Search results

That is a disadvantage of directional tires like CrossClimate and many similar winter-focused ones which seem to be directional: you can't rotate across left and right sides. Asymmetrical (but non-directional) and symmetrical you can. So if one side wears more due to less than ideal alignment...

I happen to believe it is correct. Tesla's and SpaceX's origin were 20+ years ago now, and it was the earlier period when Musk was more disciplined that he provided good leadership and value. Something bad really has happened to him, likely of his own doing. His twitter debacle means that...

Many of the FSDb problems, like this one and the lane choice issues many other people report, and the excessive speed limit changes, are a consequence of inadequate or inaccurate map data. Now we see why the other L4 robotaxi companies have to do significant geographical tuning & testing: they...

I think the idea of end-to-end *training* (not just forward evaluation/inference) is that backprop signals in a later layer (policy) can backprop back to initial perception layers and modify them to better satisfy the needs of the later layers. That is what George Hotz (OpenPilot founder)...

How? There's a giant technological problem embedded in that passive voice. Who defines the span of space of "driving situtation"? Where do the ground truth inputs come from? What if there are multiple situations? That's literally what the system Karpathy and his people designed prior to...

And how do you train a routing net to determine "the driving situation"? How do you know from raw sensor data what "the driving situation" is at any time? Is the "driving situation" something human labelled? How do you get that, what happens if there is more than one at the same time...

your average SOC is probably fairly low---do you not have home charging?

whatever has the least amount of time on average above 55%.

ION Evo are a new generation of tire tech and MXM4 are pretty old tech (10 years+). Hankook advertises heavily their uniform curing process as being a key for the ION. Presumably that's a new factory or production line with new technology.

Have you read about any means to reverse regular old calendar aging? I have never supercharged, and probably will rarely need to.

It might be true but also consider the increase in calendar aging by spending more time at high state of charge. Notice the x-axis, Full Cycle Equivalent. So I have a 2022 model 3, about 15,000 miles. With let's say 78 kWh and average 0.240 kWh/mi a Full Cycle Equivalent is 325 miles. Thats...

More the second. Look at the first graph on this thread. Eyeball at end looked like degradation rate is lowered but not by half. Let's crowdsource the estimate of degradation @ 50% vs 85%.

And possibly a cynical intentional decision that by hiring those types who might flag safety problems they expose themselves to liability before they are completely solved---a long time away if ever. We can see what Elon thinks about safety monitors at twitter. He only values 'hard core coders'.

No I still have the MXM4s from factory but these are likely my next tire. I've just been following the news and discussion.

Hankook Kinergy GT was used for some US made cars for a while, they're definitely worse than Michelin MXM4. Ion EVO is a whole new generation and they are better. I think they have switched to Ion EVO for Shanghai-built cars. Haven't yet seen it so in US. The Ion EVO is made in Korea and...

only Tesla killers will come from BYD or NIO

I think then Hankook, people are reporting better efficiency with new Hankook vs old MXM4---usually new tires are quieter but less efficient than worn ones.

I noticed from the first graph from the paper presented, over longer term the reduction in calendar aging from being <= 50% is more like 1/3rd and not 1/2 the rate of calendar aging at 85%. That may be a more realistic sense of the difference and is consistent with my personal experience.

It's obvious that he is not actually that incompetent but is likely following advice from lawyers who know the insider details of this case.

Of course they have a system that tries to detect it and it works some of the time---but it will not do as well and with as few false detections as a rain sensor, which is precise for physical reasons. Also try the above at night or in glare---I've looked at camera output in rain in the dark...

Because in the context it would make Tesla liable to lawsuits from older Autopilot which is why he was being asked anyway.

not getting fired by a petulant and vindictive boss and losing his shares

How far away is that camera from the windscreen? How far away is the Tesla built-in camera? (i'm guessing O(1 cm)) Can you record simultaneously from those in the above circumstance?

when do his Tesla shares vest?

And remember that 50% displayed on a Tesla screen is more like 52.25% in science papers because there is a lower 4.5% buffer below displayed 0% vs scientific 0%. So 53-55% displayed on Tesla might be uncomfortably close or at the shelf in degradation rate, I use exactly 50% to try to be below it.

No, roughly the degradation rate is half otherwise all else being equal. (and SOC needs to stay below 55%). The models scientifically show calendar degradation going as sqrt(time). So it's like 1/2 sqrt(time) vs sqrt(time) but that's entirely an empirical approximation and what is happening...

Wipers can't work unless they put in a ****ing rain sensor. Cameras focus at long distances, it can't see rain on the windscreen 1 cm away, unlike a human whose eyes are 25 cm back and have variable focus and look over 30cmx30cm of the screen, unlike the cameras which look through 1cm x 1cm of...

Obviously he knows what it is, as does Karpathy. A document about it might be irrelevant though in a ML context because you can't make your ML conform to some bureaucracy-decided rules no matter what lawyers say. Ashok is much more willing to BS to please the boss than Karpathy, and now...

We don't know that it's 2x, but I'd say >= 1x is probable. This is like a poorly done medical study on an area with underlying biochemical experimental results but little field data. In a scientific world, the analysts would have made stratified tests and statistical methods to attempt to...

I thought about this again and I think there is a potentially substantial confounder that should be taken into account in the data analysis: what is the typical level/distribution of state of charge between the groups? I suspect that the group that frequently supercharges, on average, does so...

there is no top buffer, and especially with LFP they need that final non-flat part of SOC vs voltage region to be accessible.

I'm sure Tesla buys them from Michelin for a much lower price, and Michelin earns a profit by pricing them very high as the 'exact OEM' replacement tire. Some people claim they don't have great grip and Pirelli Elect AS is better, but otherwise the Hankook does seem like the other optimal choice.

I know that. The point is that if the data are gathered empirically in large mass, it's difficult to impossible in many cases to distinguish which ones of them are exhibiting bad behavior and which ones are not. You can't do it by hand, you may be able to label 0.01% of the data but that's not...

curating video only works with very small datasets. The E2E training takes much more data----and finding anomalous behavior in these is just as hard as programming a robotics policy by hand with 300K lines of code and even that might not be enough.

that's L5 AGI human simultaneous linguistic annotations is easy (and already included) but human *directives* is very hard. Because now you have to ground every one of the concepts humans talk about into its correlate in the video stream and telemetry stream, and that's exactly the hard thing...

I think the T1 is the Tesla version and they may not be the same. Check specific sku/product number.

No. The E2E model creates its own policy diffused implicitly across neural networks. We don't know where it is any more than someone can look into your grey goo on your head and change a specific behavior. There's no known way to "Put that stuff in". All we get is video in and control...

Here's the problem. You have 10^8 video clips and ego telemetry. You used this for FSD 15 end to end. Training took 5 months and $50 million of electricity and occupied $1B of semiconductors for that time. Which of them has somebody "literally breaking the law"? Which law? All of them. Go.

The problem is that they tuned the behavior from observations on real data and people don't stop completely at stop signs: NHTSA forced them to do something other than what the emprical data said. This problem will grow with more end-to-end training. Thats the point, when "training it wrong"...

2029: NHTSA we dont like V14-18 behavior. Stop doing X,Y,Z. Tesla: uh, that's our entire software architecture that we've developed over 6 years. We don't know how to stop specifically X and Y and Z since drivers do that. NHTSA: tough *sugar* Tesla: does anyone know how to get V13 built back...

actually for L2+++ that is the scalable approach compared to trying to hand code those. FSD doesnt even try to work outside North America. Getting training data for those regions from existing cars and balancing the train set for their networks to emphasize those regions (probably not...

With a deterministic robotics human written policy, perhaps. But if its really trained end to end only with video input and human behavior as the training signal, there is no way to programmatically "disable any feature" instantly. It's like disabling a behavior on a non-verbal orangutan...

OK, but Elon Musk operates very differently from Waymo or Mercedes, and he's exactly the one who might be

And when these robos hit for real there will be demands for actual testing and regulation once it's more clear when average riders and other drivers experience them for real and notice their flaws. Complaints will come first from other drivers and politicians will get whinged at. Or they're...

I think this explains MobileEye's architecture then. I was surprised when they have entirely separate vision, and then on the other side direct sensing, lidar/imaging radar, stacks. I presume the vision operates the vehicle most of the time with ML heavy vision modeling, and then there is a...

Translation: they are giving up on L4 robotaxi software for a long while. Because they won't permit what would be necessary in that scenario if you consider it like testing driving of intellectually deficient humans whose brains can't be opened: adversarial human regulator fleet testing and...

That is my thesis: that end-to-end with supervised human positive driving examples helps make a more acceptable high L2 project fairly quickly which feels more natural most of the time. Current "FSD" is certainly only L2 and doesn't show evidence of advancement past that). But there's still a...

Actually you have been very nice to the battery keeping the average state of charge pretty low. You might have a bit worse BMS calibration charging to 100% less often but you'll have less battery degradation. Keep it up.

That won't make much of a difference for battery longevity. Both are considered very slow and benign charging with the size of a Tesla battery with no material effect on longevity. More important (see long threads here) is keeping average state of charge low, ideally below 50%. Most of the...

I think it's a disadvantage to train only on safe (positive) examples instead of full supervision including negative examples as policy to move away from. This requires very good simulation of negative examples, but then the system will likely learn to distinguish simulation from real from...