light these days. (Yes, the car came to a halt for the lady with the stroller on the green-light left turn but, given all the errors this car makes on a regular basis, the knuckles on the steering wheel were white and the foot poised over the brake. And it was all
very jerky.)
As for all those people who are crazy to get their hands on FSDb, along the lines of, "I paid for it! I want my feature!": be very, very careful what you wish for. My stress levels are loads higher than they were before getting the FSDb.
Now, the chatter about FSDb on these forums is kind of interesting. I've only had the current version, 2022.12.3.20. People who've been testing (emphasis,
testing) this and earlier versions do say that various previous versions had smoother turns; or hugged the center line closer on a left turn (something the current version definitely does
not do), and so on. It's been stated that each time a new FSDb hits the wall, some features get better, some features get worse. Mm.
But the release notes themselves have an eye-opener in them. One of the things in there was that the car was 41% better on certain types of left turns. Ah, yeah. 41% better. So, what about the other 59% of those types of turns?
On the one hand, I'm very happy that Tesla figured out, what with training and all of the neural network, that they could get a major improvement in a particular feature. And it looks like whoever wrote that felt equally happy and wanted to share.. and maybe alert people for feedback on that topic.
But the goal of FSDb is, well, Full Self Driving. As in Taxi service and that. And I'm watching the car and how it handles things. Nope, the car's not a human. It appears to have some short-term memory. Long term memory, as in, "I remember that screwy intersection from the last time I went through it, I won't make
that mistake again!" is Not In Evidence. Or, as I mentioned above: If one scares a student driver with the dangers of approaching the brow of a steep hill in the middle of a road, one typically doesn't have to scare them twice. The car appears to be running on a rules-based approach to chugging down the road. Which is fine.. I suppose. The rules can be very, very complex, for a computer. But car doesn't feel
danger, that's a human (or live creature) emotion based upon a zillion years of evolution. That, you know, kind of works pretty well. (Those for whom the internal manifestation of fear/danger didn't work all that well are no longer around to complain: Hello, Darwin.)
So, the real question is: How well, and how fast, is Tesla iterating on FSDb? Take that 41% number. If on every release that feature is improved by 41%, then we get 0.41^N, where N is the number of releases. So, starting from 1.0 (which is lousy, at the beginning), one goes 0.41, 0.168, .0689, .0283, .0116, .00475, 1.98e-3, 7.98e-4, 3.27e-4, and so on. that last one is 0.003%, after 9 releases - and that's not good enough. We want 99.99%, better than a human, or a mass of humans in aggregate.
There's all sorts of things wrong with that math, above. For one thing, there's nothing that says the neural net improvement process can't go from, say, 0.11% to 0.00001% on the next step. Neural nets (and computer algorithms in general) can be decidedly non-linear in that way.
And I'm personally looking at a capability sample size of one. Which is getting into a wider release with Tesla drivers. Remember: At this point in the algorithm cycles, it's all about gathering sample data and processing same for the next FSDb point release, and you gotta figure, it's Tesla requesting the data.
I've got this feeling.. that, with whatever passes for the current techniques, Tesla is not progressing all that fast. Are the people monitoring all this at Tesla
confident that they can get a non-beta, FSD out the door in the next year that works? Or are they getting, say, a little desperate? Don't know.
One possibility is that the development crowd at Tesla is betting on the Dojo system. Kind of a hope that, "If only we could train this neural net in the car at 1000X the speed we're currently doing it", all the major stumbling blocks could be fixed. Which may or may not be true, of course. What Tesla is doing is along the lines of pure research, which is defined as the Process of Running Up Alleys to Find Out If They're Blind. Nobody's really done a full FSD before.. so there could always be a stumbling block beyond which it might be near-impossible to get.
Or we could all be pleasantly surprised come December with a smooth-driving, traffic/pedestrian/what-have-you aware car that works like a dream.
At this point, I wouldn't bet either way.
In any case, I'm going to keep on testing the FSDb, not because it lowers my blood pressure or is that capable, but because the data from me and everybody else doing this may just be the push that gets Tesla over the edge.
Finally: As I said above, nobody should get the FSDb if one thinks that THIS is going to let one take one's hands off the wheel and read a book or something while driving from A to B. If you want to help Tesla, sure. If you're
really into beta-testing buggy software, sure. But if all one wants to do is get around town with one's sanity intact: Don't bother.
