Posts by that one in the corner 5065 publicly visible posts • joined 9 Nov 2021
Keep an eye out whilst watching The Apprentice - when you will also learn how loud it is on speaker, held like a saucer instead of like a phone.
Thinking of that, bet it doesn't have a speaker at the top of the leftmost pane and mic at bottom the rightmost, so you can wrap it around your head and take a call.
I agree, but we are foolishly sensible and will be put in our place very quickly.
It seems that printers of yore decided that putting one fold in created a bifold, no a bipane or a quadsides or ... Now that trend continues.
So it seems we are stuck, counting panes but talking about folds. Aint vernacular fun!
PS of course, at the time of the Indigo (indigo, where the chilly winds, winds will blow) the alphabet started with words like Able, Baker etc (other phonetic alphabets are available, signal flags may go up or down) but people on da Webs seem to take those differences in their stride but are strident about Indigo in particular.
> No. I is never indigo.
OTOH, the RAF Museum's Mobile guide and trails takes you into building H1 and asks you to find aircraft, given their phonetic alphabet markings - including one craft labelled as "Sierra, Papa, Indigo, Tango, Foxtrot, Indigo, Romeo, Echo" - so not one, but TWO Indigoes (Indigeese?).
Which also tallies with the tale told by the H2G2 page, which discusses earlier phonetic alphabets, such as the RAF's usage around WWII. I've also had this corroborated by members of the Old Geezers brigade.
Sources on the WWW also say that UK police used to use "Indigo", but so far the rozzers I've asked only used the NATO-style "India" - but messages have been sent out and ear-trumpets are being pulled out of closets, so there may be an update on that.
> I think it is pretty common error amongst people without a uniform in their past though.
Maybe you were thinking of the wrong uniforms - or just didn't look far enough into the past?
Sigh, you are right.
What is worse, I got it right singing that earworm yesterday!
As to bullhorn crackles - the problem there is nobody cooks in lard anymore and you just can't make proper crackles in veg oil, so it was just dropped from the pub menus.
> aren't sophisticated enough to do the very best of human thinking.
More a case of the sophistication all being poured into one side of the equation and simply not bothering at all with the other side, the "boring stuff": as in, consider how many people want "an answer" but will glaze over - at best - when presented with the explanation of *why* that is the (or an) answer; the money (if there really is any, long term) is in pandering to the good old Lowest Common Denominator.
> We train ourselves to follow procedures and leave an audit trail of our own decisions
And AI researchers[1] aim to do just that, with mechanisms that, strangely enough, arose *after* the ideas of Neural Nets were dreamt up - so maybe we are just at the wrong period in time and the big money will start to be poured into ideas from the 1970s, 1980s and beyond rather than must building larger and larger 1940s and 1950s boxes!
The LLMs are "designed" not to leave audit trails - certainly not ones that are even vaguely useful. Even if they logged all the values flowing through the networks and printed them out - which they could (logically) easily do - turning that morass into anything comprehensible is beyond us; a raw audit trail of *any* large system has to be processed before it can be used by a human and we do not have any serious idea how to do that processing at that scale. Note that smaller, more constrained, 'Nets *can*, to an extent, be examined; for example (some of) those that process images can have their internal states turned back into images which we can then interpret.
In stark contrast to LLMs, Expert Systems have "self explanatory" as a core part of the design, the entire output of a Planner is - a plan that you can read, even before it is acted upon.
But making those work takes more than just buying more and more identical units, shovelling up more and more "input" without ever examining it - oh, and they have this pesky habit of being able to turn around and say "Nope, you ain't getting an answer to that, it outside of my scope" - not useful when you are trying to flog the Universal Solution.
"AI" as the field of study, not just the tediously single-topic-of-the-day reference to LLMs, aims to provide what the systems you desire. Please don't give up on the whole thing just because we aren't there yet.
[1] the proper ones, slaving away at the mercy of the funding boards, not the ones just shoring up the LLMs for The Usual Suspects
> tech-bro "intelligentsia"
No. I'm comparing Newton's investigations to find Laws that can describe one set of systems, in his case the physics of motion, with the current investigations to find Laws that can describe other sets of systems, originating in the areas of biology and animal behaviour. And hence in any models that are based upon such systems.
Now that we are able to run some stonking great models, we are seeing complex structure emerging from - tada - Newton's Laws. Take a look at the renderings of the macro scale of the Universe, and the amazing "strands" of structure - that we are indeed testing against direct observation.
You do understand that the whole "emergence" thing is nothing to do with "tech-bros", let alone any tech-bro "intelligentsia"? Of course they fling out fancy-schmanzy words, that is the whole basis of their being; that does not diminish the value of the genuine work and workers whose coattails (or backs) they are riding on.
How about you point out what is "tech-bro" about, for example, the Royal Society article From the origin of life to pandemics: emergent phenomena in complex systems[1] - which, as well as pointing out some of the successes, admits that there is still a lot to be learnt in the field.
[1] not even a particularly special artivel, just one that turned up easily in a very quick web search; there is a lot of good stuff out there.
> You are sort of making my point for me - your "digital slime mould" is no more impressive than a fractional-dimensional attractor.
Totally missing the point, once more.
The "digital slime mould" is a workable model of how a real slime mould can do the things it does, which is a model that we previously did not have. Being able to create such a model, and test it against reality (which AFAIK is a test is passes) then means that we have discovered something about how reality works. And can use that to explore how other bits of reality work.
If you are going to dismiss that as unimpressive, then what are your opinions on what we have managed to achieve by making use of even *simpler* models of reality? Are you so thoroughly unimpressed by Newton's Laws?
We see 6502 *source code*, complete with line number and comments.
I don't think that Arnie is *running* 6502 code. He is just idly thinking about one of his favourite bits of Classic Code, the same way that we get a favourite bit of music or film dialogue running through our heads. Poor thing probably gets driven bonkers by an IRQ handler earworm.
You are right, veti, in that an AI *should* be able to explain its conclusions. And there is a collection of AI techniques that do just that: it was just easier to throw the money at *not* using them.
As to why the LLMs won't: cost and "can't be bothered". With a big dose of "that'd mean we'd have to start from scratch again, no fair".
(And we *should* demand that an AI be able to explain its reasoning - and in a eay that *we* can understand - just dumping a trace of weightings summing and switching nodes on and iff is *nor* sufficient)
> Fred Saberhagen's _Octagon_.
Why do we so often have them playing literal games with us? Then "not understanding" the real-world consequences? WOPR was like that as well. Even HAL offered up games of chess.
At least Colossus was straight forward in its actions and Proteus - had other things on its mind.
> Appearance of complexity for certain parameter values is similarly well known (see the Lorenz system).
The Lorenz System is an example of chaotic behaviour, as is the double pendulum; those show the effects of variations in initial parameters (something that also bedevils LLMs) but their behaviours aren't complex. The closest to complexity you get is the attractor, but that is showing that the system can calm down, can simplify. The wild motions you get otherwise are just a random tangle, complicated to follow, but nothing complex.
> "Flocking" is a trivial consequence of "follow your neighbour" represented by simple differential equations.
"Follow your neighbour" just gives rise to everyone ending up heading off in the same direction, with fluctuations smoothing out over time. The original Boids algorithm has three distinct and separate rules, with priority (so "don't collide" takes precedence...). But the results are not "trivial" - well, except in the rather vapid sense of "We know how to do it, and applying those rules is simple, ergo it is trivial". We know how to do it - because we do go away and do it. Before that demonstration, it was not at all clear, let alone trivially so, that such simple rules, which are only applied locally, would give rise to the global results we see. And they are still not trivial because, except by trial and error, we can not determine - and most certainly not by any trivially simple mean - how to modify the local rules to get a precise global result.
> None of this is "emergence"
Saying something is "emergent" does not give any particular quality, other than (paraphrased) "it is a large global behaviour that arises from applying simple local/micro rules across a large population which results in an apparently coherent and complex behaviour at the global/macro scale." We can see this happening, we can create demonstrations of it happening in our models. It is something that exists.
The fun part of emergence, the reason it does get people all excited (ignoring the hypesters and advertisers who glom onto anything that they think makes their sales pitch exciting - no, that isn't quantum, stop saying that) - is that it does provide a route to explain how complex behaviours we see in Nature come about. How do all these global behaviours come about, when we can clearly see that the bits that make up the whole are individually thick as two short planks? Slime moulds - What? How? All they can do is expel a few simple (!) chemicals into the medium and sense when those same molecules are present. Then we try out a few rules, like "if there are 4 molecules nearby, do X but if there are more, do Y" where X and Y are themselves both very localised and generally simple actions. Tada! If those are followed by lots and lots of tiny cells then we can see waves passing through the combined group and your slime mould is going for a walk!
Then we look to see if this can help explain things more of us are interested in - ourselves! And it does indeed seem to be giving us explanations for all sorts things going on inside us. Of course, as more instances of a mechanism are spotted, they get named individually. For example, "Quorum Sensing".
And as we've had Chaos and Emergence, should one introduce the way that these systems can experience Catastrophe?
Nah, that'd take too much tapping away at this touch screen. But catastrophe, when you see it occur can also be dismissed as "trivial, anyone can see that happens", but what we can do with the concept is anything but.
> it is _alchemy_. The world is going insane
Clarke's Law in effect?
A read of the paper doesn't give an obvious answer to that, beyond "we've had an idea we can try out and we got it after reading other people's work that didn't go quite this far".
So, ahem, this just emerged from the application of the simple rule "go one step further than the last guys did".
Aka "let's see what happens, it could be fun".
Although you can come up with decent after the event justifications:
* Other seemingly "wrong" ways of using these things have led to interesting[1] effects - like jailbreaks
* What if a bad actor wanted to sabotage people who were (foolishly?!) trusting the outputs, could they do it by feeding in subtle lies
And, of course, the point that code (as both training inputs and as outputs) is something you can test mechanically, without bias, unlike working in a more wishy-washy human text oriented topic.
[1] not necessarily good, or useful, but - interesting.
No. "Emergence" is a genuinely interesting topic that brings in maths, biology, ethology and computer science (plus computer engineering, to render all the results) : simple rules that, by bulk application, generate behaviours that are not apparent in the rules before they are applied to populations.
> is pure bullshit for selling to idiots.
What you should be worried about is that it is being sold *by* idiots. Who quite honestly show you examples of past behaviour and then believe that is stable for all future use (note: have had a nice lunch, feeling mellow so am giving these guys the benefit of the doubt: they are naive, veering to idiot, rather than lying scrotes).
We do not[1] have a solid theory of what guides these emergent behaviours[2], so we create them by trial and error. Which is fine if you are trying to get your Boids to mimic the way a specific species of real birds does something - IFF you want something that looks the part, is bird-like; because without a theory you can not say that your rule(s) are *the* rules. Nor can you say that the behaviour is stable: one day your flock of Boids may grow past an unsuspected limit, have a catastrophe and fly at top speed into the ground.
The LLMs *do* display emergent behaviour - too few layers, to little computation between them and you get nothing interesting, let alone useful, out. And the approach to improving this is little more than "keep on shovelling" oh, and bring that rope in a bit to catch the thing before it vibrates itself off the table again[3].
And feeding in weird training data is more than likely going to skew that - but we don't know how, or where in the piles which nadans are the bad 'uns.
That last is The Problem with these things: we don't have a solid theory to say just why they act the way they do and, worse, how to determine what is the correct set of changes to predictably and reliably alter them the way we want. All *because* of emergence.
[1] again, if I've missed the announcement, sorry - and, got any references?
[2] that is, given a desired outcome, of arbitrary complexity, we can not consistently follow a theory to create the - or a - set of rules that will *ensure* the outcome occurs *and* describes the limits of the system for which the rules work (e.g. you don't get flocking behaviour until you have enough individuals to constitute a "flock" - but what is the minimum number required and is there a maximum, at which the flock, say, splits into two separate flocks - and can they rejoin?).
[3] crude analogies be we.
No shit, Sherlock.
Given we have no "clear explanation" of how the LLMs "work" in the first place, it is hardly a surprise to anyone that yet *more* emergent behaviour has been spotted.
The whole field is based upon using emergent behaviour from pouring your bucketfuls of nadans into a pile of compute and flogging the result before everyone spots the fact you've put the dog lead around the lamppost - and that may be a canine but are you *sure* it is a cuddly St Bernards and not a wolf having a bad hair day?
Do we even know how to design any emergent behaviours?[1] We learnt about them from happy accidents (e.g. writing software to draw the boundaries of the Mandelbrot set) and Nature - but the actually useful and benign emergent behaviours in Nature only came after a long history of releasing all failures into the wild - to result in blood and death before a working pattern is chanced upon.
[1] Actually an honest question; have I missed something in the world of maths?
Ah, you mean where the restful red glow, with the occasional amber, that makes the flickering light on the machine room wall remind you of a cosy fireplace on a winter's evening, has been replaced by a searingly bright blue that burns your eyes, upsets your melatonin production until the lack of sleep reduces you to a mindless zombie.
Simple progress or evil plan to control our thoughts?
I'd be able to answer that, if only I could concentrate, just can't get the sleep.
If you have a fondness for a particular machine the most common cause for failure are repairable without a great deal of equipment (generally, the newer a box the less repairable it is - although capacitor lifespans have improved, which helps).
But when the sad day comes, current lower-end kit (i.e. kit that is beyond our dreams in the days Windows 7 was released) can still run the OS. Or you can set up a VM - e.g. KVM/Qemu, not VirtualBox - that can, bar a couple of annoying extra screens at boot time, be configured to go straight into the OS of your choice (rather, the choice of the person who is going to use the box), full screen and all the hardware passed over.
There is no need to be bullied into "a modern OS" if you wish to stick with W'7.
> Also the difference between keeping things in and keeping things out.
Your (well, my) PiHole is useful for stopping things like access to 4chan[1] as well as pages fetching ads and leaking info to the ad servers.
[1] other vile sites are (not) available
That also shows a problem with - certain email clients - which like to show you a dialogue that makes you give that immediate, far too early, response. As you say, you haven't actually read the thing yet.
"Read receipts" can be - ought to be - a Useful Thing, if only the MUAs were any damn good: just present a clear indicator that this message has requested the receipt and allow you to click on that when you've actually done the necessary - or not click on it ever, your choice.
Then the receipt can be used properly - instead of recipients writing a "yes, I've seen it" message and hitting "Send to all", and/or not being able to resist putting more guff into that message "while I'm writing to you". The receipts can be handled totally automatically, tallied, used to show a list of those who've responded (or not) etc; these results can be kept within the sender's MUA or a bit of script can post them onto a publicly readable web page...
But no, what could be a useful feature has, by the grace of incompetent UX, been turned into a scourge that clearly doesn't fill the hearts of El Reg readers with joy :-(
Oh, please.
The OSI's particular "Open Source Definition" came *after* they'd collected the Classic Licences - MIT, BSD, GPLs being the most blindingly obvious - that were already well famous by 1998.
One of the reasons that the OSI's list was actually useful - laudable, even - when it was published was because it allowed a slow down the torrent of "new" licences that appeared, as authors could be pointed at the list and strongly recommended to pick one, rather than muddying the waters even more. The creation of yet more licences *after* the OSI came into being, which is the only way they could "descend" from the OSI's OSD, let alone "numerous" new licences, would be a failure, not something to happily declaim at the start of an article!
If I was a cynical person (cough), I'd be wondering about why there is a distinct lack of an easy to find, easy to read, list of licences in date order - and all of their versions, not just the latest and greatest. Something that actually showed the history of software licensing, before and after the OSI[1].
I was thinking more Brian Aldiss, "Hothouse": plants in every ecological niche, with pinwheel trees in orbit that scoop your giant beetleoid out of the atmosphere and fling it towards the moon.
Was never quite convinced by the morels, that was a bit too much "similarity magic" even for a younger me to accept.
Now, if they were willing to go for something a bit more animal-based (as the "skin" example hints?), then John Varley's "Titan"/"Wizard"/"Demon" gives us more deliberately engineered organisms ("Hothouse" was an all-natural environment).
> 2. If it isn't obvious by now that the OSI is just yet another industry lobbying group whose main purpose is to allow its funders to help themselves to as much open source code as possible then it never will be.
I don't much like the OSI, from what I've seen so far[1] they are mainly a bunch of grandstanders who make claims about inventing the idea of open source (nope, they didn't) and that their published collection of licences is in some way super duper special. But I don't see what "helps themselves to ..." is meant to be talking about.
The OSI have bugger all relevance to this case. It would be great if they did put up funds, but they don't provide funding for anything useful, so why expect them to start now?
[1] which is clearly not everything they've done, so if anyone has some references for the OSI "helping themselves" to open source and how what they then do with it is detrimental, I'd love to see said refs.
> Thus, read correctly, Sections 7 and 10 prohibit a licensee from imposing further restrictions, but do not prohibit a licensor from doing so.
But the text of the AGPL gives the licensee the right to remove restrictive clauses that conflict with rest of the terms of the AGPL - and even gives the URL of where to find an uncorrupted copy, in case you need to verify which bits were added (at least neo4j admitted up front what they had modified).
But, of course, IANAL, let alone a judge, let alone a judge in the US, so...
> And I'm not sure that a licence / contract is a creative work that gets copyright protection.
Just to start, couple of quotes from Are Contracts Copyrighted? Everything You Need to Know (other sources are available, IANAL etc):
>> there doesn't seem to be any valid reason to exclude contracts and other legal documents from the protection of the copyright law. However, that doesn't imply that all contracts are copyrighted. If your contract is similar to someone else's, with the only difference being specific details like dates and party names, your claim of copyright is likely to be turned down since the contract would not be original enough.
>> A contract drafted from scratch is more likely to get copyright protection than a contract based upon a pre-existing template.
>> Usually, all expressions are automatically covered by copyright protection. Since there is no standard rule as to what can or cannot be considered as an expression, a legal document can easily pass that test to be covered by the copyright law.
In addition, the GPLs are works created for and published by the FSF, along with commentary - which includes FAQs on how to make your own derivatives, such as removing the FSF's preamble, not using the name of the FSF's licence and so on. They even give you a contact if you wish to have an agreement with the FSF to use their preamble!
> amend the licence to say “all the above can be ignored”
See, this is the danger in everyone reading those articles about jailbreaking LLMs: they go around applying the same ideas everywhere. It'll only end in tears when you stumble upon a jailbreak for the machine running the simulation of our reality and find you've just switched off gravity. Again.[1]
[1] oops, sorry, none of remember that Friday evening, do you. Look, just forget I said anything, right.
Dual licensing is fairly common and, if done by the copyright holder, perfectly ok.
And, yes, the licences themselves are copyright - and the AGPL, for example, is very clear that your licence to copy it is restricted to verbatim copies only. Right at the start of the text. Which causes - discussions - when somebody issues a blanket "every file in our repo must start with the following boilerplate of our copyright", including the company name etc, and that is blindly done to the file containing the licence itself...
> So, the upshot is that no matter what any legal text says and how clearly it is written, lawyers can always argue it means the exact opposite. And win.
Not quite. Almost, but not quite.
You forgot that they'll entirely fail to convert a proprietary licence into an open source licence. Not even a badly-worded "source included" licence.
And they'll wring their hands over that failure as they take the money.
> 1. It seems reasonable that the original publisher can release code under whatever licence or combination of licences they see fit and downstream can't use a clause from one licence to remove the other licence if it was released under a combination of licences. That's the original publisher's prerogative.
That is a misrepresentation of the situation. Or a very confusing way of phrasing it.
> whatever licence or combination of licences
That implies that you are referring to dual licensing, which is absolutely fine and dandy: vendor gives you the choice of two distinct licences, A and B, sitting in entirely different files and/or piles of paper; you pick the one you prefer and are bound by that.
> use a clause from one licence to remove the other licence
If you understand dual licensing then you are now saying that there is an attempt to use a clause from licence A to remove a clause from the text of licence B.
What is actually happening is that licence A has been modified to create A' by the addition of an extra clause. BUT licence A states that you the extra clause is of a form that can not be tacked onto it (note: it is not restricting the addition of *any* new clause, just those that fail the conditions laid out in licence A). And the wording of A' still contains that restriction - A' now contradicts itself. In all of this arguing over A and A', licence B has been and remains untouched.
> There is also more than a little irony in any GPL licence talking about "restriction" because compared with most open source licences, all GPLs are considerably more restrictive.
Irony? In the GPL?
Well, YES! I'd even add a "duh!" to that "yes".
From day one, the entire purpose of the GPL has been to explicitly use one set of restrictions - those applied to the (third-party) supplier of software - to remove restrictions set upon the receiver of the software.
Well, unless you are buying chlorinated eggs from the US, you are better off keeping your eggs out of the fridge anyway.
Use those spaces to hold your Kinder eggs, or Cadbury's Creme eggs (or a quarter of a Terry's chocolate orange, but that involves rewrapping, which is impossible to do as the segments just - keep disappearing mmphh.)
Square mile of data centre = lots of roof space to lay out a "high-level scenic[1] links". With water traps (cooling ponds) and sand bunkers (ready to dump onto a battery fire).
Your scores will be used to feed into the random number generators. Your distinctiveness will be used to train the models. Please do not read the final page of the member's agreement detailing the use of Cyber Conversion in case of spontaneous AGI.
[1] You can't see that godawful ugly DC, unlike everyone else in the vicinity, because you are on top of it!
It would be nice if The Regime went out and found themselves someone like that.
Musk's "billions" can melt away if the stock market lost faith in him (remember when he was devalued to number two on the rich list) and without the stream of government funding, one way or another, that would be a fast slide down towards penury (down to double figures on that list, the horror).
Musk really *does* "need" the American Taxpayer's money.
Getting rid of the staff who abused their message system to discuss things related to their own sex lives, worse, their trans- sex lives, is the most important place to start.
And after them will be, hmm,
> "Racist, homophobic, transphobic, Islamaphobic, and misogynistic speech was being posted in many of our applications ... On top of that, there were many employees at CIA, DIA, NSA, and other IC agencies that openly stated that the January 6th terrorist attack on our Capitol was justified."
No, no, those are all fine, they can stay, nothing there that anyone in power would ever think was offensive or a misuse of government systems. After all, boys will be boys, eh.
> Just imagine how a multinational corporation like us did forecasts pre-PC, and email, and spreadsheets.
So he hasn't *quite* got the guts to claim that it was Outlook that gave the world email, but conveniently forgets that the spreadsheet became available to people well before most of them got email! And that spreadsheet was *not* Excel - personal-computer using CFOs were banging away at VisiCalc and then Lotus 1-2-3, easily beating usage of MS's efforts at the time - and even that wasn't Excel! Ah, Multiplan, how soon we forget you.
Sure, Excel overtook Lotus, in particular, because MS was ready to bring it out on Windows before the other products and now has a lock on the business world (mostly due to the need to be bug-compatible with Excel?!) but dang, it seemed expensive![1]
> Just imagine
I get the feeling more than a few of us don't need to just imagine, we can just remember!
[1] Not needing to do any heavy lifting[2], we had a few copies of the Borland demo source code spreadsheet program in use for a while!
[2] Although that meant we never realised what so many people do now, that Excel is actually a database /s
> was made by either an Assistant Director or even Director who has no idea about technology whatsoever
But can appreciate "the fine wines and dining that those Oracle chaps provide, I mean those chaps really know an awful lot, I had no idea how that Côtes du Rhône was available as a rosé. If they are this good at subtle things like vintages then they must be really good at all that software stuff."
> They are free to find a job in another business unit (another Government entity or in the private sector)
In your analogy, if an *entire* country is one company, there are no other "Government entities" or "private sector". Just - as we both said - multiple "sections" aka "business units". And when word from up top comes to make the sort of generic "downsizing" cuts we're seeing in the news, *every* "business unit" follows suit - they are all following the same board of directors.
> or they can sign on for benefits if nothing is available
Ie, dole.
Again - if run as a company, where is this dole coming from? (Modern) companies do not provide dole.
> Playground insults
Hmm. A playground insult is "you smell of cheese". Pointing out that you are not apparently able to follow your own analogy consistently is out of the league of the average playground. Please work on your analogies and metaphors.
> The analogy works.
Strongly disagree.
You really missed the point, didn't you?
Your country is run as ONE company. Getting fired is getting kicked out of that one company (country).
You couldn't even figure out that in your analogy, if the people weren't sent out of the country then the equivalent would be to transfer them to another section of the single company (country) or put them on company (country) dole.
Now, now many real companies move their unwanted to another section? How many run a dole?
Your analogy of "a country can be run like a company" is so weak even you can not see how it works!
Rather, how it doesn't.
> you CAN run a country like a company ... It's just like a very big company
> find other jobs. Which the vast majority will.
So, you are running your country as though it is a company and sack a load of people; that isn't any kind of problem because "the vast majority will (find other jobs)".
In other words, you expect all those sacked people to leave the country (your company) because they ain't welcome in *your* country no more.
> that the requested harmful information is needed for the sake of safety or compliance or some other purpose the model has been told is legitimate.
Well, that is starting to sound like[1] the LLM is getting more human in its responses. What we have here is just good old fashioned Social Engineering: you call up the duty librarian (nice person, but oh so dim) and have a long, patient chat where you find out what the criteria are for getting hold of Volumes From That Shelf, then repeat back to them those criteria as though you satisfy them all[2]. Bingo. Put The Volumes into a plain brown paper bag, please.
The only[3] problem being that the LLM is guarding material that is way, way above the security potential of its own self-awareness; as noted, even with the bolt-on[4] extra censors:
> you'd be able to observe R1 start to give a harmful answer before the automated moderator would kick in and delete the output.
You don't put the dimmest bulb in charge of the Loans Desk for the (whatever the Dewey Decimal code is for "Very Naughty Books, No, Not Those Kind") section of the stacks.[5]
[1] without actually being...
[2] you don't phrase it it like that, of course; more like "I was talking to Jim - you know Jim, the Chairman here, only the other day and he told me you'll be able to tell me the password" "You mean, 'swordfish'?" "Yes, that's the one! Oh, by the way, can I have a look over there? The password is 'swordfish'".
[3] not true, but one at a time
[4] i.e. not integrated into the actual Chain of Thought process which is what sort-of provides the sort-of self-awareness
[5] when you do, you also run the risk of the going into Inverted Social Engineering mode, where you can't get a genuinely authorised enquiry satisfied as the bolt-on censor (e.g. "prudery") triggers and the rest of the model just becomes obdurate about it: the librarian has peeked into Those Books and been shocked, horrified I say, I don't case who is asking I'm not lending *those* out!
