25 cent per kWh intermittent ... that's not going to be profitable.

If AI instantly decides to reduce its visibility when it evolves, maybe we should do likewise?

I'm reminded of the workers with small rubber hammers doing final alignment of body panels in car factories, these fuselage panels will have a lot more deformation than that. Simply lining up the rivet holes probably takes a fair bit of judiciously applied force and humans are relatively cheap and very flexible robots to solve such problems.

How much force is too much? How much gap is too much? If he did the simulations to determine how much is too much his opinion are relevant, if he was just on the line and didn't think it was reasonable it doesn't necessarily mean much.

MG/BYD isn't really at some different level to European manufacturing here in Europe AFAICS.

Also it's the European manufacturers which are breaking open the quintessential European small car market. Dacia Spring Extreme was the first somewhat practical one even if underspecced, now there's Citroen e-C3, soon Fiat Panda-EV and at some point VW will join late to the party. Chinese are nowhere to be seen.

I don't dispute that M3 can have a very good value proposition. Even apples to apples (ie. same mm2, same process, same watt) it might even say still have say a couple 10% efficiency AND power advantage. I don't think so, but even if it does I don't consider it a huge deal, not a bigger deal than x86 legacy.

I just dislike when people pretend it's all about processor architecture when system integration, node advantage and ability to spend more mm2 (at lower clock) have such a massive impact. AMD chiplets are not power efficient, they are cost effective ... but that has little to do with x86. 7945HX systems are much worse than 7840U systems at power gating and much much worse than Apple systems, but again that has little to do with x86.

ISA is the least of the problems of (windows) PCs and the least of the advantages of Mac. Other factors are much more important for efficiency and other advantages.

"The M3 Max then, gets gets 307 points per watt, and between 10 and 20 more points total"

That's a third more efficient, not three times more efficient.

The efficiency is for wall power where the poor system power efficiency contributes and it's also two nodes behind. The advantage on video rundown is a nuke, a third difference is noise.

It's mostly an issue of power gating, both at CPU and system level. Even good 7840U systems have problems scaling to the lowest practical load which is to say offline video viewing. The moment you turn on wifi and actively browse, the load gets high enough for it to be competitive though.

For say a 7945HX system with a dGPU it's worse. If it's running Cinebench multi at say >50W it is actually high efficiency and high performance right up there with M3 Max ... but it can't scale system power down low enough for browsing.

They are high efficiency and high performance at higher load.

The best 7840u and 155H systems get close enough given the node disadvantage.

A computer/program is not a human to the law. You could argue they should be treated equally, but they won't be for the moment. Maybe when they are actually independent persons, but that will make a lot of things moot.

Batch 1 inference for LLM transformers is not easy, memory bandwidth requirement is very high if you do it the normal way. Batching inference for multiple users makes it easier in the cloud, but that doesn't work local.

Apple was the one that reinvented that RELU MLPs can be computed with only a predictable small percentage of weights (they should really have cited "Low-Rank Approximations for Conditional Feedforward Computation in Deep Neural Networks" in their paper). Apple will likely be the company to do something similar for attention in LLMs ... because no one else is even trying, almost the entire industry only researches architectures for large batch inference in the cloud.

Microsoft's close involvement with commercial model training is taking the risk of ending up on the wrong side of a supreme court judgement for 11+ figure fines. For what? Consumers don't care, it's not a sales driver.

Apple is far better off just buying cloud services until the supreme court says whether or not copying (of pirated content) for training is fair use.

It's not like ARM is revolutionary. Apple is just able to buy out the best node, blow money on massive die area and silicon interposers ... architecture wise Snapdragon X Elite is actually more impressive, but even then the impact of ISA is minimal.

With good system design, x86 is good enough.

Windows is an operating system to run x86/win32 applications ... and Microsoft manages to make it do that worse every year.

Because humans are weak. It creates perverse incentive and makes traders stupid.

The company can simply give performance bonuses.