Search results

I just had one of my rears fail for this reason. 15k miles and plenty of tread on the whole tire except for the inner edge. Ordered a new set all around. Can keep the old ones if Michelin wants them back.

Looks like Quartzite is open and will be through thanksgiving but is having problems so charging is slow. There are 8/12 portable urban chargers present but half the regular chargers seem to be experiencing problems and are offline. Even the urbans are slowish. So if you need it it’s here but...

That 'monster' probably can't be run on HW2+ in a fashion that would actually drive the car. It's too big. We probably only got to see it because of an accident in the distribution system. It could be something that they were trying out in anticipation of the FSD chip coming out. Maybe we'll...

I received a copy of the V10 networks but haven't had time to take them apart yet. My own experience is that V10's behavior is improved in some significant ways and I'm curious to see if this means that there have been changes in the nets that can be interpreted. It's worth mentioning that...

From the autonomy day reveal we know that Tesla had come up with a target of 50 TOPS for the FSD chip before they developed the chip back in 2016. If you build a chip with 50 TOPs it's because you want to run bigger networks - much bigger networks. AKNET_V9 is a much bigger network. That size...

I'm in the middle of a long road trip (10k miles on AP!) so I can't offer you much detail right now, but I agree with your comments about Karpathy's history and what he's doing now at Tesla. It really looks like that V9 network (I later renamed it AKNET_V9 from the definition file label) was...

Why would AP2 HW be better right now than HW3? The NN compute capability of HW3 is expected to be far superior to AP2, but the architecture also changes in ways that are not a superset of what AP2 does. Because of this the NN code has to be reimplemented on the new hardware and they may...

HW2.0 to HW2.5 sensor changes include a different part number for the radar (from a different vendor) and different part numbers for the cameras (from the same vendor). The exact spec on the radar isn't known but a similar part number from the same vendor and a comparison to the wiring harness...

That wasn’t intended as censure. Just an opinion. Please feel free to keep wasting time with the rest of us.

Oh man I seriously LOL’d on that. Good one.

So the disconnect is that we’re comparing an unfinished L4 system to a consumer released L2 system? If manufacturers final product design intent in the discriminatory factor then you’ll get weird results when you apply it to the current operating features of an incomplete test vehicle. Does...

I think this probably qualifies as a personal attack.

I agree. Producing their own NN processor required Tesla to build out a dev team and infrastructure and put money into IC development for years before the IC was ready. But all that is sunk cost and much of it was years ago - especially the risk margin. Incremental cost of fabricating units...

The vector unit in the patents is an accessory to the AME which expands the range of neural network layer operations that can be efficiently implemented with the AME. The AME design probably incorporates this vector unit in it's basic conception, as well as the various control systems and...

pdf of the main patent

I had a look through the recent Tesla patent release. There's one main patent which was filed for in Sept 2017 and 3 extensions filed six months later. The main one is titled "Accelerated Mathematical Engine" - I'll attach the pdf here if I can. The others are supplemental and offer little...

Those estimates are so rough they shouldn’t be used as the input to anything. They could easily be off by several times in either direction. The point of making them was to answer the question of whether the numbers @verygreen found are plausible- and I think they are. Apparently it’s unknown...

Your statement about off highway fatalities being higher gets quoted a lot but it’s false, at least in the US it is. IIHS statistics show rural highways, the sort of place Browns accident occurred and where AP is frequently used, are the highest fatalities per mile. Surface streets (off highway)...

Two months ago Tesla announced 10B fleet miles. Last month they announced 1B autopilot miles. But yeah, my AP usage is over 95% and I was surprised. Lex Fridman’s group at MIT was guessing 20%, which is what I expected also. They had to redo all their charts last month when Tesla announced the...

How about "Tesla Redundant Inference Processor" for trip?

Not to be a pedant, but I think it’s important to keep in mind that the appropriate comparison today is “human with AP” versus “human without AP”. There’s no “AP without human” in the current mix. I’m a human and when I drive I prefer to do it “with AP” rather than without.

Speaking of a concrete feature that requires zero-driver: I've been thinking about an analysis done by Brad Templeton from a couple of years ago. He was thinking about how self driving cars impact the economics of parking garages and worked out that - because you can redesign a parking garage...

I get motion sickness so easily that for me the answer is probably never. I can't read a book as a passenger on a straight flat road. Which really sucks because I love to read and I'm on the road a lot. Sigh. Though it does give me plenty of time to observe the surroundings. BTW I counted...

The choice of sound track is rather interesting.

Regarding the man himself - I was refreshing my memory of his history and found this recent article which I found quite illuminating. https://www.newyorker.com/magazine/2018/10/22/did-uber-steal-googles-intellectual-property It includes a lot of interesting background on Levandowski and his...

I am assuming he excludes Tesla's effort there because he believes (or maybe he knows?) that they aren't doing end-to-end as part of their effort right now. Or were you thinking of someone besides Tesla? It seems pretty likely that *someone* else is doing it. For instance, Comma AI might be...

Anthony Levandowski - arguably the most controversial and one of the most experienced program leaders in the AV business - has reemerged to start a new venture in the AV space and has (below) dropped a commentary on medium refuting the dominant narrative in the AV space. Levandowski has changed...

That statement is in reference to an earlier version of the software that was only calibrating a couple of the front cameras. Recently "calibration" means a lot more, much of which is certainly not related to stereo binocular calibration. Whether you initially calibrate at night or not is...

Sorry this is wrong (doh). Parameters get used HxW times (not HxWxC). Kernels are KxKxC in size (so KxKxC parameters per kernel) and they get used on a HxWxC input frame. So total parameters in a set of kernels for a layer is KxKxCixCo (Ci is input frame channels and Co is output channels -...

This talk is an extended simile that is explaining a proposed approach to developing SW2 (NNs in this case) to test driven development and it relates many of the recommended practices to conventional programming best practices. In the section you cite Karpathy is saying that, similar to how...

Yes, frame rate matters, but there's a more important item that you might be missing. Parameters in a CNN get used many times and how many times they get used per frame depends on the network shape. A single CNN parameter gets used HxWxC times per frame. H = layer input height, W = layer...

AK is making a comment on how performance on a particular training benchmark has evolved over the last couple of years with several specific recent examples. The table is from a recent paper by a team of people who did the work that AK is commenting on. The particular numbers aren't indicative...

Interesting - it sounds like you're saying part of the calibration process is possibly color calibration. Is that right? I'm not familiar with how much color variation affects NN performance or how much color variation there is in commercial cameras. Because NN's need to be relatively immune...

Yes, as far as I know that is true. And I don't know what the actual geometry sensitivity is as it's not something I can work out from first principles in this case. The vehicle dimensions thing was just a quick example of a different sensitivity that the networks might have. But your point...

Now *that* is an excellent observation. To be sure, it seems that the advanced AKNET_V9 network - the camera agnostic one - is probably not the one most of us are using in V9 right now. So for the time being this issue probably doesn't apply. But the observation that making the network less...

Thanks for sharing that - really interesting. AK's talks are one of the best windows we have into the technical details of AP right now. He seems to intentionally not talk about Tesla in particular but the examples he provides from his experience there provide us with some insight nonetheless...

If you're talking about Comma AI's strategy I wholeheartedly agree. It's really very clever. They get all the advantages of being open source along with IP control that hollywood can only dream about. And the ability to sell their hardware as a 'dashcam' lets them avoid control by the NHTSA...

I hear from the folks who provide me with files that hacking Tesla's has been getting harder and harder. There seems to be an expectation that at some point it'll be outside the reach of amateurs. I recall that the iPhone was relatively easy to hack for several years when it came out but that...

Agree. There are certain aspects to the use of NNs in software which is a kind of IP protection that is very powerful in and of itself. First - NNs are famously very black box in that they cannot be precisely understood in any straightforward way from an examination of their structure or...

Yeah, they really move around a lot don't they? Very small global changes in the image (even clouds and thermal noise) will cause variations in the NN outputs for segmentation and bounding boxes, but the movement in the driver display is a lot larger than that. I think the UI probably has a...

Yeah, cloud TPUs are beautiful. I actually got offered the use of a TPU Pod for a month (105 TPUs, I think) and it's so very tempting. But I'm doing RL these days so gradient descent isn't the main problem I'm facing. Sigh.

They certainly don't ignore the target application when they design the hardware. But the precise details of the hardware will be tuned to optimize the economical throughput for a set of operations that are a good match for the target application. You get into the ballpark of what you think is...

The basic dataflow in aknet_v9 is just inception V1. Tesla might be using AUTOML ideas for hyperparameter search but they haven't done it with any of the camera network architectures for which I have seen detailed architectural data. As for why they would or wouldn't be using this stuff - from...

You might want to consider that the final training run for a job is a small fraction of the total time used to develop a neural network. Often you run many variations of a training run over a long period of time before you find a formula that works and are able to generate the network you want...

AWS pricing varies a lot depending on your job constraints, machine type, interconnect and storage requirements, but the pricing you cite is in the ballpark.

I've detailed that stuff in posts near the beginning of this thread, here's an early one: #7 Bounding box outputs from NNs aren't generated as drawn boxes, generally. The output is derived from an image map that breaks the image into sections, perhaps as small as a single pixel but usually...

It's not my data to give out. That said - there seem to be a number of people who have the ability to acquire it on their own and are happy to share it. Anyone genuinely interested can acquire it if they ask around politely.

You know, I think we’ll get to full self-driving next year. As a generalized solution, I think. But that’s a ... Like we’re on track to do that next year. It's worth considering that "getting to FSD" might not be the same as "FSD shipping to customers". Having it working internally would let...

Nice work!

I was super excited to read this stuff. Of course Elon always sounds confident, even when he's talking about really hard deadlines. And people have been wrong about self driving cars for a long time so there's plenty of precedent for being overconfident. Still, I'm really happy to hear this...