@p
Gdi now I actually have to eat my own dogfood
@mangeurdenuage

@JoshuaSlocum @charlie_root @p @mangeurdenuage @mischievoustomato Strong disagree.

"A bunch of servers" has no connotation of redundancy, fail-over, orchestration, a whole bunch of things we expect from "cloud vendors." See Amazon's S3, the first "cloud" offering as I remember. You put your data on it, with a limited full object API, they replicate it across multiple availability zones (miles apart, different flood plains as far as I know), it takes care of a lot of issues with persistent storage.

Your case is much stronger for something like Grok, but it's presumably built on top of many of the same techniques. Almost certainly not multiple sites yet, but that's completely impractical due to the costs and limited availability of GPUs at the moment. Xwitter by comparison has two datacenters far apart, each able to handle the whole load of the service(s).

@p @charlie_root @JoshuaSlocum @mangeurdenuage @mischievoustomato "I can't hear you over the sound of US-EAST-CHAOS-MONKEY downtime."

FIFY. And last time I checked, it controlled some global AWS stuff....

But the rest of your rant is misplaced, it's all, has for a very long time been a set of trade offs. Using what makes sense for you has never been bad advice, and "clouds" like AWS have some good use cases, like highly variable loads, one off jobs, low upfront capital costs, and as mentioned storage.

Heck, "serverless" directly invokes https://en.wikipedia.org/wiki/Conway's_law (not my observation, BTW).

"[O]rganizations which design systems (in the broad sense used here) are constrained to produce designs which are copies of the communication structures of these organizations."

@JoshuaSlocum @charlie_root @p @mangeurdenuage @mischievoustomato "Embrace the healing power of 'and'."

A "cloud" has both real technical meaning as I've outlined and inevitably became a marketing term with limited to no meaning as you point out. I expect my interlocutors to understand this sort of thing, and @p for one doesn't disappoint.

(Check out his also tutorial https://blog.freespeechextremist.com/blog/fse-vs-fbi.html and https://blog.freespeechextremist.com/blog/about-fedilist.html as I'm reading as time allows. Although we are talking past each other, but not because I didn't read the thread, I just don't agree with some points made in it. Even the "who owns it is not a major difference" one he just emphasized.)

And here you invoke an also old maxim about data storage, one copy means zero, two copies means one...." Some of us were lucky enough to learn this when we discovered the -rf flags to rm (UNIX™ remove files, with the recursive and force flags, the latter meaning just do it, don't complain). Normies ... are they any worse than us until they too get burned??

@p @charlie_root @JoshuaSlocum @mangeurdenuage @mischievoustomato "> US-EAST-CHAOS-MONKEY

'We meant to take down half the fuckin' internets.'"

The purpose of a system is what it does!

Really, I think we're two old souls talking past each other, although I might be older, Fred Brooks' invaluable The Mythical Man Month was published late enough in 1975 it wasn't in the first set of software engineering books I read in high school at my local low tier college (and check the second edition for his later reconsideration of maxims like "plan to throw one away because you will" (your first attempt at writing a system)).

AKA I don't think either of us is really wrong, we're just emphasizing different things (and I can't say I remember all of what I read in the thread). Like for Grok I expect less censorship, albeit neither it nor "Open"AI managed to be as bad as Google:

@p @charlie_root @JoshuaSlocum @mangeurdenuage @mischievoustomato "> I just don't agree with some points made in it. Even the "who owns it is not a major difference" one he just emphasized.

"A landlord is a landlord; it's not the same as owning your land. This is a major difference."

Unfortunately for the Internet I accept the principle "You will own nothing and you will be happy." IP addresses? Only at the sufferance of your provider(s) and those like governments who can choke its throat(s). Ditto domain names, where that's quite explicit as I understand it (but that's somewhere the law may evolve, to give it a property right). And you pretty much have to be big or graf to not put your site behind someone's CDN, right? I mean, when Facebook for example will in error hammer your site for some photo optimization on their side and not give a fuck what else can you do? See also DoDS.

So if we get targeted, we find we've built our stuff on a foundation of sand, although with enough effort the industrious have been able to find new homes, for example Stormfron (white nationalist 1.0 forum, "Damn this website has absolutely no useful updates on hurricane Florence but i'm pretty sure who's responsible for it now"), Gab, The Daily Stormer, and lately KiwiFarms. But plenty died, like Parler; if I remember correctly, AWS was frantically and repeatedly asking them if they'd give Trump a platform after J6, and they were turned off on January 10th, 2021 (the revival wasn't real from everything I've read).

So it's "landlords all the way down," kinda like real property were in the US you also answer to your county at minimum on up. Except with infinitely fewer protections, from social/political to legal. But I'd agree the higher up you go in the stack the more capricious they are, and at the "AI" level that's very high.

"Hope [the FBI etc. articles] were entertaining or informative."

Essentially both, thanks!

As for "rm is forever," if your friend is around your age, I'm probably older, MIT had no serious UNIX systems at all when started there (by that I mean with split-I and D PDP-11s like the /44, /45 and /70, the point where there's just enough address space to be useful as I saw it). Multics was killed by Honeywell, UNIX™ became big, such is life.

@p @charlie_root @sicp @JoshuaSlocum @mangeurdenuage @mischievoustomato CTSS was one of the very first time sharing systems, built on an IBM 709, the third model of IBM's binary "scientific" vacuum tube computers which was announced in 1957. It was easily ported to its transistorized 70xx follow-ons,

These were 36 bit machines that could address 32K words, or 128K which is sort of equal to a split I and D PDP-11, and it could only swap, not page (MIT paid the extra money for a second bank of memory for the OS). (The U.K. Atlas first released in 1962 is credited with introducing paging to the world.)

Multics was the second system syndrome follow-on to CTSS, where a single segment of memory and the max file size was a whopping 1MB, Also 36 bit CPU, a tradition because that neatly held ten decimal digits like electromechanical calculators, IBM's 701 was originally named the Defense Calculator. But you could use many segments at once, and many would be shared in DLLs and hard core (kernel).

The AI Lab just had wildly different requirements, like hooking up a robot arm and for a demo being able to grab all the resources of a machine, very much not a polite thing to do for a time sharing service.

IBM's dogma against dynamic address translation virtual memory for the System/360 which lasted all the way through the first generation of System/370 built with ICs cost it every high end CS university's business, thus Multics started with GE, and the AI Lab bought a PDP-6, for which DEC consulted them about the instruction set, and the eventually fatal 18 bits of 36 bit words of address space, or 1MB.

(And if you read between the lines of Brook's, this plus inflexible management in module memory allocations was the key technological cause of the OS/360 five millennia nightmare for both IBM and the poor sods who had to use a completely fucked API due to this incidental complexity.

Their multiple OS nightmares and ensuring few university students would use their production shit (VM/CMS doesn't count, much???) helped make them a mind hare and eventually market afterthought in the market. DEC avoided this until Ken Olsen got too old.).

Unfortunately the PDP-6 had catastrophic industrial engineering, a recurring nightmare though at least 1990 and the VAX 9000s, which managed to sell one whole dozen more systems than the PDP-6, four vs. three! :cirnoheh:

But it got the AI Lab ready for PDP-10s, first the KA10 model with a physically easily hackable hardware architecture of Flip-Chip modules plugged into huge wire wrap backplanes.

DEC was very risk adverse about resuming this line of computers, and the big technological advance needed was from Gardner Denver, "semi-automatic" wire wrap machines which would sequentially position a pointer of sorts to the two locations to connect, which a nice working class Boston area lady would then manually wire wrap and route to the destination and finish the connection.

DEC owned the high end academic market with PDP-10s and their repackaged and upgraded OS DECSYSTEM-20s until the address space limitation became too severe and the KL10/20 CPU follow-on failed _hard and DEC abandoned both the business line and its customers.

By then they had VAXes (although a lot of burned customers avoided them), which soon got BSD UNIX, 3 was a quick hack that swapped based on Bell Labs 32V, then the BSD 4 versions paged an architecture that could address 4 GB, 8 bit byte addressed (like IBM's System/360 and successors).

And engineering workstations really became a thing with the 68000/10 and its 4GB address space, and for MIT's AI Lab long before then was the home grown TTL Lisp Machines, sort of like the Xerox PARC Altos but 32 bit vs. 16 bit and 24 bits of word addressed address space, 64MB total for the first production CADR/LM-2 model.

MIT also ran a KL10 to provide the symbolic math Macsyma program to the world through the ARPAnet etc., LCS bought a KL20 DECSYSTEM-2060 since by then TOPS-20 wasn't awful like TOPS-10 (which reserved half the by then very limited address space for the OS, see Stanford's WAITS for a hack for that), and after the AI Lab lost its engineers to Symbolics it bought another 2060 and hired an overenthusiastic nigger to be its system administrator ("six letter file names were never this bad").

I think this Wikipedia section gets the AI Lab vs. LCS story correct, certainly MIT people on the ground will do anything short of murder to get more office space on its cramped campus:

https://en.wikipedia.org/wiki/MIT_Computer_Science_and_Artificial_Intelligence_Laboratory#AI_Lab_and_LCS

Multics probably not long after the split went mostly into the hands of Honeywell, which eventually killed it for no reason other than internal politics, having 3.5 incompatible lines of computers, its own, what it bought from GE which included the Multics project which was based on GE's mainframe, and Xerox's which was SDS originally.

MIT Multics, the Pentagon's Multics used to draft their budget, much higher security requirements than anything else they did :cirnoheh: other Multics systems I think lasted through the mid-1980s until Honeywell forced everyone's hands.

https://en.wikipedia.org/wiki/PDP-10

"> Unfortunately for the Internet I accept the principle 'You will own nothing and you will be happy.'

"You're here, not Twitter."

Where Sun is my first level landlord....

"> [Stuff you're addressing with Revolver, except of course for DDoS.]

"> So it's "landlords all the way down,"

"Well, I could argue the point with you or I could hack on the thing that eliminates these problems so that I can release it and then not have these problems."

Good to hear. Even more that some of Revolver is in production.

If you decide to translate it from Go (a good enough first choice) to C, contact me. Not going to learn Go when I have much more preferred managed languages, but I know C cold after many decades of using it.

That said, as you might partly guess from my last missive, I'm not exactly a programmer and sysadmin by choice, I much prefer history and SCIENCE! But I will program for a good cause in a sane team.

"[...]"

"> MIT had no serious UNIX systems at all when started there"

"Ha, yeah, I expect they didn't exist. TOPS/20 and ITS and CTSS and VMS and Project MAC and whatnot; I hear it was a zoo back in the day, and Unix was not taken seriously until they had no choice."

UNIX™ would only fit the "had no choice" by the time Project Athena started (which paled in comparison to CMU's equivalent, and used its distributed file system).

It was so badly managed by a non-computer type who managed setting up the first prominent VMS computer system for four small engineering departments, plus the technical lead Corboto had lost some marbles, so no one good and established enough to be wise wanted to work on it, we had lots of choices (but of course undergrads worked on it and did good stuff).

So UNIX™ and patching together stuff made a lot of sense and basically worked once DEC fixed their bad glue inside disk drives problem (seriously, one RA81 almost half a GB drive failure a week in the initial fleet of time sharing systems before the move to workstations).

Before then thanks to AT&T's generous, essentially free for new installations UNIX™ licencing, it crept into many places as people bought or repurposed split I and D PDP-11s, /44, /45, /70 (although I have no idea who bought one of the latter and for what), then VAX-11/750s (much cost reduced from the original /780 with gate arrays, DEC having first used them for the PDP-11/44's FPU).

Kinda like PCs did in that period, something that bypassed a lot of issues and the IT department's expensive pay by the drink offerings, although then MIT started charging overhead for buying your own computers....

"> Multics was killed by Honeywell, UNIX™ became big, such is life.

"Well, to hear Ken or dmr or McIlroy tell the story, it was basically dead when AT&T pulled out...."

Well, it did for them, they had no access to one, plus it took a long time to mature. See the gecos field in /etc/passwd, Bell Labs/AT&T kept with the original formula mainframes. But, no, Multics became a technical and commercial success, but again Honeywell's management was terminal.

For example, the first and last pure Multics CPU project was beset by problems, the last straw being algae in the water cooling.

But the Official blame was laid on deciding to microcode it (!!! !!!). You know, one of the biggest reasons the System/360 was fantastically successful (MIT did have a 370/168 running from across the street VM/CMS, people only used that for the usual reasons of established software packages and sometimes raw number crunching).

So the max performance of a Multics CPU was severely capped at around 1/1.5 VAX-11/780, async design is hard, and a normal config had a maximum of six, plus two redundant units that connected them to memory and disks. So the only reason to continue running Multics was the software, and often the B2 level of security, see again the Pentagon's love for their system to keep the different branches and other fiefdoms from cheating when the budget was prepared.

Only one really good thing came out of this AT&T withdrawal from the project besides of course Worse Is Better survival characteristics, piping when the small address space required splitting up big tasks like making technical documentation (first formal funding for the UNIX project) into discrete parts.

Like doing eqn for equations and troff formatting for a phototypesetter, roff for a daisy wheel . And thus the small tools that do one thing well method and philosophy was born, although now "What you're referring to as Linux, is in fact, PoetteringOS/Linux."

https://en.wikipedia.org/wiki/Gecos_field

https://en.wikipedia.org/wiki/CP/CMS

https://en.wikipedia.org/wiki/Daisy_wheel_printing

@p @charlie_root @sicp @JoshuaSlocum @mangeurdenuage @mischievoustomato You're welcome!

Haven't read the book yet and its not searchable, but while discrete components would have been bad for Cray once he was able to move to supercomputers at CDC and signal lengths mattered, for less intense purposes they were OK.

Obvious disadvantages in terms of reliability due to many more connections outside the germanium/silicon, space (see above), and cost, but DEC and IBM used them in modules to mitigate many issues including maintenance, IBM starting that practice with vacuum tubes. One reason the PDP-6 failed hard was at the heart of the CPU a cage of six PCBs (each for six bits I assume) connected with multiple ribbon cables, and it was very hard to take it apart diagnose and fix a problem, put it back together, and not create another physical failure in the process.

With physically independent modules connected to wire wrapped backplanes, even big PCBs with TTL like for PDP-11s, fixes were easy ... well, as long as the wire wrapping didn't get damaged. It was so thick on the Cray 1 it was called a "mat" and the heat it dissipated wasn't trivial. Fully automated wire wrapping was also sketchy in putting too much tension on the wires. The connections were pretty much rock solid, gas tight even. On the other hand a coil of wire carrying current is axiomatically an inductor....

I don't think there was a real revolution even from the days of vacuum tubes until you could put all of a serious CPU on one chip (for a long time not including floating point, but that's slow and was OK), and the 68000 family got that going with a 32 macroarchitecture and a 16 bit microarchitecture except for the registers.

This was also RISC CPUs time in the sun, when their higher memory bandwidth and lower latency requirements weren't so bad due to comparatively fast for a while DRAM, and caching of course, including off chip, and you could get more performance out of the chip itself due to requiring fewer gates and microinstructions to get anything done. And the trade off of a 31-63 size register bank, register 0 always returning 0.

Each step prior to the single chip starting with Whirlwind just made the standard hardware architecture layout better in the obvious ways. I highly recommend Project Whirlwind: The History of a Pioneer Computer by Kent C Redmond, in his own words heard much later by my own ears, Jay Forrester solved all the problems for the foreseeable future of laying out a computer and moved on to (ugh) modeling the world (badly, AKA "Limits to Growth").

Also birthed the first good memory system for computers, magnetic core, perfected by a grad student named Ken Olsen, who later worked on a transistorized Whirlwind version and then co-founded DEC. IBM did add the optimization of making all the units of a computer small and light enough to fit in an elevator, the first UNIVAC model was so big it had to be disassembled, shipped and reassembled on site, taking a long time and likely adding damage in the process.

@p "> UNIX™ would only fit the "had no choice" by the time Project Athena started

"Yeah, I think maybe earlier, right? Athena wouldn't have gone the way it did if MIT didn't have some Unix rolling around already, I thought."

Yeah, DEC made a crazy configuration for VAX-11/750s that the Federal government wouldn't want to get the around GSA "most preferred customer pricing" requirement to sell a bunch of them to LCS.

Indifferent hard drive, and a Vaxstation 100 which was a not super smart display with a very fast fiber optic connection to the VAX. Thus the X Window System was born with a completely wrong architecture for the future of LANs and especially WANs.

"[... and what really got UNIX™ pipelining going.]

"> 'What you're referring to as Linux, is in fact, PoetteringOS/Linux.'"

Microsoft recalled him from the field, thankfully. Then again, unlike RedHat IBM it's not in Microsoft's interest to make Linux harder to maintain. Maybe, if they've given up on getting super big Windows Server bucks on Azure outside of the usual lockins.

https://en.wikipedia.org/wiki/VAXstation#VAXstation_100

@p @charlie_root @sicp @JoshuaSlocum @mangeurdenuage @mischievoustomato "> And the trade off of a 31-63 size register bank, register 0 always returning 0.

"I didn't know 68k did this...."

Sorry I wasn't clear, that's because it doesn't!

It's CISC, in the first generation addressed the maximum number of gates problem with a 16 bit microarchitecture, like the lower end System/360s, one of which had an 8 bit microarchitecture.

Slow in principle, but gets the job done with speed from putting it all on one chip. Checking just now (I barely used any for serious programming) it does have 16 registers (including the stack pointer), but they're divided into data and address ones, and I recall one or two assembler/compiler people I knew not being happy about that.

@mischievoustomato @charlie_root @p @mangeurdenuage

It's not stupid but the truth.

On the other hand, the term “cloud computing” is also (in a twisted way) a weird kind of IT humor when you realize that "raining down" could also be used as an analogy for the moment when the “cloud server” does eventually get hacked.

(Fun fact: On a German equivalent of bash.org was a quote where somebody quipped - roughly translated - "I love to mentally replace the word 'cloud' in texts about IT by the word 'clown'. "We are a clown-based businesss", "We store our data in the clown.")

@p @charlie_root @mangeurdenuage @mischievoustomato

I wonder if it was joke-stealin' or if it was derived independently.

I can't exclude it but I could imagine that it was derived independently.

In German, the connection between clown and cloud even works as a pun.

"Cloud" => "geklaut" ("stolen")

"clown" => "klauen" ("to steal")

leading to puns like "Die Daten wurden gecloud."

I think there was a browser plugin at some point.

Yeah, and some Greasemonkey script(s). Leading to things like ""Other people's computers" Strife".