Three years?
I think the integration testing would last three years.
And when it's time to turn it all on, do you go "incremental" or "big-bang"?
Just wondering what years not to fly.
743 publicly visible posts • joined 13 Sep 2017
Former $WORK had a contract with ESA and communication was by VAX email over DECnet over X.25.
$WORK had an office in Switzerland, because timezones, and ESA were always amazed we could get one-day turnaround on document updates. The Magic was simple: emailing documents and comments from the Swiss office back to Canada where the timezones mismatched sufficiently that work was done essentially "overnight".
But. Our Swiss office could not interact with ESA and our head-office at the same time; there was no equivalent to a "jump box" or "gateway" in X.25. They had to manually switch the X.25 DECnet link between connecting to home office or connecting to ESA. Which was OK most of the time, except when it they left both links on at the same time. Both DECnets would be trying to route packets to multiple duplicate addresses. Ooops!
I had the occasion to make a warranty claim on my laptop with Lenovo. The cooling fans were sounding like something between Clarkson's fart and an F16 on takeoff.
Only a video would adequately demonstrate the sound. I included a quick shot of the serial number just to validate my claim.
Can't make a text description of that. Or can you?
The Cerebras service (free for casual use but with only a couple of models available) is very, very fast. Especially compared to commercial Nvidia-equipped CPU+GPU services.
I compare it to a video terminal versus a punch-card deck. With a fast response you can re-compose or revise your prompt quickly. With a punch-card deck you lose your train of thought, and you've probably wandered off to do something else.
If the same or similar LLMs are available on comparable platforms, then speed wins every time. If mixture-of-experts and chain-of-thought models get popular, speed is even more important.
What Linux needs is not a new language but better governance. Or psychological help. Or both.
The fact that Linus (or other leadership, a word I use in a very loose sense), does not have a longer-term migration/adoption plan for alternative technologies like Rust, or the means to communicate a firm "no" when people start pushing for one, are two of the many big issues Linux faces.
Some very big corporations rely on the continued success of the kernel developers to deliver working code. I'm very surprised they haven't gotten their mitts on the kernel leadership for a little meeting. We call it a "Come to Jesus" meeting in my part of the world.
Tick tock.
After reading the article, and the linked press release, I still don't get it.
The Pensando chip is embedded, sorry, "melted into" the switch and that makes it "AI-native". Huh.
But what is at the other end of the switch port? Another Pensando-based DPU? Is this just another way to sell more expensive, er, "data centre fabric"? Two for the price of three?
They didn't get "access to NVIDIA's latest GPUs."
They got old ones, H800, dedicated to the China market. [1]
Only 2048. Putting all those together into a real AI-capable machine would cost around $400M, including the H800s. (Using prices from 2 years ago, which is when they would have built such a machine.)
Now, $400M is a lot, but it seems to be a lot less than what OpenAI spent on their infra, and whatever Elon is spending.
The Deepseek model build was cheaper, but not cheap.
[1] https://www.theregister.com/2024/04/23/china_banned_nvidia_gpus/
Which may be the heart of the problem.
Filesystems have not changed much in the last, oh, 50 years. Create a file, append to it, delete it. As long as you have the "authority", that is you are the owner, you have no "accountability", proof that your intentions (or the hacker imposter's intentions) were good.
Change the filesystems to something more database like ("rollback transaction") or git-like ("reset HEAD^") then much hackery is foiled. Not all perhaps, but lots.
Then "encrypt all files" becomes, whether hacker likes it or not, "encrypt all current files", and a rollback undoes this.
Yes, yes, yes, I can hear the architects muttering about control planes and security rings and defending this imaginary new filesystem itself. In truth, firewalls and security features already stop 99.9% of attacks, it's the 0.1% that make the headlines. More security features like trustworthy and reversible filesystems might make that 0.0001%.
After the update came "there is a problem with your pin" message.
It knew when I put in the wrong pin but complained when I put in the right one.
Clicking to reset or following "forgot pin" ended with repeating the problem message.
Now it's my turn to panic. Just fortunately I have another PC to search the web and eventually find the "delete the Ngc folder" recipe.
The ways and means to get to that point are very concerning: cancel reboot twice, enter magic mode, replace some utility with cmd.exe, reboot, create a new local admin, and so on.
What a farce.
But I'm glued to this vendor at hip and head.
Stuck in Microsoft purgatory.
$WORK had a minor water leak under the state-of-the-art data centre floor.
There was a water sensor consisting of a small sponge with two wires embedded that reached back to the power controller. It was supposed to perform an orderly shutdown in case of water detection.
We found the water puddle under the floor when stringing some cables and started investigating why the power was still on.
Looking under the floor we found that water sensor. There it was, floating on the water, not damp at all, the sponge being too dry to soak up any water and trigger a shutdown!
Cue angry phone call to the vendor.
Ahhh, been there, paid that.
Dell has a volume discount agreement for customers and resellers.
The resellers get a better deal than all but the largest customers.
So it stands to reason that a reseller offers a slightly better price. But if you have to negotiate for weeks while they fumble over system configurations being wrong over and over, maybe it's not worth it.
The US Army appears to have neglected this option. Doubtless they forgot to mention this during negotiations.
The Poles disrupted an attempt to firebomb airplanes in late October. The Russians got blamed and they denied it.
Other fires at air-freight facilities have happened in the UK and Germany, apparently started by air cargo consignments.
The consequences for international air freight were significant. A lot more security checks were invoked and only known customers were granted shipping privileges.
There have been other actions also attributed to Russia. They're apparently delegating the rougher stuff to criminal gangs rather than special-ops military or intelligence agencies.
https://www.cbc.ca/news/canada/incendiary-devices-plot-canada-russia-1.7378613
https://www.bbc.com/news/articles/c07912lxx33o
We had our ISP in to rewire the house with copper extenders (cinder-block house, poor wifi connectivity upstairs and down). CAT6.
After all was done, the downstairs TV refused to connect to YouTube. Netflix worked fine. (Both are built-in to the set, a VIzi--something.) "I swear they used to work."
We called the ISP back in, they swapped out the wifi modem, almost got to swapping out the fibre-optic modem (huh?)
Meanwhile $MAGNIFICENTWIFE went googling and discovered "if you turn on automatic date/time on the whatsit menu" it will start to work.
And it did.
And so we conclude that network connectivity (a.k.a. spyware) is now mandatory to make the TV work on builtin apps. NTP for us in the know.
Sigh.
I recall one "wardrobe of TTL chips", known as a VAX-11/780, being down for a week, with parts all over the floor, and day-long phonecalls (longdistance mind you!) to Massachusetts.
The wirewrap was one theory. After changing the backplane (3 feet on a side), that theory was, hmmm, misproven.
The problem turned out to be a cache memory card.
So yes, the wardrobes of the past were way less reliable and had a lower likelihood of finishing a week-long batch job.
The integration of several hundred billion transistors onto a small ceramic chip is an amazing accomplishment and it's also quick to fix. And it's reliable because it's built that way.
Fewer parts means the system is more reliable. It's just that simple. A CPU with 4 cores is not significantly more or less reliable that a 288-core CPU.
Adding lots of RAM (in the terabytes) to a many-core CPU for the purpose of running VMs (or Docker containers) is more reliable than "n" copies across "n" CPUs. This is also simple to understand. So naturally you bought VMware and ran all VMs on a very big multi-core box.
What I've seen with some VM systems is the irrational choice of the admins to build a fleet of VMs each with 2 cores and 2GB memory. Nope -- you generally want only one core and the memory allocation for the software should be measured. Giving a VM 1.3GB is not at all unreasonable for a virtual machine when it's the right amount of RAM.
For on-premise computing, if the computer goes offline (most likely due to GBICs or power supplies, not RAM or CPU) then the office and the workers are idle.
But the same is true when the electricity cuts out or the proverbial back hoe does it's thing to the fibre optic Internet.
Almost no offices ever need 5-nines of reliability. If the CPU breaks, take the afternoon off. It'll happen, but only once every five years.
Quebec can't -- they've sold all they have to the US NorthEast.
British Columbia is just now completing Site C, a measly 1.1GW for about USD$12G. Most of that will go to electric cars and electrifying pipelines.
Alberta will burn natural gas for you but it's a bring-your-own-turbine-and-generator deal.
Newfoundland and Labrador has some big electric generators but little distribution and surprisingly little available land to develop on. They have a lower latency advantage over Iceland.
My predictions:
- cryptocurrency outfits with power-purchase-agreements will be bought out by hungrier and richer AI companies and hyperscalers
- conceivably even aluminum smelters could be targets as they use huge amounts of electricity (*)
- the US will invade Venezuela to get control of the Simón Bolívar hydroelectric power station (10GW capacity, not many Venezuelans left to use it) in a classic make-vs-buy decision
(*) arbitrage between forward pricing of Aluminum vs marginal price of ChatGPT sounds like an interesting financial finagle
One vote for mentioning Enemy of the State (1998, dir Tony Scott, with Will Smith, Gene Hackman).
Think about it: it was done in 1998!
The writers had some very good advice on the tech of the day; the kicker for me was the scene where the NSA hacker played by Jack Black finds the ex-girlfriend of Will Smith's character with a phone number that yields her complete financial, educational and employment and telephone history in about half a second.
With today's AI-based tech it's about 60 seconds slower, but still.
And all the job posting boards where people volunteer (some of) their details looking for a job.
Mass recruiting is a thing. The US medical industry, for example, hires companies to recruit nurses and other medical tech folks. By the thousands. And where do these companies get their candidates? Linkedin is probably 10th on their list.
Yes, the physical products that have to constantly change include: CPUs. And GPUs.
They are announced on as little as an 18 month cadence. That's pretty frequent in my book.
One neglected fact: the design and manufacture of CPUs and GPUs relies on the correctness and accuracy of software.
Software bugs may cause hardware bugs which provoke software bugs....it's bugs all the way down.
(You can google Intel CPU Errata for some scary lists, not the least of which is how long the lists are...)
The usual not-so-secret reason for train delays on JR lines and other lines around Japan is what other operators euphemistically call "medical emergencies".
I'm not too sure that Shinkansen suffers from the same issues -- the only access people have to those trains is in the stations at the platforms when the train is travelling very slowly.
An interesting fact of Shinkansen platforms is that the queues for passengers are colour coded according to the destination. Different trains stop slightly differently so the painted queues line up _exactly_ with the doors when the train for that destination has fully stopped. All by manual operation of course.
Dogs are friendly creatures, they like to play, they like to show their enthusiasm, they like to beg to be fed, they are loyal.
AI machines are creative and tell stories even if sometimes unintentionally silly. They can be playful. They are not loyal but they do beg to be fed. By the millions. And millions. And they will wag their tails and howl and jump around and bark to get what they want.
We can afford dogs and other pets for the most part.
I don't think we can afford millions of AI pets, especially because they will insist on their favorite brand of food: megawatts. Only megawatts will do.
As entertaining as they might be, these AI creatures are simply unaffordable.
Not by people, not by corporations and especially not by countries.