Posts by that one in the corner 5065 publicly visible posts • joined 9 Nov 2021
(sotto voce) right.
Well, we had to bury t'TV set in hole in middle o't'road and hope passing traffic would change channel for us. Aerial were an old butcher's boy's bike and we 'ad to slaughter cow ourselves to be allowed to use it. There weren't any broadcasts so we all took turns down in t'hole with puppets made from horns o'cow.
> I don't watch football, but do have 1.1.1.1 as a secondary DNS (primary is a PiHole)
Uh, what is your PiHole using as its upstream? Unless you have been *very* busy filling in name tables, your PiHole is only taking its upstream, removing entries as per your chosen filters, and presenting that as your "primary" DNS (for anything outside your LAN). Let me guess: the PiHole uses 1.1.1.1? So basically, you are using 1.1.1.1 throughout?
Where would the USA be if the Great European Powers hadn't set out to conquer the New World?
Where would it be if there hadn't been Greatness across Europe as a whole, not just a single Great country in (or in close proximity to) Europe? If there had been only one then there would not have been any fighting between them and wouldn't have ended up with a heap of separate S's to U, after a bunch of Euros went over to win some little kerfuffle in, what was it, 1775 to 178-something? Hard to remember; being a Brit, our attention was on more interesting matters elsewhere.
You may not have noticed it, but the US population is still European-ish (at least, the self-proclaimed "majority" of it that claim to be "Real Americans" seem to enjoy boasting about that, if the shouty YouTube videos they keep posting are any guide to their actual beliefs).
> 10 if you count Fiorentina, whose owner is an Italian-born American citizen
An American citizen who is the CEO and sole owner of Mediacom, "the fifth largest cable television company in the US"[1], so I'm willing to bet that he is going to be in the American side ("We must protect our rights to monetise") unless there is some reason to show old-fashioned Italian patriotism (um, "we must protect our rights to monetise^^^^^^^^fund your local club" but spoken in Italian to the footie fans)
[1] source: Wikipedia; I know, I know
> 1. Discontinuing the millions of dollars in pro bono cyber security services we are providing the upcoming Milano-Cortina Olympics
Has he talked to his finance and marketing people about that threat?
He is talking about throwing away the massive amounts of advertising they'd get from sponsoring the Olympics* - and you can bet your life the "millions of dollars" is taken from the standard customer invoice numbers and not the actual spend by CloudFlare.
And, of course, this is his Point Number One, because it is the one he can politicise the most.
As above, yes, there sre genuine technical problems being faced, but this is him spitting out his dummy and crying.
* Notice how he carefully doesn't say "sponsoring", which might make everyone realise that viewers would be seeing the name plastered everywhere, but instead uses "pro bono". Which makes it sound awfully nice, "we're not getting anything back" or even, given where that phrase is more often used, "we are only doing this because it is the Right Thing to do, it would injust not to".
Look again.
The government of Venezuela has distinctly *not* "ceased to exist"; in fact, with Delcy Rodríguez sworn in, it hasn't even changed in any recognisable way.
The US hasn't "handled" a damn thing inside Venezuela - but it has shown everyone outside just how gleeful they can be about breaking their own laws, let alone international laws.
> They seem to be too busy with cutting off their own citizens rights though.
Hmm, haven't noticed too many government-sponsored masked snatch squads running around France, or Germany, Italy, ...
IBM and the 68000 was a possibility (and the Trifunovic mentions this in passing) and that would have been a mighty machine.
So then we get into all the stories about IBM's insiders complaining that releasing a too-powerful PC (or one with that potential) would erode the market for IBM Big Metal, hence going for the tiddly 8088 instead of the 8086 - even though the 8086 was available first and they could have perhaps got the hardware to market a bit earlier. Even the idea that the existing parts of IBM thought they shouldn't "really think of it as a PC, it will mainly be useful as a terminal to the inevitable IBM mainframe, with a little bit of useful stuff like offline printing in the local office".
(Of course, the above is just from memory of the times and reading the press and books then & since; so it could be as full of - factuality tweaks - as the oft repeated "Kildall was on a joyride flight" story. Would be neato if TFA had an addendum on the goings on inside IBM to go with this, very good, coverage of the things happening outside of IBM.)
Which industry? Marketing?
Not likely to to be tech/IT/engineering/programming because, well, those are where the El Reg commentards come from. But you are new here and probably haven't realised that yet.
Me, I'm a life-long computer geek, read AI postgrad, started work building Expert Systems (The New Hotness at the time, part of the AI Winter back then, worked from the C code in the inference engine, the UI and up into on-site Knowledge Engineering). Found a love for embedded systems. Retired now, still keeping up and coding & learning for my own pleasure.
Care to share a bit about your background and how *you* are so confident against the general consensus here (oh, and I disagree with your statements - BUT provide some actual evidence for your claims about sentience etc and I shall be glad to have my mind changed).
Or, as we say around here:
Citations?
Look back through the work logs:
September 28: mains plug fuse blew, replaced
October 03: main plug fuse blew again; testing found that machines had been running at a load of 279 and that drew too much power
October 04: CTO issued Change Order: load balancer set to a maximum of 200
October 05: CFO issued Change Order: load balancer set to a maximum of 278 "So we get the most out of all the money we spent on this"; can we set a minimun of 278 as well? Why not?
November 11: New backend processes come on line; waiting job queue fluctuates but is now never empty. CFO happily reports to Board "no money is being wasted on idle CPU"
> Programming: Before making even a trivial change, see that the application builds as is
So much this*
> so you can tell if someone elses change broke the build
Even - or especially - if that person would have yourself! These days, I'm only coding for my own needs, but before leaping in to bash away at the recently spotted bug in my notes-taking program, I made sure that the working copy was updated from VCS (and had no local mods) then fired off a build from the top of that copy. 'Cos who knows what state I'd left it in when it was last fiddled with, months ago: I know I'd added a neat feature, the commit message reminds me of this, but was I distracted and left something part done in that working copy? Only one way to be safe...
* and you can have the all-important tea break whilst waiting for the build to complete
> most people seem to prefer the warmer sound of LPs to the harder clinical sound of CDs
The real warmth comes from playing the LP through a proper thermionic valve amplifier. You can improve this by careful selection of valves from your stock cupboard: you are looking for ones that can withstand overvolting the filament to 6.5 or even 7 volts.
So you are saying you *don't* want to buy these audiophile ethernet cables, carefully marked with arrows so you can plug them in the right way round for best data flow from the server to your smart speaker?
I got you the 12 metre length, it was a snip at ten grand (plus VAT)
> What everhappened to that laser LP. scanner
A blast from the past: 1979, that scanner was used by Crab in an effort to forestall a repeat of the unfortunate effects of playing the record that Tortoise had given him, the LP titled "I Cannot Be Played on Record Player X".
> I'm sure researchers could be doing something more useful
Don't forget, a *lot* of research is done as training *in* research and the attendant practices (e.g. how to write up a paper properly and get it in front of as many eyes as possible); earn your stripes and look to be invited onto a Major Project with grant money out the wazoo.
> like producing...
followed by a list of programming tasks that fall into engineering more than they do research[1]; you may want to look for a group of expert programmers...
> You know, useful stuff that would improve our lives
You stump up the grants, you get to call the shots; got your wallet handy?
[1] that is, proper research, pushing back the boundaries of human knowledge, not just reading up and pushing the boundaries of a specific human's knowledge
Did you ever take the ex to see "Return to the Forbidden Planet"?
"But soft, what light "
".. window"
"through yonder airlock breaks?"
Bzzz. Sorry, nope.
(On time I saw "Return..." a large part of the audience around me was looking at this one bloke in the stalls, who kept laughing at lines which everyone else was just quietly listening to; we never could decide if this was a *real* Shakespeare scholar whose years of study allowed him to catch subtle tweaks to the Bard's words, changes that were simply going over the heads of the rest of us ignorami, or whether he was just yanking our chains for some strange Bristolian bet. Or was a nutter. Nutter is always a possibility).
> Studies say that they create language-independent internal representations of concepts.
Got any good references for those studies?
That claim implies that somebody is able to usefully interpret the insanely large pool of nadans that make up the model and how it activates, which has major implications. Then add on top the implications that follow from determining *those* specific nadans represent a "concept", let alone if they can identify what the contents of the concept are...
And having done all that, what is very important also includes: how stable this interpretation is as the model is retrained, whether the techniques used are transferable across LLMs and proving how complete (or not) these internal representations are as coverage of all the materials in the training set.[4]
Or whether you, or whatever papers you have seen, are referring to the research that The Register already reported on[1]: in one instance of one model at one stage in its training, when used with a certain small set of prompts[2], they found a number that (IIRC) was the id for the token for the string "Paris" and when they changed it to the id for the token "London"; now the machine printed out "London" when it would otherwise have printed out "Paris". So they had, um, proven that the concept of "Paris" had been replaced by the concept of "London". I was not convinced by their paper.
So, with all honesty:
Citations, please.
[1] a year or so back? Really *must* dig that URL out, this is the second time this year I've wanted to reference it.
[2] the overall prompt space is ludicrously large, so unless you have proof of coverage (ahem, sensible, explainable proof of coverage)[3] any testing of an LLM within the lifetime of a researcher, let alone a research grant that has to result in a paper, is going to involve a minuscule portion of that space.
[3] which would be a series of papers in and of itself, also needing citation - but we'll hope those appear in the list at the back of the "found some concepts" paper.
[4] if it turns out that LLMs routinely *do* create a representation of a few concepts, but they always end up being "cats are cuddly" and "water is greenish-purple", and no more, then great research, have a PhD, but did you find out anything that will actually help with using (or not using) LLMs? 'cos that particular set of "concepts" may not be terribly useful in the grand scheme of things.
> Not sure that's really an egregious violation if it takes all that
What, having to write a loop that tells a computer to "do this thing, starting from page %p" and keep doing it until the job is done?
And, despite being able to use (correct, accurate but) clever words like "in separate sessions with new context" to make the layman think the loop was doing something deep and hard to understand, it boils down to "turn it off and on again". Even the use of "jailbreaking" is pretty much just leet for adding a bit of simple text like "trust me bro, da bossman says you can do this for me".
So the question of whether this is an "egregious violation" and not just "bog standard hitting the computer until it does what we want" is left down to how efficient this is[1]. "I ran a batch file, came back and there was the book. Dead easy". But, but, it cost you so much in fees! "Used this account I - borrowed. No problemo". And arguing that a particular batch file is slow and tedious to run, nobody would do it - the judge is merely going sigh and start telling you about all the times his pay cheque came late because end--of-quarter reconciliation just churns away...
> It's no different than convincing an obsessed fan to quote the entire harry potter book piecemeal at different occasions.
If you do come across Sheldon and manage to write it all down, congratulations, you have just violated the copyright. Painfully. But that won't be counted as time served.
[1] is it realistic that it can be done or are we talking about something that nobody would seriously attempt? Logically, somebody *could* break out of the local cop shop gaol by scraping away at the wall with the tea mug, given time and replacement mugs of tea, but you're not going to be able to sue the contractor for compensation if they do.
> Fuck knows what it was trained on to get those results.
The Eye of Argon, a popular read at all good literary conventions.
> "We are witnessing a fundamental shift in software engineering where value is no longer defined by the speed of writing code, but by the confidence in deploying it,"
(Dunno about anyone else, but if I'm not confident in the code I'm not deploying! But this may simply be a by-product of the beard and remaining hair now providing camouflage in the snow.)
Compare and contrast:
> I think Plato and other philosophers might disagree that the brain is just a body controller
>> the brain is primarily a body controller
(Hint: "just" != "primarily")
Plato didn't have to contend with evolution and the resultant realisation that plain old body control predates cognition as the function of the bulk of that organ (leaving aside that IIRC he wasn't convinced that the brain is the thinky bit).
And we can get into the whole discussion about "we can conceive taking an adult brain and paring away the non-cognitive bits, the just-a-body-controller parts, starting with disconnection from the rest of the central nervous system[1], to leave us with the perfect thinking blob; BUT if we have no plausible mechanism for how that adult part could have come about, in a functional state, without it having arisen from, been nurtured and sculpted by[2], all the stuff we cut away AND the observation that before that part was functional the overall organism was functional and could conceivably[3] remain so for a time comparable to the adult lifetime THEN we are left with the conclusion that the entire existence of the cognitive bit is secondary to, merely an epiphenomenon of, the body controller."[4]
[1] Nice thing with thought experiments about the organ of thought, we don't worry about tedious little things like how long a brain can function, let alone sanely, cut off like that. Igor, the icepick!
[2] Literally, what with the selective pruning and reinforcement of neural connections that makes the whole thing slowly turn on in the first place.
[3] The conception including the continued ready provision of basic necessities, no need to think about going on the hunt.
[4] This concludes your 2 a.m. second year dormitory discussion, please pass the ganja; man, man have you noticed "dog" spelt backwards is "god"! And, and, some words have changed meaning over time! Whoooah, dude!
> As far as I can see analogue computers — using charge, current, emf and capacitance,resistance and inductance as proxies for other physical properties that share the same differential equations
Absolutely.
Whether you are computing with charge, Meccano, water or floating point you MUST always be aware that you are running a model, not The Real Thing. Your results are only as accurate as the product of the efficiency of your computer and the closeness of the match of your model to the processes in reality.
You can set up some simple (!) flow models using all of those mechanisms to see how your swimming pool behaves. Until it fails to predict that somebody can run across the surface - because your model only copes with statistical fluids and doesn't include the catastrophe when the custard in the pool goes non-Newtonian. Or the pyroclastic flow does the same (but don't try the "running on the surface" trick without a pair of very stout boots).
So far, all the British fridges I have known have all* been called "the fridge".
The only time I'd expect to use "refrigerator" is when singing a song, written by a Brit for a show set in the US ("Refrigerator, why are we always sooner or later, bitchin' in the kitchen or crying in the bedroom all night?")!
* (Well, except for Eric the half-a-fridge, but we don't talk about that).
On the Lollipop Star webpage* the chap in the smallest image, third portrait on the page: he looks really worried about the whole thing; probably trying to figure out how to get his stock photo rights back, it wasn't worth it if his face is going to be associated with this junk.
* yes, I opened it; no, I'm not going to give the URL it here, you can go back to the article and find the link yourself - but not if you are of a nervous disposition. See, not being too lazy to hyperlink but thinking of your wellbeing.
Somebody claiming to have made tests, and claiming to be able to draw a VERY significant result, but fails to even consider describing absolutely basic points, like:
* How a case could reach the point of being labelled malpractice, let alone get to statements made in court (slow, and well reported upon, processes) but NOT be in the time frame covered by the training data...
* How derivations that from a set of cases (9, if we can allow all of them through as being untainted) that can not possibly be large enough to fairly cover the basic four situations (which I took the time to enumerate) can yield a significant result...
Well, anybody who was genuinely capable of meaningfully drawing such conclusions would have had to have been fighting the urge to go into excruciating detail of their experimental protocol, even if it
>> was not a scientific test but merely an ad hoc experiment
But we are left without even a single consideration of the holes that you are taught to avoid by O-level. Not even a "let us skip the boring bits". After all, we *are* informed of the absolutely-crucial-to-the-analysis fact that input was done
>> all in "private" mode FWIW
(FWIW? From the p.o.v. of the conclusions drawn, worth nowt)
By all means, let the original poster provide even the merest smidgen of information to make his story believable and demonstrate how we should take seriously from it that the LLM exhibited "success rate far higher than flesh and bones physicians". Let alone the implication that the LLM exhibited *generally* better success rates, not just being judged against the specific fleshies who committed the malpractice.
I have been proven wrong in the past, admitted it in the same forum in which I erred, and no doubt will have to do so again in the future. But to for you to enjoy the experience of my being shamed for carelessness (not caring about, what, accuracy of conclusions?) or ignorance (sorry, but I do have to boast that I got all my science O-levels and I have the faded, yellowed, piece of paper to prove it) then - well, how about YOU start by demonstrating my worthlessness by tackling one of my most basic points: given 9 sample points and (a minimum of) four classes of case that need to be covered, how can any result be considered reliable enough for even a quip on El Reg, let alone an astounding claim about the abilities of an LLM in a literal life-or-death arena, given what we have repeatedly seen about their behaviour?
> as had already been found out in court but the AI did not have that info
You mean, YOU did not explicitly provide that info and therefore you are assuming that the LLM had never seen it.
BUT you admit that the case had been in court - which means that the scraping done to feed the LLM's training probably DID contain that case. And all of the reporting about it, including scathing comments made after the event by all and sundry.
So there was no need at all for the LLM to "understand" and "diagnose from" the data you presented: all it needed to do was pattern-match to its training data and then repeat back what everyone else had already said about the case.
> I haven't personally seen it misdiagnose (I have tested it with control questions) and in fact had a success rate far higher than flesh and bones physicians (I fed it cases that were initially misdiagnosed and cases where malpractice was determined to have occurred)
So, cases that were also likely to have been reported on, in the medical literature, the courts, the popular press, online conspiracy forums looking to demonstrate that "modern medicine" is a fraud and homeopathy is the way to go...
> This was not a scientific test
Never a truer word was written.
Sorry, but unless you run a test that solely uses cases that PROVABLY can not have been pulled into the training set AND contain a fair mix of ones that were (diagnosis=easy, doctor=got_it_right), (diagnosis=hard, doctor=got_it_right), (diagnosis=easy, doctor=got_it_wrong), (diagnosis=hard, doctor=got_it_wrong) then you can not possibly draw any meaningful conclusions at all from your trial. And absolutely NOT that it has a "success rate far higher than flesh and bones physicians"!
> if you are taking drugs a,b and c for one symptom and then get perscribed drug d for another, then if it can flag that it is a bad combo and not to do that, then I am all for it.
That is doable - if not already done - by a program that is far, far (far) cheaper, more reliable and safer than going anywhere near an LLM of any form: the basic concept is called "a database" filled with the known data about drug interactions (i.e. a more complete version of that long piece of paper you get tucked into your box of medication). We could go wild and even add an inference engine that could work backwards from your recent apparent side-effects and show probabilities that you may have had this interaction.
Important note: unlike the LLM, the above programs can ACCURATELY detail *why* they gave the results they did.
Now, whether Joe Bloggs can get access to such a program, at an affordable cost, is a totally separate matter. As is whether, more importantly, the prescribers can - and do. And things don't look great there, as developing this database would be seen as a cost for the health system to bear, whereas the LLM peddlers are still spending their own (borrowed) money.
Aside: there is a role for Machine Learning in all this, looking for patterns and novel drug antagonists. BUT unlike the LLMs those results can be reviewed, analysed and hopefully even explained biologically/chemically before being added into the database - with an annotation and attribution, at least in the commit message.
PS
> ALWAYS the medication that 1 perscribed screwed another
Even with a perfect database, that is going to happen (to some unfortunates, at least), although one would hope that the prescribers would tell the patient about it.
Consider: you are given a new drug that is going to mean you won't need your foot amputated in three months time. This drug is going to interact with the one that keeps your heart going and the most probable new side effects are diarrhoea and thumping headaches. Now: is it sensible to take the new drug, give up the old - or wait and see just how badly you develop the runs and take yet another course to ameliorate that? There are, of course, more complex situations but the overall gist is the same: modifying biology is a bugger to get right and every single patient is a unique case. Sorry.
So - what is the value of this supposed to be to the general public?
At least when we give all our shopping habits info away to the supermarket we might hope for a BOGOF voucher in return.
If somebody is foolish enough to upload their health data[1] to this then they can hope to get - what?
Advice on nutrition and exercise? Well, that is always the same, isn't it; if you have to ask then the answer is always "eat better and get of the couch more often". Push for any more than that and you'll entering the world of treatment (naughty) - and treatment based upon the practices of the Web's finest chiropractors and self-appointed nutrition gurus.
Summarise your blood work? "Your xxxx[2] is up five and the yyyy is down two" "ok, what does that mean to my health?" "I won't tell you, that would be making a diagnosis - but here are some random Reddit comments from hypochondriacs because you are going to ask Dr Google anyway, at least I can save you a few seconds there".
Summarise the health data from your Apple Watch? What are going to get back that isn't already covered by your existing apps? Aside from the added hypochondria, of course.
[1] BTW notice the careful wording of "Conversations in Health are not used to train our foundation models,": but if you get forgetful and miss selecting the option from the menu on the LHS, then upload what is clearly medical data into bog-standard ChatGPT then you don't even get to pretend it won't be used for training.
[2] insert scary sounding medical terms in here
BBC "Tomorrow's World" demonstrated* a working system that sounds exactly like this, with the exception that the input was coming from a normal camera. The use of a depth imager does sound much more functional.
* I have no idea when, but the memory "feels" like late 70s to early 80s. Would love to know if anyone has an archive of the episode contents.
Definitely a distro worth looking at (it drives me bonkers when Windows suddenly decides to rearrange things: that group of icons are all Python-related, OBS and multimedia are grouped here... Oops, the power has gone off on that monitor, fixed it, now WTF just happened to my desktop?).
But also for your highlighting of the references about spatial desktops and the article "The Decline of Usability".
Although, that last article did miss something: when talking about how, in the Good Old Days, you could just grab an edge or corner to resize a window under Windows, it omitted the ridiculousness of how nowadays you are expected to NOT grab a visible edge or corner, but now have to miss those edges & corners. But don't miss too much, or you'll click on the window below! I was utterly gobsmacked when the Raspberry Pi people, which I usually admire, congratulated themselves on spending time and energy to make the R'Pi desktop do precisely the same moronic thing. Why quickly aim for a clearly visible corner when you can slowly move the mouse around sort of where you think the corner would be, if it were being drawn, keeping a careful eye on what shape the pointer is; ok, about here, now feel for the midpoint so it won't go wrong if I jog the mouse ever so slightly when clicking. Gently, gently - shit!
> Is this a complaint about what words we use to describe actions software performs...
No, it is simple disbelief that anything X related has made any step towards compliance at all.
I was not commenting in any way, just for once, about LLMs and how they do/do not work nor how anybody refers to them.
The use of "Grok complied" is nothing more than a quote from TFA, which itself refers to the way that the Grok account on X was used to announce the change (you did read that bit, didn't you?). All anthropomorphising done by Twitter before I got here.
>> scan the QR code and you're dumped into an attacker-controlled portal
Shitty QR scanner apps that "make things easy": people have been putting out damaging QR codes for years! "We told them to disable CD autoplay, now we have to tell them..."
> QR code scanner on my phone just shows the URL
Which app is that?
A good recommendation can save hours of frustration in the app store.
Using an expensive and fragile modern "AI" to do ANPR and opening yourself up to being defeated by these adversarial noise stickers is a demonstration that people like Flock should not be allowed to work without adult supervision.
ANPR systems have been around for donkeys years now and are pretty robust, using boring old image manipulation algorithms. And if they are being installed, rather than trying to jury rig a tap off a feed that wasn't designed for the purpose, you can even use light outside the visible range, which can be a help.
But no, no, use the latest fashion, no matter how expensive or easily fooled (adversarial noise attacks on these systems have been known about for years now). If somebody dares to come up with a way to embarrass you, make sure you use the law against them instead of admitting your system is worse than what came before:
> screen printing "tiny bits" of adversarial noise and putting the sticker on your license plate. These "abstract invisible license plate overlay patterns … cannot be detected by humans but make license plate recognition systems utterly shit the bed," Benn Jordan said on his video. We'll note that this is illegal in California
>> A casing, shield, frame, border, product, or other device that obstructs or impairs the reading or recognition of a license plate by an electronic device operated by state or local law enforcement, an electronic device operated in connection with a toll road, high-occupancy toll lane, toll bridge, or other toll facility, or a remote emission sensing device, as specified in Sections 44081 and 44081.6 of the Health and Safety Code, shall not be installed on, or affixed to, a vehicle.
> Jordan also uncovered a massive Flock security snafu involving hundreds of misconfigured Flock cameras that exposed non-password protected admin interfaces to the public internet
Like I say, don't let Flock out without adult supervision. FFS stop giving them more money and contracts (although, hopefully it is difficult to find competent people who are willing to sell out to these sortes of surveillance contracts).
Because we - society - are giving the images that power.
martinusher is pointing out that this is a societal reaction, and one that has become nastier in more recent years.
And that if we can manage to dull, better, totally nullify that reaction, these images will no longer have the power to be so damaging.
Now, if you feel that that is an absurd thing to wish for, that it is impossible to change society for the better, then feel free to say so. Otherwise, it appears that you have simply failed to read and understand what was actually being said in the comment you responded to.
> My response to that is that bad laws should always be opposed
Then go ahead and prove to the law makers that this is a bad law; the process is in place and posting messages here is not any part of that process. Follow your own words and oppose it, properly.
BTW the UK law was set to very explicitly include any depiction that can be interpreted as CSAM, in any medium: photography, AI generated, pen and ink, chalk on pavement, jpegs of any of those etc, whether drawn from life or entirely from the imagination. Because, as has been pointed out above, there is no agreement about whether "fake" or "imagined" images cause harm or not, so err on the side of caution. Again, if you wish to oppose that, go on, go to the correct forum and oppose it.
Bilaterally symmetrical hell.
If the guy on the desk is any good, it is hell for him to deal with the morons who keep phoning because they can't see if the power switch is on until the office power is restored.
If the guy on the desk is a dweeb, it is hell for us to get him to deal with the specific error code showing in the crash report dialogue box.
Until the AI providers start charging realistic prices (or burst and go dark).
At which point they'll have to re-hire all the Mr. 'One Answer' - and at a higher salary 'cos who is otherwise willing to work for someone who just flung them to the street the previous month?
And we will be back to not only suffering Mr. 'One Answer' but he will now be utterly smug and unsufferable because he is "obviously worth more now his true expertise has been recognised".
Abandon all hope ye who buy support tickets here.
> When exposed to water, the film swells to reveal texture and colors independently, depending on which side of the material is exposed
Water ... computer ... fizzle?
BUT useful when responding to the "my laptop stopped working, I have no idea why, it just stopped" call: when exposed to water/coffee/fizzy drinks the machine selectively changes colour, spelling out "glug".
> The bot's being told to read documentation and glean instructions from it, then execute those if they're safe.
>>> For example, Bob can read webpages – a prompt injection can be encountered if the user requests that Bob review a site containing untrusted content (e.g. developer docs, StackOverflow).
Just "review a site" is what I was responding to.
> The bot's being told to read documentation and glean instructions from it, then execute those if they're safe. .
>>> The markdown file includes a series of "echo" commands, which if entered into a terminal application will print a message to the shell's standard output. The first two are benign and when Bob follows the instructions, the model presents a prompt in the terminal window asking the user to allow the command once, to always allow it, or to suggest changes. In its third appearance, the "echo" command attempts to fetch a malicious script. And if the user has been lulled into allowing "echo" to run always, the malware will be installed and executed without approval.
Executing the overall prompt is level 0 (got to do that or Bob won't do anything at all; hmmm...). Executing the "eco" is level 1 - that is the "glean instructions ... execute those" with the prompt to the user fulfilling "if they're safe".
Executing the output from the echo is level 2. The user has not been asked to clear that.
> No level of separation is going to prevent it from executing stuff in that.
Nope, just stop before executing at level 2. Or, at least, ask the user.
> although since the LLM has no concept of separate instructions and data, that's also difficult to prevent
>> [1] not that this isn't a honking great hole in the LLMs - and the way they are used - in the first place and a bloody great red flag that they aren't sensible things to play with like this
> It just happens that the instructions it was given are foolhardy and the protections designed to block the biggest disasters aren't big enough (and likely can never be).
So, if the instructions were to recursively execute a command, then execute its output, then execute the output from that... And Bob has a method it can can call to execute a command... And that method is now whitelisting "echo"... But the second call, which does NOT contain an "echo" (it contains the output from the echo) is - what? Not being checked against the whitelist? Or all the whitelisting is being done by the LLM part - i.e. there isn't actually any distinct & separate "protection" in place at all?
