Search results

The limiting problem is data availability, and proper data availability and conceptual progress. Look at the large language models. By assimilating nearly all human readable text, they can simulate text very effectively. The central problem technologically/conceptually is that they have no...

I just saw on SimpleTire that in the USA the summer version (coded IK01, not IH01) is available for the 18" wheels. Unfortunately, it has no mileage warranty and tread starts at 8.5/32" vs 10/32" in the all seasons. On the upside this is the tire that beat the Michelin Pilot Sport 5 in most...

Almost always the same tire when new will have lower efficiency than the same tire worn (as higher tread blocks move more during load, dissipating energy). So a new iON having better efficiency than an older tire, particularly a low rolling resistance tire, is an impressive achievement.

This sounds like you weren't actually using FSD Beta. The highway performance of real FSDb is better than that. Your car can be entitled to Full Self Driving (getting you autopark, summon and navigate on autopilot) but still not using the new FSDb stack. I mean FSDb has a long way to go, but...

The real issue to me is 'how numerous, and how *good* are the supervision labels". The quantity of bits on the input is a burden---the quantity/quality of bits on the labels is the gold. The ML performance is limited by the quantity of good labels in the low cardinality classes/rare cases...

Yes. There is a step function in degradation rate vs SOC because of complicated chemistry. Best to be on the clear side of that.

The question is what is the "test data" and loss function on it? Test data of driving policy is much less dense than test data of generative video---predicting video 2 seconds ahead has tons of self-supervision labels, no argument. Driving policy is more important, risky, and difficult to...

I personally think Tesla will give up on driverless (but never admit so) software---they will sell the robotaxi platform to others who want to try and lose money doing it. E2E will make it easier to make a pretty decent L2++++ driver assistance product across many localities with much less effort.

Battery chemistry doesn't change that much. Maybe the general rate of aging in production might be slightly lower thanks to proprietary CATL additives and production techniques, but the trend and the change in rate vs state of charge will still stay there as that's a fundamental chemistry...

Ashok was saying that these might be intermediate concepts it figures out---but the real issue is if there is any modeler added additional loss function and training signal applied to them. HIs answer seems to be "no" and that's different from what Karpathy did.

The hack around this, in a densely populated area, is crowdsourced routing. I.e. not a neural network approach, but a big data approach. Back at the Mothership Cloud, the server would know that most people who wanted to end up in a location which is along your desired route (say a waypoint...

The fundamental reason the LLMs 'hallucinate' is that they have no valid internal knowledge that distinguishes "this is real" vs "this is something that I've read". The distinction that even children understand as they can truthfully say something is "pretend" vs "real". It's a major...

Follow what AAKEE says. Plugging in all the time is not the same as charging all the time after you reach the charge limit. Keeping it plugged in at a low charge level (50%) so any accessory use in on wall current and not the low-voltage battery is beneficial. Setting max charge level to...

I think it would be almost the same, yes. Much of the degradation at higher states of charge is reactions at the graphite anode, which is the same in NMC vs NCA.

Staying at 100% does harm the LFP battery, because it uses the same graphite anode as NCM/NCA, but at lower voltage. Generally overall LFP degradation is lower for calendar and cyclic though.

Thanks for the update. Science is mostly consistent. Though the graph doesn't show as much sqrt(T) effect vs linear in T effect for calendar aging as I would suspect. The practical message stays the same to me: set charge limit to 50% if you don't need longer distances. Helps for both...

It will not. It may learn from a few cars driving in your local region if they happen to sample them and upload their data, and have it selected for use in the next update's training. Online local learning of ML models is a very difficult and risky prospect, especially as there is no easy way...

I've been re-thinking the end-to-end neural network training debate (on other threads), i.e. Tesla's new strategy of building in nothing in any intermediate perception or labeling stages, and only using observations and human driving in a generative/corrective architecture. The quoted example...

Everyone uses machine learning in perception. How much machine learning is there in the drive policy? That's the critical question.

Nobody knows how to make "rails" for such a system, without re-building the conventional policy with deterministic coding. Now if Tesla's problem was that they have a deterministic solution with sufficient performane (like Waymo) but it's too computationally challenging to run on-board...

That's just the reality of LFP cells vs NCM/NCA cells---LFP is very good at getting charge in and out with low resistance & heat, at cost of lower energy density. LFP is better at everything than NCM/NCA except cold-weather performance and energy density (mass and volumetric). I don't think...

It's so logistically difficult and also prone to liability-inducing problems: your teenage son drives your car like a drunken maniac, and then when you go to use FSD for a nice calm commute it acts up. Humans have a way of trolling and poisoning self-learning AI systems very quickly: look at...

At a technical level the question is "what is are the loss functions, where are they and where is the ground-truth label coming from?" That's the core decision for the nnet training.

That's the more useful intermediate situation. Technically its "model distillation"---and they could estimate effects a more expensive deterministic code/optimization basted planner that has higher driving performance but worse computational properties that made it infeasible on-board. But the...

That will be a major problem trying to connect it to the maps and desired routing information. Humans are explicitly instructed on what a lane is, a stop sign and all the other semantic content of safe driving is about. Going to V12 helps get rid of labelers, but it's going to make engineering...

It limits the maximum power output and changes the mapping from pedal depression angle to power output.

Chill mode does reduce max power output. Not quite a half but significantly.

I assume that's 18.5C and not 18.5F ? I don't think that will be true once you start driving and using the batteries. get ready for your arteries to pop, there's >1 with physics PhD here exactly yes. The battery could be damaged by asking for a very high current when it's too cold. almost...

I think the battery heat losses to environment are significant and there isn't significant insulation, and they want it that way. Generally too high temperatures, not low temperatures, are more of a concern as they cause permanent damage to batteries, even though for the short run it's easier...

In a dual motor car, one motor is permanent magnet and the other is induction. The induction motor is typically used only for higher acceleration as it's less efficient, as it needs to consume and dissipate electricity to magnetize the rotor, whereas in an permanent magnet motor (with either...

The individual quantum nature of the EM field is not measurable practically in radar, but it is so in CCD detection in optical frequencies. CCDs will read out counts (though with well under 100%) efficiency and there is stochastic noise that's not due to sensor noise, unlike radar. The words...

All nets can work, but only if there are lots of rules which went into making simulated data and filtering/scoring real data as "yes drive like that/no that's a bad idea", with the nets 'distilling' the rules. The off-line optimization/rules can be too computationally intensive to do on-board...

What is not clear at all is if the neural planner is a distillation of results from an offline (train-time) physics-based numerical optimization done at higher-resolution/computational load, or if the planner is trained only from empirical observations? The second is implied by Elon but I have...

What "Changed the World" the most is top level research from large entirely PRC teams is the norm. That organization (SenseTime Research) has hundreds of very good ML papers, much more than Tesla.

That's a very good self-supervised heuristic, I like it. Now the problem is "draw a box around the good drivers". As in literally---if you have no idea about the structure of the internal representations because it's all been trained end-to-end, how do you make a metric on something to cluster...

Another bad Elon decision. The cameras cant do it because the focus is wrong, they're focused at hundreds of meters instead of 5 millimeters. You as a human have an eyeball 50 cm away from the windshield and can see that there is rain on it. If your eyeball were on the glass it would be hard...

Perception isn't the issue, but control---you need to output different control signals and have different control algorithms in different regions. The training targets need to be different and some of the networks would have to be segmented and conditioned on regions.

And how exactly does one sort those out by the millions in an automated way? Build a machine learning classifier? Well, how do you judge the good ones? Maybe you need a quantitative optimization target and explicit physics-representation and some rules about ideal driving behavior?

I agree 100% with Karpathy here, I had some thoughts but he said it even better. Karpathy is a major genius---clearly Ashoks hype agreed more with Elon's attitude than Karpathy's clear science focus and it's a shame that Elon couldn't keep Karpathy. (And Elon is jealous of Sam Altman) There's...

Some of us do machine learning for a living you know

Specifically what they mean by "direct photon" is that older versions of the nets used images that had been post processed through some standard image filtering/representation libraries before presenting to the net, but they got better performance (both computationally and in ML) by skipping...

High inference (on-board) power consumption is a big deal in an end-user owned EV and would significantly reduce driving efficiency and range. 1000+W (as I expect Waymo/Cruise to be) would That's actually something that does make sense, it means that the control latency is shorter than human...

And some time back the previous versions were a "kludgy mess" and V10 or V11 was going to be the awesome sauce

Technologically that can be solved by having some nets be segmented (whether fully or partially) or fine-tuned by region and other parts remain the same in an end to end training. Or, like the LLM fine-tuning, a low-rank overlay delta that's region specific on top of a generic base model. The...

Sure, or many other cases. Because our perception of the world "driving off a cliff" doesn't necessarily match the internal perceptions and intuitions of a big bag of grey neural goo whose representations we don't understand. Like it's never seen cliffs before or know what a concept of a...

For this problem, any novel architecture will take literally years of scut work to tweak and tune it. A big rewrite coming out now means they're pretty far away from a reliable system OMG they're going to need way more time. At a minimum there needs to be tons of examples of bad drivers and...

San Diego to Los Angeles round trip. 22 M3LR. 260+ physical miles. Two segments, one stop. Started 100%, ended about 13%, but I don't remember exactly. Average OC/LA freeway traffic. Very pleased with freeway efficiency, better than my old BMW i3 which had tiny skinny tires and a much...

The range loss from calendar degradation is highest the first year and slows down after that. Its 303 miles new. 285 is slightly higher degradation that I would expect (and there is some inaccuracy in the BMS) but not a sign of any obvious major defect. Model Y P is not so efficient, being a...

In USA there is no objective noise tests, but in Europe there are. The noise tests are outside the vehicle though, not inside. That said, everyone reports here on TMC note how the Hankook iON are very quiet. They'll probably be my next one. There haven't yet been reports on Bridgestone...

10 month vs 20 month: 4 mi 0 month vs 10 month: 9 mi Academic models approximate calendar degradation as proportional to sqrt(t) with t as age of cell. sqrt(2) = 1.41 sqrt(2) - sqrt(1) = 0.41 (20 montbs vs 10 months) 9 mi * 0.41 = 3.69. 4 mi degradation observed. Limited data here but...