Posts by that one in the corner 5493 publicly visible posts • joined 9 Nov 2021
Two quick downvote on tales that admit causing problems with extra routers and DHCP.
I wonder if there is someone here with a bad conscience about doing the same thing but never admitting it; now they're worried that you'll give the game away by describing the symptoms, making a colleague think "Hey, that sounds just like what happened last month then Joe plugged in that router, which he swore he knows how to configure".
> I don't envisage AI ever getting anywhere near such an emotional performance.
Not from the drinkers, but you'll find your phone batteries are completely drained by the performance.
Oh, those poor, sad, zeroes, left alone on the streets while we stay safely backed up in The Cloud.
> . If what you produce makes it easier to find cancer tumours, or improves a production process or makes a self-driving car safer then kudos to you, but we don't seem to hear much about those.
Well, no, we wouldn't hear much about those (except perhaps the safer self-driving, if that ever comes to fruition).
Aside from the old "if it works, it isn't AI anymore" aspect, which accounts for all the useful spinoffs quietly working away as "just normal stuff we use to get the job done", none of the rest are things that you would hear being shouted about in the general media (or even on a tech site, to be honest).
Why?
Mainly, because they work and don't need to be throwing tantrums in order to get some attention or any money from a sugar daddy flashing the cash and hoping to impress his pals with the stunning platinum blonde[1] AI hanging on his arm.
The production process was improved by 9% and costs will be recouped in 18 months? There was a notice in the relevant glossy to announce a price freeze to customers (whilst competitors prices went up) to use the savings by improving retention: a staid business, nobody trusts flash adverts about "revolutionary new methods".
Helping find tumours? Well, there was a report in July just gone: New AI tool can help treat brain tumors more quickly and accurately, study finds but that is too boring for any more shouting: medial stuff like this needs years of testing and approvals, and in the end it will "only" improve outcomes, it won't raise Lazarus. No immediate monetary gains? No immediate screaming from the rooftops by marketing.
[1] pssst: platinum blonde my backside! Rumour has it that model was trained using software intended for a different field. They reformatted the data to trick the program into reading it: the AI equivalent of dying your hair.
Have to admit, I am now intrigued.
FIA, this particular "thing" apparently exists already, so can you point us towards the reported customer need for an "AI" that can make a noise like a siren or some rather dreary music?
Preferably, a massive pent up demand which has been filling the relevant trade papers for months, if not years, with calls to replace existing foley artists, sound effects samples - and all those annoying session musicians. One that justifies the costs of creating and running a neural net on 20,000 hours of samples (presumably at a 44.1kHz sample rate).
And not just "me too" calls for the wonders of AI, this Miracle Of The Modern Age, to be applied to this industry sector just because all the other sectors are getting to play, we want to as well.
> More experimentation would be needed to check it does not perform ChatGPT-levels of plagiarism.
Ah, there is the clever step in this particular approach (not that I approve of what I'm about to describe; let's swap "crafty" for "clever").
By training the model on audio samples, they can (and reportedly have) restricted themselves to material they own the copyright to. When they recorded the music, for which they (may) have had to pay performance rights for modern pieces, they get reproduction rights over that recording.
If the model starts spitting out untransformed audio waveforms then those are simply reproductions of their own recordings. There is no performance going on, just pseudo-random reproduction of materials that they own the rights to reproduce. No plagiarism is possible!
Now, if you want to argue against that interpretation, then you are going to weaken the argument against language-oriented models (i.e. LLMs and the ongoing copyright suits).
So, crafty: they'll have us either coming or going. The sods.
> I wonder if anyone has tried training one of these things on sheet music rather than audio samples? There are centuries of that, mostly out of copyright, that could be used for training.
I believe they have (although don't press me for a citation right now): there is a long history of analysing music properly in maths and computer science (including all shades of AI). Given the stochastic performance of these neural net models you could probably get away with including the Musikalisches Würfelspiel dice-throwing parlour game from the 18th century (using the very best AI tech they had to hand).
Although music is a "serious business", especially if you use old pieces, classical, baroque et al and the "pile it in with a shovel" approach has been fine for a student project (like the singing dog, we are not so much impressed by the quality but...) but serious people take a serious approach and like to hand-craft their rules or, at least, have a result that can explain to people how the piece works (which, of course, neural nets generally can not). That wasn't meant to be (too) snide, by the way: I thoroughly approve of musical analysis, without the shovel.
> That said, the model weights are not open source. They are shared under a Creative Commons license that specifically forbids commercial use. As we saw with Llama 2, whenever Meta talks about open sourcing stuff, check the fine print.
(Haven't checked reality - the repo - yet, but just going from the above sentence and talking it as gospel)
> check the fine print
You've just said that all the software tools are under the MIT licence. Complaining that the weights are not similarly licenced is like praising a oddball programming language's compiler for being open source then damning it because they haven't applied the same licence to all of the programs written in that language!
Although that is not a strict analogy, of course, as the weights are less program and more of a data dump. Which at least makes the use of a Creative Commons licence a sensible usage (CC licences are not a good match for program source code but are fine for data dumps, whether that be text, images, any old pile of nadans, your database of top-ten hits since 1972, ...).
There is a difference between being supportive of open source and trying to stole the flame wars (and I'm aware of the irony that I'm probably doing exactly that).
In days gone by the community had a sense of humour and instead of talking about "punching robots in the face" this would see this as an opportunity to Hack The Planet (one small step at a time).
Hofu Hackers, now is your Time To Play!
Gather your friends and family to create a steady stream of people through the supermarket, all behaving in the same not-quite-as-expected fashion. Can you convince the AI Assistant to send everyone looking at rubber household gloves over to the olive oil ("Customers found it made the experience smoother")? Does it seem to be judging people by appearance? Let's find out: everyone wear something that looks like the supermarket uniform and see what happens when you all buy one bottle of antiperspirant and nothing else. It is looking at how people act around the shelves? Have half the group hopping down the cereal aisle and picking up branded corn flakes, the other half skipping and picking up discount own-brand cornflakes.
Ah, for the heady days of the Improv Everywhere group and classics such as Best Buy uniform prank
"I name this desk 'HMSO Red Stapler' and wish godspeed to all who toil on her"
A mug of lukewarm canteen tea is majestically (how else?) smashed over the desk's secretarial return, that magnificent structure that marks this as Design Classic. A hoarse cheer rings out and is echoed across the watching crowd as the chair majestically rolls down the unloading ramp and docks into place.
As dignitaries leave their platform, small children and other urchins gawp up at the long, smooth sides of the modesty panel, until shooed away by gruff yet friendly working men carrying hessian covered panels. These are carefully arranged around the shiny new desk to form the easily recognisable shape of a cubicle that appears to be an quiet and efficient place to work and yet, by the careful selection of a too low divider, is neither.
Their jobs now formally over, the men file out of the yard and head back up the steep cobbled streets for home. As the last of them leave, the heavy gates close slowly behind them, shutting with a final ringing clang.
A cold and foggy quiet now falls over the rows of terraces that line the grey streets. Streets that all lead down to the great lumber yards and chipboard pressing sheds that form the heart of the community. Yards that now stand silent, a community that stands silent, silent to ensure they'll not miss the call, the call they all pray to hear yet fear may not come again:
"They're takin' on men, they're takin' on men down Staverton yard again. Hunts and Hawk have opened the gates, for contracts to Staples and Equipu came in and they're takin' on men."
[1] oh, you just meant employes with an assigned desk, not actually a named one!
or the car makers will take it away but *still* insist on having "AI enabled" *something* (just for market differentiation, as the quote in the article notes).
At least pictures of flying pandas are safe. Encourage this idea in cars across the world (ok, it'll be pictures of eagles and/or wolves in other countries, or flying teapots in certain other regions-with-obvious-stereotypes).
Otherwise we're going to get "enhanced" wipers (trained in California weather, confuse Manchester for the car wash and refuse to come out of hiding); headlights (that "helpfully blink to warn the car turning across your lane that you are approaching fast, to make sure that they have seen you"); windows (that roll down every time you crawl past the Bradford and Bingley "drive in bank" during rush hour in the snow).
Or worse, some maniac is going to connect the AI to the steering!
> "A malicious actor, or a competitor brand intentionally creates inaccurate or malicious documents which are targeted at a model’s training data. The victim model trains using falsified information which is reflected in outputs of generative AI prompts to its consumers."
Which neatly puts all the blame onto that mean old competitor brand - when *all* the blame would lie on the shoulders of the idiots who just sucked up every random bit of garbage they could find to use in their training set.
In comparison, suppose we heard that the FBI and CIA announce that, acting upon information they received from a bound manuscript found lying on a park bench, they are creating a major joint taskforce to hunt down a "Mr Scaramanga"; this individual is described as an internationally wanted assassin who is believed to use a custom weapon assembled from a gold pen and a gold cigarette lighter. Would we blame Ian Fleming for deliberately misleading the Forces of Law and Order or should the finger be pointed at whoever picked up a discarded paperback and dropped it into the case files?
Addendum: the PDF does mention "training the model on unverified data" *but* that is treated as a separate example from the situation above.
Will that interface be a commonly understood API, say a connection to an SQL database that will allow flexibility to the consumers, allowing them to quickly respond to new circumstances?
Or will it be a strict set of queries, one per paragraph of the requirements spec, each with its own unique style of signature (this one uses JSON, this one XML, this one is in binary and only half the fields are in network byte order. This one - no idea, but that is not a problem because you have a full set of examples in portable COBOL. Here is our phone number in case you want to buy some support time, some more examples in another language or, for the really brave, a new query added to the set.
ESA remembered to send the heavy wooden tripod, got everything nicely focussed on the ground-glass screen and removed the dark slide, but the heavy black cloth had failed to unfurl properly.
Luckily the older galaxies have stopped the young ones in the front row from fidgeting so they can take the shot again.
If we could just have the First XV cluster together now, very nice: hold it, hold it...
Data slowly starts arriving again.
October 17th: Jackie, the junior researcher, who nobody ever listens to, notices anomalies from the magnetometer.
October 28th: Europa occluded. Jackie knocks excitedly on door to gamma ray observatory
November 20th: Xeelee Nightfighter, wings furled, settles into Lunar orbit. Pilot apologises for bumping into Voyager, supplies NASA with insurance details
November 22nd: first attempts to sue Xeelee pilot for whiplash arrive at JPL.
December 9th: Xeelee insurance company responds to the claims.
Date unknown: remnants of humanity stare up at gaps in the dark, dark clouds, thinning at last, as the first stars are seen to shine through once more.
What "duff kit"?
Do you have any evidence that the fTPM is not correctly functioning as a TPM?
Did you understand that the 'f' in 'fTPM' stands for firmware? I.e. this is a bit of code (in the BIOS AFAIK) that provides, basically, a cheap'n'cheerful implementation of TPM.
Ryzen motherboards can (and mine certainly does)[1] provide a connector for a hardware TPM - but even that is unlikely to be a speed demon (although, as a separate piece of hardware, it will run in parallel with the CPU so its slowness won't block everything and cause this stuttering (probably). Still, calling into an TPM more frequently than you really need to seems like it'd slow down your kernel's RNG.
[1] had to check the manual, I know I haven't bothered to buy a h/w TPM.
As far as I can tell, baseline Linux has never claimed to be FIPS-140 capable: Linus removing a performance hog won't affect that.
For anyone who *does* need to certify themselves against FIPS they can get old of a replacement RNG and either use it to replace the kernel routine or just use it within their compliant code.
Or buy the required mods & support from a third-party.
> so i am not sure why someone would access the random number generator continually.
There is a long history of squabbling over the RNG in Linux (and quite probably in other places) about what "must" be included in the RNG in order to "obviously make it better". This is not the first time that something has been added at one time and then pulled out later.
Without checking the commits (which anyone can do) we don't know if the too-high polling was in there at start of the fTPM support or if someone just saw a call that could read the value, without understanding it would be slow, and whacked it into (what we now know to be) the wrong place.
I don't believe that Linus has claimed to be a mathematician and able to judge on those issues on the RNG, he (his lieutenants) check the code for programming errors and has to accept the originators claims of usefulness. Then real life usage shows up problems that *do* fall into Linus wheelhouse and he speaks out...
To start, I admit I've done access control in the past that I now know wasn't great (and am glad that those installs don't get connected to the Internet).
Nowadays, I'd quite like (it isn't necessary, but would be nice) to access some of my LAN boxes from the Internet - and I've got loads of material saying how to set it all up, thanks, but don't trust it enough to take the risk (for the perceived gains).
> You need to address expertise
Gaining expertise, particularly as an autodidact[1], in online security seems strangely hard, compared to pretty much any other area of ops (for a small LAN) and/or programming.
First is a lack of confidence in proving that the methods being proposed actually work, and against what: basically, you have to be able to demonstrate that you can break into the "unprotected but otherwise correctly set up and working system" first, then show that the added protection fixes the issue. I.e. turn a claim (or even just a vague worry) about a security issue into a testable issue in the bug tracker. To do that, I first need to be able to break into the online system like a real Bad Guy and for some reason the books[2] on securing your Apache server has recipes for setting permissions but none for smashing down the door in the first place!
Second, to be blunt, is a dismissive tone about the subject in forums where you'd hope to see better. Even in Register forums, there tends to be many replies that basically boil down to "well, I do better than that"[3] and no pointers to practical sources of learning. Compare that to other subjects (h/w and s/w) where you can often get useable tips and tricks.
To be frank, the end result is that I have very little confidence in any of the "how to do online securely" claims :-(
[1] I'm not in a position to just be put onto an expensive course at company expense
[2] tutorials etc; unless you have some references to better materials.
[3] comments on a Reg story last week (URL) even had someone else pointing out this attitude
And for all of those solutions, the way the OP phrased the puzzle, the final mile west has the person walking in a very, very tight circle around the North Pole. In fact, spinning on one toe, the other toe pushing until it has completed the mile.
In which case, the bear can be any damn colour it wants to be, just stop the world spinning! Heave, yuuurrrrk: oh, look, the bear is multicoloured, with chunks of carrot.
LLMs are amenable to direct analysis of their internals, just on the basis that they are software (even though most of their bulk is, to us, an undifferentiated mass of nadans).
However, it would be very costly to perform such an analysis and be able to then predict what paths the next run will go down[1], especially as the LLMs are knowingly built without anything akin to helpful debugging aids (the AI-speak for this is that the models have no explanatory power). Some of that analysis has been undertaken but that is fragile and very limited (it doesn't trace any of the dynamic behaviour of the model).
So, instead of actually *looking* at what the LLM does (or simply admitting that they don't have access to its internals) they are in great danger of doing exactly what we hate users doing when they try to describe how any other program works: anthropomorphising it and ascribing it some complex behaviour.
An LLM *may* well have something inside it that has discovered a pattern in its inputs that matches what we could describe as reasoning by analogy. But we ought to be examining these things as software, including building them with the intent of formally analysing their behaviour[1, again]
Aside from anything else, if the training process *has* set up as useful an ability as analogising, wouldn't it be jolly useful to be able to re-use and refine that for a new model, rather than just keeping fingers crossed that the training manages to recreate it?
[1] after taking control of the random number sequence
The responses from Twitter were all gathered into a suitable container and are now used to predict the future of X.com (so far, with startling results).
One senior editor did ask if they were sure that it was safe to do this, as the guzunder is getting seriously full, but was told, as the song has it, "it's my potty and I'll scry if I want to".
Well, to be fair, you did forget to mention the story in TV21 where they modified TB2 to go into space as well. IIRC they also (of course) packed Pod 4 and took a modified TB4 to do the actual rescue in an icy "sea" (mind is telling me it was Titan, but given the existence of Rock Snakes on Mars it may have been an ocean of Venus).
Not sure if that story counts a canon, but maybe we have a *very* avid reader here.
Well, if we take the 8 light minutes (quoted a few messages back) from the Sun to the disc, add in the flat Earth claim the the Sun is somewhere from 30 to 37 miles above the surface, I believe that gives us a speed of light around 215 miles per hour.
If the Sun also takes a mere 24 hours to circle the Equator then, um, something about the light not reaching the other side of the disc until the Sun has itself already got there (it might have gone around 1 1/2 times, the maths gets a little - odd).
If the light from "night" reaches the other side when the Sun is also there, it must be "day" light after all, not "night" light - so the night's *are* dark because the Sun isn't there and neither is the light.
Or something like that.
> You start spouting phrases like "sentiment analysis" at IT admin types and they think you're taking the piss.
Why do I get the impression that you also don't see password management as being something that requires empathy? Although I suspect it may occasionally inspire a bit of creativity, eh!
"More recently, Jumpsec security researchers showed how an IDOR vulnerability in Microsoft Teams could be exploited to bypass security controls and send files — specifically malware — to any organization that uses Redmond's chat app."
<cough/>Malware? But they already have Teams installed!
> Surely Tom Lehrer's Elements Song (recently placed by the author and performer in the public domain) will clue in the younger Left Pondians to Gilbert and Sullivan.
The danger is that they'll wonder why the G&S society have changed the words from Lehrer's original.
Similar to David Bowie being congratulated for including a Nirvana[1] song in his sets (Bowie[2] got rather ticked off when that happened)
[1] The Man Who Sold The World
[2] autocorrect tried to change "Bowie" into "Unique", which I suppose is accurate.
From the comments so far (and to come, no doubt) I get the suspicion that neither these particular AI "specialists" nor the people they are addressing have ever actually poked their heads out into the real world to see what is out here. Unlike Register readers.
As in, pretty much everything "apocalyptic" is already known to everyone who reads, watches decent documentaries and is just generally observant!
Bioweapons - look up one comment, check.
Pipe bombs - well, the interesting part appears in the actual name, so check. Bombs in general? Well, we just used to chat with one of the sixth formers doing chemistry and physics double science; then again, we had proper text books in those days, with cross-sections and everything.
Full armed insurrection - read the history books; not the vague "it's page 23, this must be the Austro-Hungarian Empire" ones from school but ones with names like "A detailed account of the uprising in X on date", with addenda by Gen. Markham, rtd.
Nuclear weapons - ok, tricky; but not for any reasons that LLMs could really help with.
No, clearly the real danger is LLMs spouting out a garbled versions of the above.
> a bit of a pain as needed to swap in 78 specific stylus unit
Ah, you didn't have the tone arm head shell that held a rotating end, which allowed you to have both needles fitted at the same time. Very useful addition to the Garrard (sp?) deck, made sure we didn't put the not-in-use needle down "somewhere sensible", as that inevitably meant we'd not find it for a week.
Even better, meant no more stroppy tweenage girl being told that, no, we couldn't listen to any pop music this evening, but there is plenty of big band jazz or show tunes on the 78s to while away the evening. Me? I like the big band better than her choice of pop and would never have thought of moving the little tin from one shelf to the next. Ahem.
> you couldn’t balance a 12” turntable on the device shown in the first diagram.
12"? Nah. But would you believe 7"?
You didn't notice that there is no scale provided and a number of the turntable parts are carefully labelled, indicating that they make up an important part of the patent, being commented upon in the text?
Welcome to the true innovation, Apple's new Compact Vinyl Disc player (or as we called it, "that stupid toy which only plays singles"). The comments note that what appears to be the RPM selector is actually to reverse the platter for listening to demonic invocations (or " You mother sells whelks in Hull", apparently).
PS
The sides of the platter appear to show strobe markings for three speeds (three? Who would trust antique shellacs to this?! Oh,that third must be for reverse but notice part 808 is also labelled: it is tweaked to fake whatever speed you put into the iPad, just or the effect. Or just to have a unique, patentable feature no one else has done...
As you reach past the turntable to tap the iPhone screen.
Next patent: we turn the UI text upside down then tell you to turn the iPad over: ta da, all the Ok/Cancel/... buttons are at the top of the display and you can reach them the tone arm.[1]
[1] multi-line dialogues being against the minimalist design ethic, so if the text is now backwards return the software as "not cool enough to be on the iPad" (Ts&Cs of the Apple Store, clause 17d).
All the painful hoops they had to go through to add the new stuff whilst still keeping the entirety of the rest of the x86 model working, which is a massive pile of - random stuff - nowadays.
The i960 was attempting to get away from all that - and history demonstrated that people just wanted to keep using x86, with its limited and rather non-generic registers 'cos that is what all of the code in the world actually runs on.
Plus they seem to need to announce "big special changes" like "new AI data types" (huh?[1] you mean signed/unsigned/fixed-point values in various bitwidths? how very novel and AI'ish) rather than "guys, you can use R8 to R15 now, okay". Even if Real Programmers end up using the new registers/opcodes "just" to speed up str(n)len (which is actually a really useful thing to do, but doesn't fill a four-colour glossy on drool-proof paper very well).
[1] What follows is just my guessing; the AVX10 Version 2 release is going to be *so* exciting|
Hardly "gymnastics" - there are specific registers that can be read that allow you to determine what parts of the instruction set are available - as the article says:
> Developer code will only need to check three fields, according to Intel: A CPUID feature bit indicating that AVX10 is supported, the AVX10 version number, and a bit indicating the maximum supported vector length.
You only need to write the detection code into one library (and then keep it updated as new material arrives) - or find someone who has already done that - and then use the flags that extracts as required.
If you want to do some gymnastics, you can set up your build system to auto-generate multiple object files from recompiling the source once per whatever architecture variants you want to use and also generate the single function that the rest of your code will call; that function simply checks the flags (from the library, above) sets a pointer to the appropriate one of the multiple objects and then invokes via that pointer (you only do the check and assign on the first call, then the pointer is null, of course).
Again, you set up that build the once and then just need to maintain it in sync with the identification library.
As far as tricksy coding goes, that isn't really very hard gymnastics - call it a gate-vault rather than jumping over the box.
Not quite accurate.
The registers (can) live in a generic array, the "register file" - and the size of that is certainly set by the hardware design.
However, they *can* modify how the register file is accessed via the Instruction Set Architecture (i.e. the ASM opcodes we use to program the beast) and by adding/removing/modifying entries in the ISA they can change how many registers our code can see, how those are used and so forth. How much of that is possible is down to how flexible/generic their microcode design is - and I have no idea how flexible any of the Intel microarchitectures are.
(got a bit distracted, meant to get back here earlier - hope I'm not too late)
ISBN 978-0128119051 - yeah, pretty much so. "Computer Architecture: A Quantitative Approach" by John L. Hennessy, David A. Patterson. Although, just to be weird, I have the first edition of this, which is rather old now, but I looked at the later editions yesterday and came to the conclusion that the newer editions have lost some of the introductory material in order to squeeze in the more modern material! If you're not too worried about being totally up to date the older edition(s) are really cheap on Abe Books.
The same authors also have another series out on the hardware/software interface, which may be more interesting/accessible to programmers - they still delve into the architecture below the ASM opcodes (i.e. microcode) and there are now separate editions that cover "general" computers (e.g. the Intel Core i7, IBM Cell ...), ARM and RISC-V. Apparently also a MIPS edition. I haven't looked through those thoroughly, but one or more of those looks like it would be a good companion for my older "Quantitative Approach". Although second-hand 'cos they are rather pricey :-(
Also:
I haven't looked at this one, but it is apparently worth a look (and it is cheaper than the ones above): Microprocessor Architecture: From Simple Pipelines to Chip Multiprocessors by Jean-Loup Baer. 2009 vintage, it is said to concentrate on the micro(code) architecture level, rather than any particular ASM opcodes presented to the programmer. It does refer to the Alpha, P6 and Athlon as examples (vintage!).
> Chinese made AI-enabled products should spark similar concerns to Middle Kingdom sourced 5G equipment
Like rampant paranoia.
> AI-enabled products present perhaps an even greater risk than 5G which is also more difficult to mitigate ... AI is embarking on a point where it will touch all aspects of human life in a manner that humans will grow accustomed to ... lurking in the background while potentially shaping the way societies think by influencing behavior online, enacted in ways that gatekeep jobs and credit, and also implemented in systems like traffic grids, maritime operations or rail systems - it will be impossible to control.
All of the concerns apply equally to the stuff already coming out of OpenAI et al!
> AI-enabled products and services from authoritarian countries are likely to be overlooked
What? Why?
Oh, because they'll get past the regulations - the same way that crap Chinese kettle leads get past the regulations (and inevitably catch fire)? You mean because the Western buyers demanded cheap, got shoddy and sold that on to us? Whilst those who took care got perfectly serviceable wares from China.
On the other hand, if the only way to get regulators off their arses and willing to pay to have people figure out some *real* controls over *all* of the ways these beasts are deployed, then maybe paranoia is the only way to go.
There can't be any downsides to this approach, can there.
Take an inexplicable phenomenon[1], sprinkle in some random behaviour[2], leaven with the unexplained[3] and feed it the unexpected[4]. Stoke the flames[5].
It is hard enough to be sure that our carefully planned[6], deliberately coded[6] and painstakingly tested[6] systems are safe to release to the public.
The hubris of the Madmen, releasing their Thinking Machines[7] upon an unwary World, is staggering.
[1] LLMs - neural nets in general - have no ability to explain how they reach their results and are just huge piles of nadans when you look at them yourself
[2] as per the article, a seed value plus an internal stochastic walk
[3] guardrails - nice phrase, sounds reassuring; details, please? Apparently added quite quickly, compared to the boasts about how much effort it was to create the LLMs in the first place, so tacked on or really fundamental to the way the model works internally (mot that we'd ever know if the latter worked, see [1])
[4] well, did you expect that suffix to work?
[5] set another machine the goal of finding the "unwanted" results, by a process that looks rather like a maximising "fuzzing" process.
[6] ever the optimist, ignore more of the rest of The Register's articles about systems falling over in embarrassing ways.
[7] yes, yes, these don't think, they aren't really intelligent - trying to be poetic here! Yeesh.
