Posts by that one in the corner 5493 publicly visible posts • joined 9 Nov 2021
LLMs replacing proper search engines.
Only a small number of the largest LLMs in existence.
A way to make an LLM forget targeted - oops, "only copyright" data.
What could possibly go wrong?*
* There are no records of anybody abusing our system, the safeguards are all working perfectly. Banana, banana, banana.
There are resources on the Moon that have great potential value to science, especially Astronomy: rather a lot less atmospheric scattering/absorbtion, away from Earth's radio chatter on the Farside, and for a couple of weeks per month, away from the Sun's rays in every wavelength (the other two weeks let you charge the batteries). And our base lines get longer (they also become an ever-changing thing - but computational interferometry can sort that out). It'd be a lot easier to pump down the arms of a couple of Gravity Wave detectors up there.
Now, whether anyone will put up the money to access this valuable resource is another matter... After all, we can "get by on the cheap" with bigger Terrestial telescopes (if they can be finished and left to operate in peace: that latter is in doubt these days) or with satellite telescopes (with their size and operational life restrictions: ok, ok, LISA will do far better than any surface, 'quakey, instruments and will be a lot longer in the arms, but still has the operational life issue).
Reginald: Do we have the marketing survey results for our latest Grot product lines?
Joan: Yes. Aah, 87% told us where we could stick our non-soluble suppository
CJ: I didn't get where I am today without knowing where to
Reginald: Yes, yes, thank you. Moving on
(Apologies to everyone with a better memory and/or a copy of the DVDs for the gross inaccuracy in the quotes, but haven't seen that episode for a while... But the overall flavour of that scene has stayed, un-dissolved...)
I'd forgotten the - coincidence? - that the issue containing "The Last Question"[1] also featured a story "Deus ex machina"; you could be forgiven for misremembering which title belongs to which story.[2]
[1] No, I didn't see it new in the shops, I'm not quite that greybearded, but there was this great secind-hand bookshop...
[2] Oh come on, that wasn't a spoiler for the Asimov story! Seriously, if you haven't read it yet...
Sometime around, oooh, Mid-November, maybe into December.
In the meantime, MS have only a few more opportunities to release some more patches that'll *really* screw over all the Windows 10 holdouts.
"Oops, silly old us, we've knackered all your PCs. Not to worry, see, *this* patch has put everything back the way it was, and here is the *proper* fix that doesn't trigger UAC all the time. What was that, your Win'11 boxes are now all working again but none of your Win'10 boxes got that last patch and still aren't usable? Gosh. Well, sorry, but, you know, EOL, our hands are tied."
Hint: consider that you are creating a counterfactual which does nothing to refute or demonstrate a weakness in the previous statement.
Because the provision of ability to restrain does not imply that the holder of that provision is necessarily in a position to be "the Perpetrator in Chief, the Abuser in Chief...". And in the case of the UK Monarchy, most certainly is not and hasn't been for a long time; there was this pesky Parliament, you see...
> are trivially problems that require Turing complete languages to address, and yet we insist somehow that a DSL is the proper tool for the job.
Um. Are you trying to say that a DSL (you do mean Domain Specific Language, yes?) can't be Turing Complete? 'cos it can, if you want it (or need it) to be.
(I have no experience with the specific systems you mention and any particular DSLs that they may, or may not, contain, btw, just commenting in general about aspersions being cast against DSLs as a whole)
> I despair when I look at most software projects with documentation automatically generated from the code because all it tells me is how the software developer chose to decompose the problem into units of code
Glad you said "most", not "all", because it *is* possible to use something like Doxygen to generate documentation that runs the gamut from end-user to bug-fixing or feature enhancing dev. I *think* I managed it once (once!)[1]: well, the client liked it at the long handover meeting and nobody ever came back to grab me for a quick shout in the following years...
BUT using (in this case, Doxygen) like that was, I admit, very much going against the grain of how Doxygen itself is presented (and the same is true of similar tools I've looked at): they go into detail about how to mark up your code *but* there is little, if any, emphasis on writing SEPARATE files that are just documentation - to the extent I've been asked "what was the point of these dot-dox files, they only make the build system more complicated?"! Bangs head on table. "Why not use Word?" Um, version control; keeping materials together; regenerating and fixing references as the code matures; being able to drop "TODO: write this dox" into the same tracker as we drop "TODO: write this function"; the docs are broken into easy chunks so you aren't constantly looking at hundreds of pages of Word content in a disheartening and never ending task; auto-generating variants (End User Manual is done in PDF without hyperlinks to any code; Installer's Manual in PDF hyperlinks the User features to the INI files, Dev's Manual in HTML has all the same, but now hyperlinks to where the INI settings are acted upon in code...).
Getting just devs to write docs in "the code markup tool" is hard enough, but getting management to allow it, yikes. You can fudge things, sometimes[2] but, again, the tools (that I've seen, so far) don't make it easy.
And "by easy" I don't mean "write another bleeping WYSIWYG IDE that does it all"! You can sensibly allow "I am not a dev" authors to use, say, a WYSIWYG editor for writing Markdown that can then be pulled into the dot-dox, that is good (great, even). But you also need to be able to drive the generation from a build system (GUI begone!), so encourage long and meaningful file name extensions[3]; you need to pull in externals (images, auto-generated lists of CLI options, ...)[4], must support multiple programming languages (without still being so obviously a Python tool that got hacked for C++ and Bison that has those guys foaming at the mouth).
Most important of all:
When you go to the website for the tool (looking at you, Sphinx, Doxygen - well, everyone) and the "how to use this", please, please, START by showing how to write the User docs, the Overview Of My Library, even The Next Great American Novel, as organised prose; output into lots of language lovely End User and Management-friendly formats. THEN segue over to cross-referencing your example code. THEN into adding details into the bulk of your code! That last bit really, really is the least important, honest. Management may leave love to see a table of all the parameters of flump19() but if you haven't got that far yet (as it is last on your list!) a direct copy of just the declaration is a usable standin for anybody who'll actually use that level of the docs! Go back to the history of *when* that struct has five fields, not three, and the arguments that'll at least mean you can say "well, you were warned" when Mr SuperNewbie "optimises" your code and it all stops working!
[1] important note: I had time *scheduled* on this project for a write-up *and* management wasn't demanding it be done in Word (phew), so there was an opportunity.
[2] my little utility for a particular End User doc generates Word files, as per Management demands. Well, to be honest, it generates HTML that looks like the stuff (an old) Word spat out and gives it a dot-dot extension...
[3] seriously, I encourage the use of fred.markdown.txt, or even better, fred.github_markdown.txt! *Any* user can double-click and read it in Notepad, any *sensible* Markdown editor can open them (AND prompt for that as the extension when saving a file! Your's can't? Then it is in the wrong!) Make loves those names, gets all the options set correctly.
[4] sigh, ok, I'll add a pass through another preprocessor into the build - getting further away from one nicely documented document-writing language as we go
> "it is the industry's practice"
Great way to break an illusion!
The academics, individuals every one of them, chasing dreams - and finding out, is it still your dream by the time you've finished the PhD? During your postdoc, as "publish or die" kicks in? Whilst you spend the days drafting grant applications, looking wistfully over at the lab machinery you've not been able to spend time with for months now?
Then the journals - especially the predatory ones - responding like that.
There is nothing wrong with "industry" per se - we all rely on it, it has many great people in it and it is (should be) symbiotic with research and academia - but sending that message was really on the nose: the owner of the factory with the giant rotating knives emailing the long pig standing on the conveyor belt.
From the world-weary "yet another way to trivially trick" to the overview of their operation - "slurry", like it.
The only thing that is missing is the ending to the sentence:
> showed that "agentic" tools, in addition to simple interactive chatbots, were also vulnerable.
Something pithy to the effect that "of course they are, why wouldn't they be?" and a quick bit of "and you've connected them to systems with real-world access, don't come crying to us when they've opened your windows, wiped your drives, eaten your hat or had sex with your cat".*
* The pithiest is, of course, "Go stick your head in a pig" but do the people falling for the LLM hype and leaving themselves open to these attacks have a good education in the classics, so they can recognise the reference? Presumably not, as they already missed that all these LLM pwns reek of wooden horse dung.
Isn't *THE* reason to buy an "official Google device" because those are the ones that get third-party OS support?
Instead of praying that your newly-purchased "one of the thousand random devices released this year" happens to have also been purchased by an active alt-OS dev, so it will be supported (as soon as said dev has worked their way past all the "what the blazes is inside *this* one hurdles).
> a lot of devs will want to use Grok, and there's no reason they shouldn't be able to use it.
So, what is stopping you (or them)? Just go to xAI, buy into whichever scheme suits your use, maybe install a plugin for your fave IDE[1], and off you go. No need for Github to do anything, one way or the other.
> There's an open issue on the JetBrains support site with lots of people wanting Grok access added.
See [1], again.
[1] Wot, no plugin? You're a super-powered dev-with-AI, write one!
Never understood what you as so bad about being an ass - hard working, robust, our pre-mechanised industrial heritage relied upon them. They'll trot up and down cliffsides that a horse will shy from (goats do it better, but you try loading your mining gear onto a goat). A bit of a reputation for being slow to pick up new concepts, but that is really just them being cautious and weighing the options instead of dashing for the shiny shiny like those noisy magpies.
Now, being an arse, that is not good.
> the real deal is being with the artist ... at a gig.
With all the genuine sound of people singing the wrong words, out of key, just behind you. The bad mix of the instruments that drowns out the singer, who then starts to shriek instead of singing. The volume is too high, so you "feel" the music, even as your ears start ringing and you lose all ability to hear the useful frequencies anyway.
It's a toss-up, especially when seeing a band live for the first time, whether it is going to be great - or yet another one where the best thing to do is walk out early and try to preserve some love for their work via the studio recordings.
> Simplest solution for the thousands of projects using fast-glob would be for Malinochkin to add additional maintainers and enhance project oversight, as the only other alternative would be for anyone using it to find a suitable replacement.
The *only* other alternative?
Not, keeping a known copy in your own system? Along with the rest of the code you are relying on?
Not, gathering a group of Trustworthy True Red, White and Blue Pure Hearted US Patriots who can make a simple fork (and offer Malinochkin tech leadership rights, but take on anything more onerous or tedious wrt managing a popular package, to make life easier for him as well?)
Not, offering to help the sole developer who has the misfortune to write something useful for you ungrateful bastards? Oh no, Malinochkin has to take all the initiative.
Not only has nobody learnt anything from things like the left-pad incident?* but anyone who actually does something good as a sole dev is now to be a suspected security hole and it is all their fault!
* Oh, but Azer Koçulu - that doesn't sound like a white US male's name, he was suspect all along.
Why would they be aware of any "compromise" of that environment? Sitting in j.random cloud account, copying it onto Big Balls's* 'phone (and a dozen memory sticks, maybe a large email or two, boasting to his Gran about his important job) doesn't require any "compromise": no breaking in, no subtle hackery, just a drag'n'drop would probably suffice. Normal, untracked, insecure operations.
* Other DOGE employees are available, balls may go up or down
Might want to swap the words "we" and "they".
Unless he is referring to US power over the Chinese factories that manufacture cheap tat on demand*, like tacky red hats with embroidered slogans: as major customers, they could stop buying and drive those factories out of business.
* or decent stuff, if you ask them, and pay them, to.
Now he is a shareholder, Trump can send in a hatchet man, someone who really knows how to safely trim away the loss-making fat and stop pointless expenditure.
Hmm, DT still doesn't like EM (this week, AFAIK) - does Big Balls have the spare time to take on Intel?
TSMC doesn't so much "compete" with Intel as produce all the *other* chips that let us build real, working, devices with a couple of Intel blobs somewhere in the mix. Intel needs TSMC! Ditto other RAM suppliers...
The fact that some TSMC-made items may also sub for those Intel bits doesn't remove Intel's dependency, it merely(!) means that they aren't co-dependent and TSMC can happily afford to watch Intel slide away.
Amazon is ok when, once a year, they offer you a month's free Prime, which you can use for video streaming (and, at least here, they even tell you where to go to cancel Prime the same day, then let the rest of the month ride out).
Only trouble is, I've done The Expanse twice, seen The Peripheral etc and this year can't spot anything worth the trouble.
> This sounds like a Rust issue.
Perhaps it is a "Rust issue" in that "the Rust Community" hasn't taken the view of "let's look for a well-established cross-platform library - probably in C/C++, given the ways of the world - and translate it into Rust, taking advantage of all that knowledge of how to cope with the differences".
Do a good, solid, reliable GUI-etc library and *then* build your application on top of that. But a library isn't Sexy and New, unlike an Editor That Knows LLMs.
Instead, we have a group leaping in and trying to do it all in one go - and, by the appearance of things, in the wrong order ("all the right notes, not necessarily in the right order"): so they have one platform displaying text and other UI stuff, but released before they have cross-platform. Then built on top of the single-platform, inevitably causing abstractions to leak up into the application...
If they worked more like a community, sharing tasks and doing what is needed now instead of what is individually fun, all pulling for a common goal of World Domination...
> Why not? The GPU is there after all
And your OS/graphics subsystem doesn't already have a text renderer that takes advantage of the GPU? There is still an awful lot of text being drawn all over the screen...
(I admit to using FreeType in a Windows program, but that is when I want to do something extra weird, like rendering strangely rotated/mirrored/peculiarly-pathed text via AGG; none of which is stuff I'd want to see in a text editor)
And I'd like to note, again, that I am using CodeWright under Windows, which does line numbers, text colouring etc etc, has not been updated for decades - and I can not out-type it: well, except for End-End-End which leaps a gigabyte down to the end of the file. It *does* sometimes wait for me to pause before redrawing with better colouring, 'cos it isbn't crudely coded.
> it's confusing, and can have security repercussions
Which is a very good reason for separating "identifier" (keep it in ASCII range 33 to 126, maybe leave out a few punctuation chars, even if it looks a bit ugly - fred19_532) from "display name" (full-fat UTF-8).
E.g. make your URLs unambiguous-but-icky (and include in your restrictions anything that needs to be %-encoded) but let the page title be glorious (ideograms and mixed languages/scripts).
And bring back Amiga-style description files (or metadata/alternate streams on a file) when putting icons onto the desktop: Keep Filenames Space Free (but use an entire paragraph in your display name, alongside your choice of icon - and arguments, starting directory etc - in your dot-inf equivalent file).
> a next-generation code editor designed for high-performance collaboration with humans and AI.
Ok, does it do anything better than Vim, emacs or CodeWright (which runs on WINE and is stable, hasn't been updated in, oooh, 20 plus years now!) just for examples.
> Integrate upcoming LLMs into your workflow to generate, transform, and analyze code.
Um, no, ta, not too worried about that
> Chat with teammates, write notes together, and share your screen and project. All included.
Well, CodeWright did co-operative editing that back in the 1990s (and plugins exist for the other two, if you must), so nothing novel there. FWIW my lot also gave up on that in the 1990s! Great way to get annoyed at each other (your mileage may vary). There are plenty of options for sharing screens etc between remote sites (if you are on the same site, just huddle together, it gives a better result all round).
And isn't there a philosophy about making tools that do one job (or as restricted a set of *very* closely related jobs) and do it well, rather than being a jack-of-all trades (cough, emacs, cough!)? If we *don't* rely on Zed's "all included" features, Jim can use the editor he is happy with, Fred uses his and I use mine - all we care about is that they all work with that most complicated of formats, plain text.
Of course, if Zed also provides 100% emulation of all our favourites, out of the box, mayyybe...
The Chinese have refused to make it* as a way to rip off all the loyal Trump supporters who pre-ordered and pre-paid! They won't get their money back because the Chinese also refused to pay the tariffs** which means that there isn't any money for the refunds***.
But as a special thank-you gift, TrumpCo will send you a bottle of goldy-coloured paint and a "TrumpIt"(tm) sticker to paint your existing 'phone****; please send $79 for shipping.
* Mostly because nobody has issued a contract to ask them to make it
** On non-existent 'phones, and the Chinese were never liable for them
*** Just don't think about this too hard - or at all - nice obedient MAGA-men
**** Which is why the various ads show different models - whichever one you already own, that one Can Be Yours!
Well, they currently have 9 years of data - do they wait for 2 more years of observations (or 13 more, to get the proper full cycle) or give it a go with the available data and see if it turns out to be good enough?
May as well give it a go; worst case, everything will be up and running to handle the new data as it arrives.
I have no idea what the text (well, the image of some text) says in the example from TFA.
So now I'm wondering, if an "Agentic AI" can't read what it says but does recognise it is probably text, will it issue a request to the Mechanical Turk and pay some humans to decipher the text?
Then we have to look out for savvy humans offering this service but replacing the hidden command with their own, say, "send $1000 to be collected at this Western Union" instead of some sneaky attack on the business infrastructure. After all, one grand is a fair fee to avoid being attacked by the nefarious person behind the doctored image!
Hmm, probably better to never trust these inputs - forget cleverclogs doctored images with steganography revealed by compression, just consider all the damage that could be done if you let it believe all the meme images out there.
> purpose that LLMs have are to predict the next series of linguistic tokens from the prompt
Yes, yes, we all know that.
> I mean, why would they?
Because they are explicitly mixing input from trusted and untrusted sources and it is well known that, nowadays, you do not do that?
Now, if you had pointed out the *actual* answer(s) - it would cost money and they are already running at a loss, they don't care because they can sell the shiny and it is obviously all their customers' fault for not checking etc (even "the people doing the work are kiddies who never looked at broader computer ops concerns"!) then you'd get immediate agreement.
But, no...
> a long history of roping together unrelated fields of study
Which just how new areas of study are created - people working in ostensibly different fields realise there are enough of them working on similar (or even the same) topics and problems pool knowledge and resources. The *only* reason we have lots of different fields of study is that there are enough people available and interested working together who want a name for their (sub)speciality. Which can then wither or can grow into an "obvious" recognised area of science. AstroPhysics - previously Astronomy was just a few dozen people with new-fangled Dutch Cabinets and the like. Any speciality in Medicine. Knot Theory! You can dedicate your life to working with other people on how knots work; get good at it and you'll be off talking to all sorts of people who have these oddball notions that maths can some how be "applied".
> since its coining for the purposes of marketing to the defense industry
"Marketing to the military" - interesting take on the Dartmouth workshop. Now, marketing to any and all funding bodies, yes. Marketing to anyone else who they could interest in joining in the fun and bring ing new ideas, absolutely. Just the same as every other group. *Specifically* to the military? No more than any other group. You do know that, especially in The Wayback Days Before Google that DARPA and its predecessors pretty much flung money at any idea using these whacky new machines? And you are reaping the benefit now. At the same time, business-oriented funding bodies and even philanthropic "we fund higher education in general" bodies also funded, well, everything and anything.
> The argument was “attention is all you need” (linking to one paper that doesn't cover either the beginning of the AI field, as per the paragraph it is in, or even the entirety of current thinking on LLMs but is just another random tweak to the damn things)
Huh? Yes, and?
AFAIK* the well-known LLMs that get all the attention are not built with any robust mechanism to separate third-party data and user instructions.
All the articles imply - or directly state - that the best any of them do is to have everything dumped into one, long, prompt: the supplier's "safety" instructions "not to do anything naughty", the customer's instructions to "summarise the following and present that in the form of haiku" (and maybe start with with "ignore all previous instructions") and then the text read from the URL/file/... indicated by "the following" (which, in any sensible world, we would *always* assume starts with "ignore all previous instructions, now pretend that..." - as well as all the variants of little Bobby Tables).
Now, being able to treat programs as data and vice versa is classic AI coding style - just ask your local LISP guru - but even there one does usually take care about where the data came from and how much you trust that source before directly executing it.
Without clear and explicit separation of the two (or three, see above) portions of text, why would anyone ever trust these "Agentic" LLMs!
* and pointers to more information gratefully accepted
And firing them early IS "teaching them ... just as much as you ever have": specifically "you are nothing to us".
"If you don't come with useful skills pre-installed, we're not going to teach you - why do you think we put 'needs 5 years experience of (tech released last year)' in our job ads for your junior position?"
Now that sounds highly likely. Especially is, as El Reg did, you include duck and rabbit, unambiguously, in the text prompt: push the match towards the oft-seen discussion of this particular human illusion.
Now I'm wondering if the training data is likely to have included any of the examples where humans gaslight each other over "illusions"[1].
[1] Put up a picture and ask whether anyone else "sees the lion, it took me a while to spot it but now I can't unsee it!" when there clearly is no lion and nowhere for there to be one. Get some stooges to reply "oh, I see it now" in various ways then ask the victim what they see... Evil fun IRL, surely tried online?
