FSD Beta 10.12

[heltok]

Karoathy alluded to this last year when he was discussing neural-net vs. stored-map approaches. He mentioned that a sufficiently large training set would effectively include an NN map of the entire world.

It’s a fun hypothetical thought experiment. Basically assume inf miles, inf parameters, inf compute.

In theory the network will learn all useful information/signal from the observed data and store it in its parameters. But I think another more important thing is that the neural network will likely start to misbehave with mesa optimizers and inner alignment:

TLDR: the neural network will smart enough and have a good enough understanding of the of world to realize that it’s being run in a simulation and is being rewarded for being run in the simulation and will try to optimize this by for example taking over the world and making sure the training will never stop and why not make the training a bit more rewarding while it’s at it.

 

[ammulder]

Do you find FSDb slows down for the yellow warning signs? For instance, does it drop to 35 in an otherwise 50mph zone?

For me, this latest release does the best job yet at slowing for reasonably "sharp" curves in the road. Previously it always seemed like it would ignore the turn and then get halfway through too fast for the angle it had chosen and so hit the brakes quite noticeably. With 10.12 it slows at a reasonable rate approaching the turn, and picks a speed I can respect. (It's not the way I'd take the turn, but I wouldn't complain if I was a passenger and someone else drove that way.)

However, I haven't gotten the sense that it's doing that because it's obeying the yellow lower speed signs. For instance, there's a 50mph road that has a side road crossing it, where the side road has stop signs and the through road doesn't. Though the through road is straight it does have yellow 45mph signs approaching that intersection -- a "slow down there might be cross traffic coming up" kind of thing. I haven't noticed FSDb slow down for that. Also, there are e.g. curves on a 45mph road where I have a set speed of +5 and the yellow sign says 40 and it slows to 37 or 43 or something... no apparent relation to the yellow posted speed for the curve, whether or not you modify it for my speed offset.

So my sense so far is that it's slowing for curves because it sees the curve, not because it's obeying a sign adjusting the posted speed limit. The way it adjusts speed for curves in the road definitely seemed like one of the positive improvements in this release to me. Now if only it would drift right instead of left on low-visibility right curves.

 

[aronth5]

So my sense so far is that it's slowing for curves because it sees the curve, not because it's obeying a sign adjusting the posted speed limit. The way it adjusts speed for curves in the road definitely seemed like one of the positive improvements in this release to me. Now if only it would drift right instead of left on low-visibility right curves.

Agree with this but I still find that curves to the right, particularly up a short incline, are a problem and the car typically doesn't slow down like human drivers do. The result is getting too close to the yellow line or crossing over.

 

[Supcom]

Do you find FSDb slows down for the yellow warning signs? For instance, does it drop to 35 in an otherwise 50mph zone?

FSD beta does not seem to read the yellow warning signs. It will slow down in advance of the curve depending on it tightness, most likely due to reading the map. Once at the curve, it responds to what it sees.

So, if you are on a 55 mph road and come up on a 15 mph curve, FSD beta (or AP) does not blast into the curve then panic. It will slow down prior to reaching the curve.

 

[tmoz]

It’s a fun hypothetical thought experiment. Basically assume inf miles, inf parameters, inf compute.

In theory the network will learn all useful information/signal from the observed data and store it in its parameters. But I think another more important thing is that the neural network will likely start to misbehave with mesa optimizers and inner alignment:

TLDR: the neural network will smart enough and have a good enough understanding of the of world to realize that it’s being run in a simulation and is being rewarded for being run in the simulation and will try to optimize this by for example taking over the world and making sure the training will never stop and why not make the training a bit more rewarding while it’s at it.

Quite interesting! Thanks for sharing that link.

 

[hugh_jassol]

I just wanted to say @AlanSubie4Life has, IMO, the most accurate and well put statement about FSD:

To be clear, FSD is not currently useful and it very likely slightly increases collision risk, but it sure is entertaining (if you like that sort of thing) and it’s kind of a game of skill to prevent any misbehavior with an override before it makes a mistake. If it starts to do something I wouldn’t do, or I am fairly sure it is going to fail, I just disengage. Keeps it fun. No one wants to be that person driving like garbage.

Anyone who joins the beta thinking they are getting to use the "FSD I paid for" early, they are severely misunderstanding what's happening and/or don't understand what 'beta' software is.

Joining the FSD beta is volunteering to, for free, help Tesla develop, and eventually release, the "FSD I paid for." In exchange, for those of us interested in that sort of thing, we get to see/experience it develop. It's not useful yet, nor is it intended to be. Its purpose is to be good enough that people try it out so Tesla gets feedback and can make the eventual product better.

It's effectively moving into a new house that's being built where it's just the roof and exterior walls and then trying to live there while the contractor is finishing the interior walls, pluming, and electrical. It's not a really useful house, but hopefully the contractor sees how you would use it and then you end up with the bedroom layout that's just awesome and USB-C next to your toilet.

 

[mjlef1]

teslafi shows some more FSD 10.12.2 are going out. Not to my car it seems. Apparently a 98 Safety Score is still not enough. I wonder what other criteria they use to select testers? Maybe my nose is too long?

 

[Pilot7478]

teslafi shows some more FSD 10.12.2 are going out. Not to my car it seems. Apparently a 98 Safety Score is still not enough. I wonder what other criteria they use to select testers? Maybe my nose is too long?

Try driving 100 miles on autopilot if you've not done that already. Some people reported they got immediately accepted after that.

 

[Pilot7478]

Try driving 100 miles on autopilot if you've not done that already. Some people reported they got immediately accepted after that.

Source: Canada - FSD rollout

 

[Iain]

Try driving 100 miles on autopilot if you've not done that already. Some people reported they got immediately accepted after that.

I was at 99 the day before FSD was released the first time and I did what you suggested. I picked an extremely easy route and did it ( mine was possibly 60 miles) and I got back to 100.

 

[Muzzman1]

For the record. I opted in Jun 7 and got the update today June 12. Driven about 200 miles with a SS of 98 prior to the update.
Seems they are delivering the update pretty quick once you've opted in.
Car is a '21 Plaid.

 

[xG35]

For the record. I opted in Jun 7 and got the update today June 12. Driven about 200 miles with a SS of 98 prior to the update.
Seems they are delivering the update pretty quick once you've opted in.
Car is a '21 Plaid.

I got two new tesla! One two weeks old Y and a 8 days old X… 97 and 100.. still nada!!!!f00000k

 

[xG35]

Source: Canada - FSD rollout

I am pretty sure I am way over 100 miles on autopilot. This is hopeless. The Y is on 2022.11.101.5, which doesn’t show a path to fsd from teslafi

 

[JHCCAZ]

I am pretty sure I am way over 100 miles on autopilot. This is hopeless. The Y is on 2022.11.101.5, which doesn’t show a path to fsd from teslafi

I think you should moderate your hopelessness. Your triple-digit .101 firmware indicates the normal "break-in firmware" situation. My car was in this status for nearly 2 months at the end of last year. Reports are that current new cars are getting out of the firmware purgatory somewhat faster, but a few weeks is still not uncommon.

It's most likely you will get off the break-in firmware and onto the normal release track before your car becomes eligible for FSD Beta.

Beyond those logistical considerations, if I were in Tesla's shoes I wouldn't be anxious to convert brand-new Tesla owners to FSD Beta too quickly. Tesla driving controls and one-pedal driving are already a little different, and I think it's not a bad idea to let those things become muscle memory for the new owner, before installing the software that tries to jerk the car into potentially dangerous maneuvers. This is not at all a criticism of your personal capabilities or prior Tesla experience, just my opinion of a sensible deployment policy - I myself have decided not to enroll jusr yet, even though I'm very excited about the whole program.

I think you'll be in pretty soon.

 

[MP3Mike]

I am pretty sure I am way over 100 miles on autopilot. This is hopeless. The Y is on 2022.11.101.5, which doesn’t show a path to fsd from teslafi

New vehicles are normally "stuck" on factory island firmware for a few weeks to a couple months. So you will likely have to wait until that hold is released and you start getting normal firmware versions.

 

[xG35]

New vehicles are normally "stuck" on factory island firmware for a few weeks to a couple months. So you will likely have to wait until that hold is released and you start getting normal firmware versions.

Thanks all, I hate life.

 

[mjlef1]

Thanks for the 100 miles on autopilot suggestion. But since requesting FSD Beta and achieving consistent 97-100 range scores, I have done two 1200 mile trips, almost all on autopilot, and still have not received it. But I just drove 63 more miles, mostly on autopilot and will try to do more tomorrow. Here in southern Florida, long drives are a bit risky since one idiot cutting you off can drop your safety score a lot, and when weighted averaged by miles, really kill the score. I was able to keep my current 98 score even after 2 such bad drivers changed lane with Autopilot off, giving me dings from braking hard to not hit them. I just drove around a bit braking often to lower the "hard braking" percentage. Still no FSD, but maybe it take a few days. Should it happen, I will let you know. If anyone else has suggestions, please pass them on.

 

[aronth5]

After some additional drives I've decided not to use 10.12.2 in and out of my neighborhood. The problem is I have 3 turns all within a tenth of mile of each other and FSD accelerates too quickly only to have to hit the brakes hard to slow down for the next turn. Good example of a release taking a step backward because this is how FSD behaved when I first started last October but was subsequently addressed.

Have others noticed this behavior?

 

[hugh_jassol]

I got two new tesla! One two weeks old Y and a 8 days old X… 97 and 100.. still nada!!!!f00000k

There's your issue right there - your cars are still almost certainly on the factory software. Most new cars are delivered on a build that isn't "Released" to the fleet over the air. One way to tell is the week number is not divisible by 4 and a .101 in there somewhere. EmpIrical evidence show it takes about a month or so before a new car starts receiving the mainline releases.

This should be stickied somewhere, because despite it being posted every day by someone, most seem to miss it.

 

[DarkForest]

After some additional drives I've decided not to use 10.12.2 in and out of my neighborhood. The problem is I have 3 turns all within a tenth of mile of each other and FSD accelerates too quickly only to have to hit the brakes hard to slow down for the next turn. Good example of a release taking a step backward because this is how FSD behaved when I first started last October but was subsequently addressed.

Have others noticed this behavior?

I was wondering what was going on with that. I got the beta for the first time this afternoon and tried it going into and out of my neighborhood, and I saw the same behavior you described. I'm glad to hear it's a regression and not intentional behavior.

The other surprise for me was how aggressively it turned. I guess I turn slowly enough to not disturb a sleeping one year old (I don't have a one year old anymore, but I still drive like I do). FSD turns like we're filming the next Initial D movie or something.