2,317 karma · joined August 2, 2012
- Larger MoE models only need compute proportional to the number of active parameter, so Qwen3.8-Flash-Next (180b with 6b active) runs more tok/sec than Qwen3.8-27b on the same hardware.
- Larger models typically have higher quality output and require fewer thinking tokens to get to an answer.
Consequently, measuring time to complete a task with a wall clock, the speed difference is less than you would expect.
MoE models want a different ratio of memory/compute/$ than dense models like a Gemma4 31B or Qwen3.8 27B.
Since only some of the experts are active each round, you don’t need as much compute to get good tok/sec, but you do need a lot of memory to avoid streaming weights from SSD or system RAM.
You could run a full Muse Spark on one of these for $24k at probably around 50 tok/sec.
You’d need 5x RTX 6000 Pros at $80k for Muse Spark (albeit at much higher tok/sec).
On a $24k budget, you’re looking at a maximum of 96GB from Nvidia or 384GB from Intel or AMD. So you’re bottlenecked by PCIe bandwidth, and all those TFLOPS go to waste.
Other configurations of the Studio have similar logic for other MoE models. For instance, Qwen3.8 flash-next runs at a very acceptable 75 tok/sec on a 128GB M5 Max for about 1/3rd the price of an RTX 6000 or 40% less than 4x B70’s.
Even using the sale prices, an iMac is still cheaper than a Mini with the Kuycon.
My absolute favorite feature of Apple displays is that I don't need to have the brightness cranked to max all the time, even when sitting next to a window. Some of them are, dare I say it, too bright for a normal office at maximum brightness (yet still dim for comfort in a pitch black room).
There are some good alternatives if your budget stretches to $1000, though: Samsung's 27" ViewFinity S9 comes to mind, with 2880p and 600 nits.
Every enrichment process drops in efficiency as the enrichment increases, asymptotically approaching that process's enrichment ceiling.
That's why nobody uses thermal diffusion to enrich uranium: It's by far the easiest method (you can do it with 1800's technology), but the practical maximum enrichment fraction is too low for anything useful (other than providing feedstock to better enrichment processes).
That's also the reason a centrifuge enrichment plant uses warehouses full of cascaded centrifuges: The centripetal force inside the centrifuge is working against a diffusion gradient, so each stage only gives a small percentage decrease of the "derichment" fraction.
Once you pass 50%, you stop looking for U-235 needles in a haystack to keep and start looking for U-238 needles in a haystack to throw away.
If a company pays you $1M of shares, the IRS treats it exactly the same as if you were paid $1M of cash and you used it to buy the shares. The capital gains work the same too: you pay capital gains tax on the difference in value between when you receive the shares and when you sell them. That’s how it already works.
When a billionaire moves to Wyoming, the marginal tax revenue is far greater than the marginal cost of police services, for example.
There’s some manifold of Pareto-optimal solutions to the problem of taxes and services.
California, for example, spends a lot of money to get not-great public schools, and has housing costs that are out of control. So if you’re importing new hires from other states, and pay them a lot to live here, and personally pay high taxes on top of that… that doesn’t seem like a Pareto-optimal point.
At some point, the wealthy will decide to move someplace that looks more Pareto-optimal. Better externalities, lower taxes, or both.
It is a fiscal detriment to CA and NY though.
The problem is a lot of AirBnB hosts want to offer the hostel experience for hotel prices.
It has long been obvious that many scammy hosts exist on the platform, and it ought to be easy to detect such hosts if they bother to try. Yet, that never happened until now.
So what changed? I would bet that until now they have always worried about having enough inventory to sell, and that now they are noticing they have too much inventory. That changes the dynamics from "be extra nice to those who bring the inventory, and even tolerate a little fraud from them" to "be extra nice to those bringing the dollars".
I do wonder if this is an early sign of a coming recession.
In other parts of this thread we have people saying that the AfD must be banned because they are a threat to democracy and that the German constitution has provisions which prevent amending it to weaken the guarantee of democracy...
But how do you then reconcile the idea that people must be prevented from voting for the party they wish? Is that not undermining democracy when you only let them vote when they vote "correctly"?
Splitting audio into bands and voiced/unvoiced content is basically how many "modern" low-bandwidth voice coding techniques like LPC, MELP, or CLEP work (see https://en.wikipedia.org/wiki/Linear_predictive_coding). The point of that processing is to compress, not encipher, the voice data.
Analog scramblers, like A-3, by contrast, shift the frequency content of the speech as the means of encipherment.
The point of the SIGSALLY voice coder is simply to keep the amount of key material for the one time pad reasonable. Even with that compression, each OTP record lasted only 12 minutes. MELP, which is a current DoD vocoder standard for use in secure voice channels, produces intelligible speech rates as low as 600 bps.
The part that is definitely a kluge is the key synchronization method. It seems that they did not use a PLL (which had already been invented by that point) to keep both ends in sync and instead relied on highly precise frequency references and turntables driven by synchronous motors with hundreds of poles.
That creates knock-on effects throughout the healthcare system:
* Doctor time is worth $xxx/hour, so every second counts. We hire a bunch of support staff to make sure doctors aren't "wasting" time on paperwork, taking vitals or talking to patients. Now it costs more to run a practice.
* Doctor time is worth $xxx/hour, so we need to make sure that health insurance only pays for procedures that people truly need. That leads to insurance companies hiring a bunch of doctors and nurses to adjudicate claims (that's expensive, plus fewer doctors practice medicine). It also leads to denied care.
* Doctors have a lot of money, so if a procedure goes bad and you want to sue someone, they make a juicy target. Now you have higher malpractice insurance, so doctors bill even more.
* Doctors are afraid of being sued (see previous), so they often order more tests than necessary to prove they are providing adequate care. Those tests cost money.
The cause of high doctor salaries is an artificial shortage of doctors. Medical schools collude to limit the number of medical students who matriculate each year, which squeezes supply.
We could very easily fix the American healthcare system in about 10 years by simply allowing doctors and nurses trained in other developed countries to practice here, plus forcing medical schools to accept more students.
Regarding the latter, there is no shortage of qualified applicants:
1. Nearly everyone who starts medical school finishes. Matriculants are so statistically overqualified that hardly anyone washes out.
2. Medical schools have silly-stupid entrance requirements which have little to do with being a good doctor. Beyond GPA, you need many hours of volunteer work outside of class, extracurricular activities, and often research experience. These admissions requirements are imposed because there are simply too many qualified applicants for the number of slots and med schools need some "fair" way to cull the herd.
His software company, named Green Hills, supplies the RTOS for Mobileye, which formerly supplied driver-assistance hardware and software to Tesla (and continues to do so for other car companies). Mobileye is no longer a Tesla supplier because Tesla insourced their ADAS.
So he has a very real financial incentive to convince the world that Tesla's system is bad.
With a modern 4 GHz CPU, that means your budget is only 4000 clocks per interrupt. The context save alone can make it hard to hit 1MHz, and you really want to have your interrupt code pinned to the L1 cache because a few memory reads will blow your entire budget.
As nice as USB4 is, I’ve never heard anyone claim it has single-digit microsecond latency, to say nothing of sub-microsecond latency. The typically-quoted ~20 usec latency of USB4 limits it to around 50 kHz for an emulated ISA card.
If you restrict the domain to nighttime or exoatmospheric, a high-resolution wide-angle camera paired with an offline astrometry database has the makings of a strapdown astronav.
Widening the domain again, the natural motion of the vehicle is probably enough, when combined with inertial sensors to aid pixel-shift image stabilization, to achieve the same kind of processing gain as the older systems got with shutter-synchronous detection.
Literally 5 consecutive minutes of making 100s of wrong decisions snatching defeat from the jaws of victory.
According to news reports, a magistrate judge at OSHA has been sandbagging an investigation into retaliatory firings by this airline (of pilots who reported unsafe practices with maintenance, CRM, training, scheduling, everything). I wonder if the NTSB will include that judge in the investigation. It seems like he or she has a lot to answer for.
https://www.cbsnews.com/news/miami-plane-crash-21-air-safety...
In almost every state, you may post a bond in lieu of insurance, for an amount as low as $35,000!
You don’t even pay for the fraudulent credit card transactions in the first place! There’s no money for them to return!