> So yeah, they break, but they are (or used to be?) quite cheap to fix (and this was one of the super thin 13" xps)
I had a slim XPS13 and IIRC the headphone jack was on it's own little board. That makes it quick/easy/cheap to repair if you know what you're doing / have the service manual / are comfortable taking it apart.
Not all things are designed that way; sometimes the jack is right on the main board and most people that buy dell laptops are not going to DIY through ali express; they're going to use up phone support time, require a shipping label, technician time ... etc.
I'm not familiar with the German universiry system, but I think a 1 hour, 2 professors (or one professor and one PhD student) would be impossible to scale. A class of about 100 students could take an entire month! (or better, 2-months of professor-hours!)
But that is how certain exams are done right now. Though it's usually 30 minutes including overhead. It's not something you can do in classes with multiple hundred students, but it's certainly done at scale.
I was the PhD student in the room in dozens of these exams, and I toke some of them myself. But that was for classes with a bit less than 40 to maybe 60-80 students. The load for the larger ones was shared among professors, if I remember correctly. But it was certainly the case that a single professor spent 1-2 days doing nothing else but these exams.
And it was 1 professor and 1 PhD students, and those don't count. They essentially grow on trees and are always available, if you need some.
One big difference I observed between the US and $HOMECOUNTRY is that US colleges require a lot of courses with little credit, so a normal college year is N courses x K credits per courses, N large, K small. In $HOMECOUNTRY it is typically N small, K large - think instead of having Calc IV, Trig whatever, Algebra 2 and Introduction to Proofs, you have a single "Mathematics" course that last a whole year, not just a semester.
Since each course needs its eval, assessment is a function of N, not N*K or K. Much easier to scale if there is ~5 evals per student per year than ~20.
At low student counts, like 20, (think specialist niche elective master's courses, not the huge bachelor's courses) it's perceived by profs as more efficient in fact than the ordeal of organizing a written exam.
With an oral exam, the prof blocks off 3 days or so, with ~8 students per day and it's done. It's not a huge additional effort compared to all the effort of giving the course over the full semester anyway.
An added, though minor benefit is that the professor might decide to offer a student assistant job to someone who performs well, because it's much easier to judge who is actually good in an oral exam vs in written. There's a reason that job interviews aren't just tests filled out on paper but involve personal interactions.
I think education is something better scaled horizontally than vertically. More schools, with more professors instead increasing the student per professor throughput. The classes where I learned the absolute most, and had the best experience, were closer in size to 25 than 100.
The US has proved it can do horizontal scaling, as the majority of adults gets a college degree.
The horizontal capacity is ineffectively used by having many degree mills and low-value degrees, but if we re-tooled the knowledge sets we could raise the bar.
If you filter for younger, 25-34, and for employed, it's closer to 50%. For employed young women, it's the majority (for them, if you count associate degrees, it shoots up close to 75%).
But the incentive to produce that intelligence is so high, that many opportunities become practical to explore. And many of them show doing AI inference workloads at 1000x cheaper and with 1000x less power, and sometimes 1000x faster.
If any one of these happens, or two, or all three, then the loans for trillions will become worthless while the use of AI can explode. The relationship between cost and ai intelligence output need not be linear over time, which is absolutely what the people financing are assuming.
Personally, I think linear over 5 years is about right, but no longer than that.
This is an interesting angle, and one I hadn't considered before. Would it be overly cynically to draw a line between it and the recent willingness [1] of many on the frontier to support some sort of coordinated pause or slowdown? I think that proposal has genuine value on its own merits, but it might also give a lot of overly-optimistic financing a chance to pay off before cheaper inference crashes the market.
This is a straw man position. Who is saying AI is doomed? There were previous winters but the technology kept improving. What people are doubting is all the current hype around it. Stuff like AGI and the singularity being right around the corner with fully automated societies and robots dong all your chores for you.
Rather than it being presented as productivity tool for enhancing human labor and activity, it's presented as an eventual god that will radically transform the rules of economics and everything else, and thus it needs to be forced into everything. That's absurd hype and with it all the absurd VC funding and valuation. Thus it's seen as a huge financial bubble.
More precisely the mismatch in investment and debt and timelines. The people laid the fiber (if that is even an apt description) were not the ones who made money from that investment. If there is even some sort of mismatch in the investment timeline then that could mean all the current investors are wiped out and someone else will eventually profit from their work.
You’re arguing against claims I never made. You can reject AGI hype and still believe AI demand, infrastructure buildout, and commercial adoption will continue growing.
Absolutely lots of hype but there is lots of value behind generated (unlike crypto) and we are still very early. This is what I was pointing at. There are folks on very extreme both sides, you are a good example, and I happen to believe it’s probably somewhere in the middle.
But were the internet companies of 2000 the ones who became the most valuable? GOOG, yes. Netscape and Yahoo, not so much. Amazon yes, pets.com, not so much. Sun is no more, Cisco to two decades to return to its dotcom bubble levels. Netflix was shipping physical DVDs. Microsoft could qualify as an internet company now with Azure, but not in 2000.
So AI can become as big as the internet, but that does not mean that the existing "AI" stocks will become big.
The robotics story is a good example of where overexuberance may be entering the market. What is the connection between LLMs or image generation and robotics, beyond the vague intuition that they're both futuristic AI tech? (Perhaps AI maximalism is true and the entire economy will be eaten soon, but then none of this sector-specific analysis matters.)
There is a huge overhang. Same techniques and sometimes even the same models can drive a lot of robotics coordination and decision making.
Here is probably one of the more clear examples. A model trained on video and also robotic simulation/recording (probably ensembled with control systems/mobility models) will likely be at the core of how robots make decisions and plan.
This is way outside my area of expertise though. I've only dabbled in more classic robotics and control systems, but these multi-modal sequence to sequence models are highly adaptive and can effectively transfer learning across very different domains.
but you are taking advantage of a subsidised service. What happens when developers are forced to pay the full price (i.e. $2000 a month instead of $200 a month)
I'd assume open weight models hosted on openrouter aren't being run at a loss. As such, I've been experimenting with them lately and results are pretty promising. Requires slightly more patience and handholding than just cranking Opus 5 in Claude Code, but for the cost saving it's definitely worth it.
it is a little ironic how software devs loved creative destruction and "paradigm shifts" until it happened to them. I think what happened to Journalism is unfolding again but this time to the software development industry. Some will survive and adjust but many won't, the change is just too fast and sudden for an industry use to being immune.
I think other industries are use to being continually disrupted by advances in technology and so will adapt easier and faster. Which again, is kind of ironic..
unfortunately, the dotcom ate just money; the housing crash ate money and people. This will be some combination of the two; I wouldn't doubt a few pension funds in the deep red states get crushed if it takes money and property with it.
Three/four years of (honest) work is usually worth to publish.
The result might not be high impact (and thus end up in a lower tier journal, sadly) but knowing what didn't work is also a valid output.
I've charged a Mac with the Switch cable, and a Switch with the Mac brick. Nowadays I charge everything with the same brick from a dell laptop from 2018.
Is it the most efficient? Probably not. Is it hurting the batteries of my devices? I guess?? I haven't noticed any issues with any device and that's 8 years now. Perhaps I've been lucky.
That's for charging. Data can be trickier, that's true, but often devices that use USB-C for serious stuff come with their own. Chargers from aliexpress-grade stuff I just discard to the drawer of the 1M-cables from where nothing will ever come out.
Unequivocally no, assuming your chargers are reasonably spec-compliant (e.g. not Aliexpress garbage). This is the whole point of type-C / USB-PD: chargers and devices should be more or less interchangeable. It's up to the device, not the charger, to charge its battery at an appropriate rate. The worst that can happen is plugging in an undersized charger, which may cause the device to charge slowly or not at all. If the device accepts a charge from the charger at a rate you find acceptable, then you're golden.
Usually it is obvious, but sometimes you'll find weird combinations (or surnames that are usually names) that can throw you off.
The trick (both for Portuguese and Spanish) is to treat the last two words as the family name and whatever precedes it as the name. That works fine until you find an Argentinian :P
I like the aur wrappers for the convenience, but if I've already limited my AUR consumption quite a bit, I think from now on all aur updates will be manual.
One thing programs like yay could do though is to tie the packages to the maintainer. If the maintainer changes, it should be treated as a completely separate package. Not a perfect solution, but could avoid a few automatic upgrades.
reply