Recent content by DrChaos

Agree. I don't know the full algorithm, but when the lead acid is spontaneously losing charge the main battery will come on more often to charge it. But when the SOC of the main battery is low it probably stops trying to top off the 12V, and now the 12V runs down quick and died. Or the power...

Potentially but distributed parked computers don't have that much value. The cases it could handle are limited---and useless unless there is at least high reliable download bandwith. The tasks have to be trivially paralellizable. The value would be large scale scoring/evaluation of models...

High latency low reliability network connections aren't so useful for very large scale training. There's a good reason the highly intimately linked supercomputer NVidia boards binding coprocessors very tightly and other connections binding boards closely are preferred and most useful. This...

He said that AI will not learn spatial reasoning purely from text and he is correct. The LLMs can sort of guess at some proxy to spatial reasoning when expressed through text on simple examples---and the GPT-4 system is opaque how it is trained but it is not only on text. LeCun does not...

tire compound, tread design and internal construction (and shape, i.e. wheel diameter a bit) influences rolling resistance. Amount of open area on tire + wheel strongly influences aerodynamics. Tires are always fully closed so going to smaller wheels improves aero unless your wheel is flat...

unless there's some sort of emergent property breakthrough (and that would be on training data size first) I conjecture first order scaling is going to be logarithmic in compute performance: like log(mistake_rate) = A - B*log log flops And you need the clean dataset size and model memory size...

Worried this might be like Elon doubling down on the Cybertruck, he's going to push on what he personally thinks is Cool but not appealing to the masses.

https://www.hankooktire.com/us/en/tire/ventus/s1as-h125.html Material Code 10300060, 98V is the Tesla tire with foam and 8.5 tread depth. Material Code 1028523 98W is the non-foam generic tire with 10 tread depth.

Which specific tire model? Hankook Ventus S1 A/S? There are indeed two distinct Hankook part numbers for 235/45R18. Only one is Tesla OEM with the foam---and lower tread depth for efficiency. So that can't be foam that is added after manufacturing.

My theory: it's trained heavily on Model 3 data (most numerous) and doesn't adjust for car model. People who complain about hitting curbs seems to be Model Y, S and X, which are wider. It's taking the centerline as learned from human 3 drivers. It's been fine for me, but I have a 3.

Not good for taking luggage inside. Those upswing doors are expensive and prone to abuse damage and will hit a neighboring car or pole. And why low profile tires? Why not 16" with big fat rubber like in the old days? A traditional London Taxi is near the optimal configuration in the...

I believe these are the same as the iON EVO---just the earlier name. I think you got an earlier manufactured tire at a giant discount.

I'm disappointed the summer iON have no treadlife warranty at all. Even if were 35k or 40k vs 50k I would consider it. I drive in southern California and have no need for any snow or cold weather performance ever. That the summers have no treadlife warranty is Hankook signaling to me that...

It's never a numerical issue. The issue is the type of people applying for those. History shows by far the best don't work at an environment with aggressive nasty leaders. They work at Bell Labs type or Los Alamos Labs with real scientists running the place and insulating the creators from...

His personality and management is dissuading the high end ML scientists who have options. Who would choose Tesla over OpenAI, Anthropic, FAIR, Microsoft or Deep Mind? Particularly with Musk's ill informed hot takes and aggressive firing and chaotic demands. With musk running things more...