Search results

How is fault proven for non-AV collisions? There is no ground truth in those cases either; the best case is that there's dashcam video, which may be sufficient to give an approximate idea of what happened and who is at fault. If Tesla ever achieves a pure-vision Robotaxi, its eight cameras...

This describes a classical database lookup, not a neural network. A neural network does not maintain a database of its inputs; all the information in the input training set gets irreversibly (and lossily) condensed and mashed together into the neural network, with no way to retrieve or extract...

The point of the pure E2E approach (photons in, control out) is that there's no need for the network to explicitly calculate distances at all. (Unless it's a two-stage modular network with a human-designed occupancy layer in the middle.) Elon's statement about"LiDAR no longer being necessary"...

LiDAR is closer to ground truth than depth inferred from pure vision, but neither is exact or deterministic (due to sensor noise, etc). The real world is far too complicated to have a formal process for a "right outcome" (control-wise) for any given input, otherwise the C++ coding approach would...

"Can't explain" is only in the semantic sense, not in the arithmetic sense. The neural network is deterministic and will always produce exactly the same output for exactly the same input. This can be mirrored in equivalent C code that is guaranteed to have functional equality (though drastically...

Sort of. The network doesn't maintain any independent memory of the inputs; it just synthesizes them all together into a combined network. So there's no such thing as "identify the closest response that it was trained on". Rather, through the training on millions of examples, it's learned to...

Correct, you could. Of course, the C code would be absolute unintelligible gibberish (or at best, a sequence of gigantic matrix multiplies with random-looking numbers), even if you could compile and run it and get the same results as the neural network. Put another way, you could show that...

I don't think this is what Tesla is using the LiDAR for. Neural nets are extremely opaque; even with LiDAR ground truth it's essentially impossible to "know why" the neural net makes a particular decision. Rather, they were likely using LiDAR to train v11, which had an intermediate "occupancy...

True that any Von Neumann machine is at a very abstract level equivalent to any other. The difference is that an "IF ELSEIF WHEN DO LOOP" is by its nature very serial; all the processing goes through a single tiny bottleneck on the chip that does the testing and branching. A neural network, on...

We are there for simple reasoning. Complex reasoning, not yet. This sounds like a "lookup-table", which is not really a good analogy to how modern neural networks work. Rather, a neural network synthesizes its vast amount of training data (which could be exabytes, orders of magnitude more...

Loosely, "AGI" means the cognitive ability to solve most problems that a typical human can solve. It may or may not include the ability to "learn" (self-training of its own neural network), or to have a "mind", or "consciousness", though those terms are also quite ill-defined. True that no one...

Training the next iteration of the software will always (or at least for the foreseeable future) be centralized. The software will be identical for each car for each drive; it won't "learn" as it goes. Otherwise you'd end up with millions of cars, each gradually developing its own distinct...

“This rocket sure has a short development arc.”

There are two cameras in the front housing. Is there any way to check whether the other front camera has a similarly faded view? My 2022 MY had always had poor front camera image quality, which turned out to be due to chemical offgassing residue inside the housing. May be worth checking for in...

AGI will require a neural net that can take its own output into account. In other words it must be able to iteratively process information in order to "think" about it, and must have a memory of its own thought process. (ChatGPT sort of does this; I'm not sure if the current FSD architecture...

A useful definition of intelligence is: "The ability to solve problems that one hasn't encountered before." The video makes the same questionable assumption I mentioned earlier, which is that it assumes that the FSD network can't be good enough to extrapolate solutions to not-yet-seen edge...

No, it’s part of the long tail.

This was a really interesting read, thanks for sharing! The blog post does make one very questionable assumption, which is that “long tail” events are independent and uncorrelated. To give an illustrative example, suppose two events with 10^-4 probability are “struck by lightning” and “attacked...

I’d like to see its, er, decision tree. 🌲

“Officer, an invisible tree came out of nowhere, hit my car, and vanished.”

https://en.wikipedia.org/wiki/Trolley_problem

Completely agreed. L4/L5 FSD is still several years off, and almost certainly unachievable to the required degree of reliability with current hardware. There's always going to be a multi-way tension between pre-existing map data, cached databases of other drivers' previous behavior...

It's sort of incongruous that NHTSA is just fine with allowing the driver to override the speed limit via preferences (e.g. you can set FSD's top speed to "Speed Limit + 15%"), but they're absolutely opposed to overriding stop sign behavior via preferences, even though the former is arguably...

It just doesn't roll off the tongue the same way.

Different shape, but probably a similar number of pixels on the camera feed. (Kids are taller but skinnier.) My point is that the network has probably already been aggressively trained to avoid hitting humans specifically. (Or if it hasn't, it should be.) E.g if the choice is to hit a human or...

NHTSA did for all practical purposes force Tesla to enforce complete stops: Tesla to disable ‘rolling stop’ feature after NHTSA says it can ‘increase the risk of a crash’ The stop-10-feet-before-the-intersection-then-creep-forward-then-stop-again behavior is different though. Maybe we're saying...

Not all obstacles are treated (or recognized) equally. My Model Y on v12.3.4 recently ran over an easily-avoidable 5-gallon paint bucket (thankfully empty). If it had been a person of similar size, I'm sure it would have recognized it and tried as hard as possible to avoid or minimize a...

It's easier to solve hard problems 99% of the time (useful L2) than it is to solve slightly easier problems 99.99999% of the time (useful L4). So there's not a contradiction here. It may even be that once L4 is achieved, the driver will still have the option to enable "L2/L3 mode" for this...

It's end-to-end AI, just with carefully curated (and/or simulated) training data, particularly for stop signs. I'm not sure what the stop-10-feet-too-early-at-intersections thing is about though; it seems unlikely they would deliberately program or train that in? Perhaps it's an unintended...

Is this what happens when you pulse DC rapidly on and off 60 times per second? I always wondered.

It’s not the ease in which it was done. It was apparent lack of forethought or planning or consideration in which it was done. Musk: “This 747 jumbo jet we’re flying is too inefficient. Reduce the size of all of its engines by 14% immediately!” Lead engineer: “But sir…” Musk: [chops off one...

This technically describes Level 4 as well. The distinction between them is that an L3 system may request on very short notice that the driver take control, even while the car is in motion, and even within its ODD (the “specific conditions”), whereas an L4 system must be able to at least pull...

Which evidence? The VTTI study seems to show that accident rates decrease roughly in proportion for all severity levels (for autonomous car vs human), at least for levels 1 - 3 in one of their charts. This seems to imply that e.g. halving the minor collision rate (through software/hardware...

I think the most likely change in outcome (if any) is that some manufacturers decide to leave CCS ports on their vehicles for a while longer and provide clunky NACS adapters, rather than build in NACS directly. Also, as a multi-Tesla owner I'm disappointed that the network probably won't...

I'm not so sure it is either. But I'm also not so sure that it isn't. The consequences are unpredictable. Perhaps manufacturers will decide to keep CCS connectors as the default on their cars (plus clunky adapters for NACS), rather than switching the cars' built-in ports to NACS, because they no...

Betamax was a standard, too. (Technically superior to VHS. We all know how that went.) The point is that there's substantial overhead for other manufacturers to switch to NACS as their charging standard. They might have been willing to take that hit because access to Tesla's Supercharger...

This is what stall-sharing is for. It allows the network to theoretically operate at maximum power output (maximum profitability) even when only half the stalls are occupied, yet still accommodates the full number of cars at peak times (with some drawing less power as their packs fill up)...

This is reversed. Tesla was on the cusp of getting buy-in and support from all the other manufacturers for the NACS standard, which would have ensured that it would become a thriving ecosystem with support and contribution from all sides in the future. But by his recent actions, Elon has...

It didn't crash, it just added more polygons.

To be fair, "imminently" is also one of the accepted definitions: Definition of MOMENTARILY Obligatory viewing:

Reporting a pothole is a trivial amount of information; it would fit in a text message. (Just the GPS coordinates, and perhaps a few bytes to describe the size / position of the pothole in the lanes.) The car should be able to do much of the real-time detection and avoidance on its own, but with...

My guess is that v12, being E2E trained, will have quite a lot more headroom than the earlier brittle heuristic-based approaches. Perhaps the current approach will eventually reach 1 disengagement in 1000 miles. But I don't think it can plausibly reach RT level (1 safety-critical...

The robot will do whatever it's programmed (or for a neural network, trained) to do. The question of how to appropriately balance risk with reward is a highly nontrivial one, for both humans and robots. An AV (or human) trained for zero risk would never leave the driveway. So no, we don't expect...

I've experienced the same thing, for both v12 on city streets and v11 on highways. It seems to strongly prefer the leftmost and rightmost lanes, leading it to frequently make such mistakes (for both left-turnoffs and right-turnoffs), when arguably the default bias should be (when a lane splits)...

You're right, it would probably make it 99% of the time. So do absolutely terrible human drivers, even drunk ones. The 99% statistic doesn't make FSD a good or safe driver. A significant number of accidents are caused by two cars making otherwise-minor mistakes at the same time and place. If...

It's much more than just a shape difference. One of the significant innovations of NACS was that it uses the same physical pins for both AC and DC, whereas no previous connectors (including CCS) had attempted to do that. That's part of what allows NACS to have its compact form factor.

My point was that Tesla FSD is still regularly failing even in really easy scenarios. If they can't solve these common issues, they will not be ready (in my book) to even remotely consider L3 or L4, or even to be considered a "useful" city-streets L2 system, regardless of whom else has been...

I don't short stock on principle (and because of the unlimited downside risk), but I've divested significantly over the past couple years as Elon has become more and more erratic. Tesla as a whole still has tremendous potential, but Elon's recent decisions have made achieving this full potential...

I interpreted the post as taking a few minor but real v12 flaws, and overexaggerating them for attempted comedic effect. The current state of FSD (to me) is that it's incredibly mind-blowingly amazing, yet at the same time, not nearly where it needs to be. The incredible part is the super-wide...