cross-posted from: https://hexbear.net/post/8768696

I visited the largest electric truck charging site I have ever seen, located in Beichuan, Sichuan, around 150 km north of Chengdu, China.

The site is planned for 100 MW scale. In stage one, around 51 MW is already installed from 72 Huawei 720 kW power units. I counted 108 charging spots with dual dispensers, meaning 216 plugs, plus 18 megawatt charging positions. Stage two is planned to add 33 more 1.44 MW charging systems, bringing the site close to 100 MW.

This information is 7 days old but I didn't see any info posted here or in c/libre about it, so here we are.

Also, I don't know who Vincent Lapierre is, so can someone here more in the know school us on this person?

It's a Chinese Wi-Fi module for a PC! With Linux drivers!

There’s a really interesting quirk in modern architecture that a lot of people have been noticing lately referred to as the Curse of Depth in the paper. Basically if you look at popular models like Llama or Qwen or DeepSeek you will find that the deeper layers are surprisingly useless. You can completely prune away huge chunks of the later transformer blocks without actually hurting the performance of the model. The representations in these deep layers end up looking practically identical to each other, and it’s a massive waste of GPU hours because we are training billions of parameters that end up doing almost nothing.

The authors trace the root cause directly to Pre-Layer Normalization. Pre-LN makes training massive transformers way more stable than the old Post-LN setups, but the catch is that as you pass data through more and more Pre-LN layers the output variance explodes exponentially. Because of how the math works out this exploding variance forces the derivatives in deep blocks to essentially become an identity matrix turning the layer into a pass-through filter that cannot learn any meaningful new transformations.

And turns out that the problem can be fixed using a remarkably simple tweak called Layer Norm Scaling. They literally just scale the output of the layer norm inversely by the square root of the layer depth. This completely stops the variance from blowing up as you go deeper into the network. Because the variance stays under control the deep layers actually wake up and start contributing to the representation learning.

They tested this trick on models ranging from tiny 130M parameter setups all the way to 7B parameter models. In every case Layer Norm Scaling beat out standard Pre-LN and other normalization tricks. The pre-training loss drops significantly and those gains carry right over into supervised fine-tuning tasks. Best of all it requires zero new hyperparameters or learnable weights. It is just a clean mathematical fix to a fundamental architectural flaw.

They're really getting desperate.

https://archive.ph/uMSep

The anger and anxiety over electricity in America is palpable. You can see it in packed utility commission hearings, in protests against companies, and in furious reactions on social media.

And you can see it in the polling. Across poll after poll, more people are saying that they can’t afford their bills and they think utilities need to change how they make money. And they are also very cynical about data centers.

So will this be the push utilities need to finally change the way utilities pay for infrastructure?

This week, we dig into three indicators. First: 75% of Americans say their home energy costs have gone up, and a quarter of Americans now consider utility bills unaffordable. Second: 86% of California voters said executive pay should be tied to affordability. And third: 71% of Americans would now oppose a data center being built near their home, a 49-point swing in less than a year.

A Colorado State University researcher didn't plan to study rooftop agrivoltaics. She was furious when solar panels appeared over her research plots without warning. What her data showed next changed the direction of her career — and quietly launched a new field of urban agriculture.

Denver alone has an estimated 5,000 acres of flat rooftop space sitting largely unused above a city that needs both clean energy and local food.

cross-posted from: https://hexbear.net/post/8761674

Two Bit da Vinci's Ricky Roy partners with Elliot Richards (Everything Electric, Fully Charged) — who's lived in China for 18 years — to lay out exactly how China quietly built the world's most powerful auto industry, why every legacy brand is bleeding share in China, what their export wave is doing to Europe and Russia, why the US tariff wall is more of a stall than a strategy, and what it all means for Detroit and the American consumer.

Western drug companies are pouring tens of billions of dollars into China, to develop medicines that can be resold at premium prices in Europe and the United States.

China dominates all phases of the pharmaceutical industry, from new drug development, testing, and mass production.