The model is explicitly trained to produce as uniform a distribution as possible, because it's designed for batched inference with a batch size much larger than the expert count, so that all experts are constantly activated and latency is determined by the highest-loaded expert, so you want to distribute the load evenly to maximize utilization.
Prompt ingestion is still fairly similar to that setting, so you can first compute the expert routing for all tokens, load the first set of expert weights and process only those tokens that selected the first expert, then load the second expert and so on.
But if you want to optimize for single-stream token generation, you need a completely different model design. E.g. PowerInfer's SmallThinker moved expert routing to a previous layer, so that the expert weights can be prefetched asynchronously while another layer is still executing: https://arxiv.org/abs/2507.20984
BYD is selling a lot of cars, but they're also making a lot more cars than they can sell at sticker price in China, as does every other Chinese car company. This oversupply leads to all kinds of distortions, like dealerships registering cars as "sold" to make their sales targets and then selling those brand-new cars as "used" at a discount. https://www.reuters.com/business/autos-transportation/local-...
Maybe your YouTube echo chamber additionally thinks that this will cause BYD to collapse, but I doubt that. There are about a hundred Chinese EV manufacturers in worse financial shape, who're likely to go bankrupt first, which should reduce oversupply enough for BYD to survive.
There's not really NEV oversupply, the EV adoption curve will keep shooting up from 50% to 100%, many PRC EV companies will die from competition/pricewar consolidation not broader oversupply, since EV industrial base is still undersupply relative to 100% adoption curve.
Oversupply is in legacy ICE displacement due to rapid domestic EV penetration. "Zero-milage used car" accounting trick is primarily to export excess capacity of gasoline cars (now that EV has taken over) that aren't moving domestically anymore. MOST of PRC exports are ICE, IIRC 60-80%, there's plenty of global demand for ICE still. Pushing domestic sku new car with crushed domestic demand as "used" exports where there's plenty of demand = meet sales target, but less through discounts but import fees engineering - used cars circumvent import duties, certifications, warranty requirements etc. It's a lifehack to unload domestic ICE inventory, not EV. This also likely transient effect because NEV transition in PRC happened so fast ICE manufacturer that target domestic market caught flat footed. They need a few years to either retool to EV or shift primarily to target export markets that still has appetite for affordable gasoline cars.
The Reuters article focuses on exports because that's what their target audience is likely to be interested in, but the People's Daily article they reference in passing https://paper.people.com.cn/rmrb/pc/content/202506/10/conten... is treating it as a domestic problem, highlighting the appearance of "zero-mileage used cars" on the second-hand market, specifically calls out EV companies for putting quantity over quality, and floats stimulating used car exports as part of the solution.
Also, at least the EU doesn't distinguish between new and old cars when importing from outside the EU https://taxation-customs.ec.europa.eu/buying-and-selling-car... presumably precisely to avoid creating the kind of loophole you imagine exporters are exploiting. But they aren't; they just hurt their own profitability to juice their numbers and chase growth.
Domestic problem as in domestic involution problem not a EV overcapacity problem. I'm pointing out important to distinguish between ICE and EV, ICE is structural overcapacity, EV is under capacity but involution. ICE zero mile shenanigans is to liquidate/ unload excess inventory that doesn't have demand in PRC, mostly targeting loopholes in russia, central asia, mena. There's (less regulates) markets other than EU.
BYD/EV zero mile hack is more for subsidy harvesting, they're trying to dip on domestic subsidies for new sales then skip on stuff like servicing/warranty because used sale, i.e. shady involution behavior that abuse national policies, hence crackdown. It's not because EV are excess capacity, but so many EV players involuting and accountants find creative ways to carve margin. It's not EV sector makes 100 cars can only sell 50, because overcapacity, so discount to sell 100. It's EV sector makes 100 cars, sell 100 @1% margin but with some engineering they sell 100 @2% margin.
Yes, there are some brands that does use it for price cuts, premium brands, i.e. nio/xpeng, who won't lower sticker price, so they have to flip to lower price in used market. But that's also less overcapacity, more PRC EV moving so fast last few years no one wants 12 month old models. AKA tech stack progressing so fast / involution pressures so hard consumers won't pay sticker for anything older than 1 gen. What happens PRC manufacturing turns car cycles into phones obsolescence cycles. The broader the point here is EV makers are generally not doing accounting tricks to sell more cars due overcapacity / lack of demand (what over capacity implies), they're doing accounting tricks to sell the same amount of cars to make their razer thin involution margins slightly less razer thin, at the expense of the government.
Right. I can imagine seeing lots full of unsold cars might be interpreted as "there is no demand", especially if they are trying to push a particular narrative.
GPUs are more efficient than CPUs for LLM inference, using less energy per token and being cheaper overall. Yes, a single data center GPU draws a lot of power and costs a fortune, but it can also serve a lot more people in the time your CPU or consumer GPU needs to respond to a single prompt.
For a person finding bugs for a living, an up-front fee to have their report reviewed by a maintainer would amount to an investment towards receiving a bug bounty if their report is valid and valuable. Just the cost of doing business.
It would discourage drive-by reports by people who just happened to notice a bug and want to let the maintainers know, but I think for a project that's high-profile enough to be flooded by bogus bug reports, bugs that random users just happen to notice will probably also get found by professional bug hunters at some point.
Only if the system is fair. If I as a maintainer want to scam I can just close the report as invalid, collect the $$$. Then a week latter I fix the issue with a commit that looks like it is unrelated.
I wouldn't do the above, but it is easy to see how I could run that scam.
You can look at how the maintainer dealt with previous bug reports to decide whether you can trust them or not. If there haven't been any previous bug reports but they nonetheless ask for a fee to help deal with the large volume of bug reports, yeah, that might be a scam. If you're running their software, maybe also check whether it's full of malware.
Good scammers pay out once in a while. Casinos hate it when everbody body loses, and love it when one person wins big - they need just enough losers to make money while enough winners to show off.
The typical graduation-requirement paper doesn't get published in a professional journal, so I think professional journals do provide significant curation.
Competition reduces profits, which is good for consumers, but bad for investors. China has a bit of a problem with hype-driven investment, where something becomes a new trend (solar energy, bike sharing, electric cars...), lots of new companies get started, followed by oversupply, followed by price wars, followed by bankruptcies. Of course every company hopes that it will be the others that go bankrupt, so during the price-war phase, they take on a lot of debt to keep operating just a bit longer. In the worst case, there could be a wave of bankruptcies developing into a full-blown economic crisis.
Nuclear fuel reprocessing plants release iodine into the atmosphere. The prevailing wind direction at European latitudes is west to east, so when the iodine comes back down in rain, it'll likely do so over Russia (emptying into the Arctic Ocean), Central Asia (emptying into the Caspian Sea and various lakes) or Northeast Asia (emptying into the nortwestern Pacific Ocean, including the Yellow Sea.) This paper has an illustration: https://www.sciencedirect.com/science/article/abs/pii/S03043...
Identifiers correspond to nodes and a mention of an identifier in the definition of another corresponds to a directed edge. The resulting graph won't necessarily be acyclic, but you can still use it to inform the order in which you present definitions, e.g. newspaper style starts with the most high-level function and puts the low-level details at the end: https://pypi.org/project/flake8-newspaper-style/
reply