Posts by that one in the corner

Microsoft and GitHub are still trying to derail Copilot code copyright legal fight

All for analogies, but can we bit a bit accurate (or at least explain our analogy)?

> Model training might be fairly described as a process of encoding the whole work/s, rather like jpeg encoding a copyright artwork

Uh, no, sorry, that is taking analogies too far. And it most certainly isn't a "fair" description.

> The fact that it produces imperfect copies of it's source is an encoding limitation, the same as other highly compressed copying and storage methods.

Bad analogy. The lossy nature of JPEG is well-defined in the quantisation step and even the worst over-compressed JPEG still spits out a recognisable copy of the *whole*. If you can suppress the ringing, a lossy image compressor will literally spit out a copy "as though you were standing x metres away" and can not resolve the high frequency components. Still designed with the sole purpose of representing the whole.

Note that we only apply lossy compression (of a form arguably similar to JPEG...) to audio and visual, not text. "Strip out the high frequency components" from text and you get gibberish, especially when trying to compile the results.

> The model itself is a copyright violation

I have argued before (and it was not well received, sob) that the model *might* be compared to the Huffman tree you can find in many compressors: as you traverse the tree, you hit a leaf node and spit out the tiny (just a few characters, maybe a word) on that leaf. The same tree is used (in this analogy) and traversed a *lot* to finally spit out a decent chunk of material. In the decompressor, it is the input bitstream that causes a specific traversal - one bitstream outputs "Moby Dick", another " Life of Brian" - and does so reliably and repeatedly.

It is those input bitstreams, the traversal pattern, that are the copyright violation. Not the Hufman tree.

In the analogous traversal of the Copilot (or other language model), the model itself should be as innocent as the simple tree (just a ot bigger and mor graphy than treey). So it must be what is guiding the traversal that is to blame, much as we did the bitstream above? But the traversal process is stochastic: until you finish it you have no idea precisely what you will get out. Unlike the compression example.

UNLESS, of course, in Copilot "the model" is a really crap model and the whole thing is not behaving like a "well behaved" deep learning model ought to and Copilot is all just a fake.

Why isn't there enough strong, clear, evidence from the plaintiffs?

I've noted before that there are reasons for models like Copilot to spit out "recognisable" code, most of which boil down to a naff model (shoved out too early, surprise) and/or the monkeys on typewriters: there is a random component selecting from "I've just put out A now that is usually followed by W, X, Y or Z, roll a die...", do that often enough and you'll be able to see strong matches with inputs.[1]

We've heard lots of stories about people seeing their code being spat out again and maybe it happened.[2]

But "maybe" isn't useful in court.

Why have the plaintiffs gone in with such feeble evidence, against a monolith like Microsoft?

[1] In fact, every line Copilot spits out *has* to be recognisable, or it won't compile (assuming that the complaints are over code that does, in fact, compile). Ah ha, that is a for-loop, I recognise it. Obviously, we are concerned about non-trivial, longer, sections of output.

[2] Sorry, I have to say "maybe" because so far I have not seen a fully-described example of it happening, complete with the whole conversation. Please, if you have them, give citations.

Copying or transformative?

From https://www.copyright.gov/circs/circ14.pdf via abend0c4

> "Only the owner of copyright in a work has the right to prepare, or to authorize someone else to create, an adaptation of that work". It also gives as an example "A new version of an existing computer program".

I can not copy a work but I can study it, see how it was constructed and use those methods.

For example, I can study the works of a living artist, a painter in oils, and learn from them how layering of colour works, when you mix paints to match a shade and when you place unmixed shades side by side to create the impression of the mix.

I can take those techniques and create a whole new piece, applying it to a subject that the other artist would never touch.

So long as the original allows me - and probably everyone else - to view his works, this is all perfectly legitimate, even fully expected, behaviour. I do NOT need the artist's exolucit permission to do so.

A model like Copilot is *meant* (if done well) to act similarly: so long as the source code is made available to read, it is acting entirely legally and your quoted paragraph simply does not apply.

The question you (and the court) have to apply becomes: *is* Copilot acting as it is *meant* to do? Or is it just a bad/incomplete implementation of the idea of something that can synthesise code? If it *can* synthesise, at what point does it inevitably spit out "recognisable" code because that is the only known (or simply the "obvious to any practitioner") way to do the job?

Microsoft signs 1.5 million seat contract for Office 365 and more

Re: Windows?

> Orifice 365

As used in dental surgeries, Ear Nose and Throat and, of course, Proctology.

Re: Headline?

The purpose of a headline is to pique your curiosity, so you want to open the article and see if your guess as to who this contract is for was correct.

1.5 million seats? Could it be the Burundi Civil Service? Nah, that is a silly suggestion. The EU Cross Party Conference? Oooh, oooh, the North Korean Secret Police: "Open the online form in Word 365 and Bob was your uncle"!

A great parlour game, fun for all the family.

Re: $940 million for a million and a half users ?

We just need to find that Bytes Software Services is a subsidiary of Capita and it will be the perfect exemplar of Government Spending.

VMware, AMD, Samsung and RISC-V push for confidential computing standards

But how could companies collaborate in a "multi-party computation and analytics" project?

They would have to come up with some way of copying each others' datasets onto all those machines. Maybe invent some kind of "mobile storage medium" that exists outside of the datacentre?

You know, that is a crazy idea, but it might just be crazy enough to work!

Forget these apps and AI, where's my flying car? Ah, here's one with an FAA license

Obviously designed to appear in the next Mission: Impossible film

Tom Cruise will fit inside, with room for the nuclear bioweaponised list of enemy agents in the set next to him. He already has experience with a gimballed cockpit from "Oblivion".

The whole thing is just an elaborate action movie teaser.

> Potential market is going to be small.

"Small" - I see what you did there!

Cops told: Er, no, you need a wiretap order if you want real-time Facebook snooping

Turn over access to a user's future posts

Facebook is employing Time Lords now?

What if I then choose *not* to post the message, does FB vanish forever into a temporal paradox?

Or having vanished, will it be immediately replaced by something even more bizarre?

Microsoft puts profanity filter on %@!#ing Teams transcripts

Well ******* along with ****** ****** **!

This is simply *** ****** *** ******* in the end, and we will all realise that *** ********* *** **** on Wednesday!

Pah. Get ***** *** ***** bucket of ice.

Ripoff Vuitton handbag smaller than a grain of salt fetches $63,750 at auction

Re: It does not look to be hollow

Just so long as you can dance around in a disco or use it to slap someone at dawn[1], it qualifies.

[1] not much heft in this one, but those edges look razor sharp.

Quirky QWERTY killed a password in Paris

Re: Paris...

We had a flight booked to go from our regional airport (less than 4 miles from home) to Schipol, "suffer" an 8 hour layover (we had plans for that) and meet the rest of the group for the connection that evening.

The travel agent was really happy (fine for him) when he managed, at short notice to us, to rearange the tickets "let us meet the whole group early in the morning at Heathrow, so we could all go to Schipol together, no bothersome layover".

Cue one hurried booking for an overnight stay near London (so cheap!) and a few hundred extra miles on the car, drive to drop the car off in a free spot we have access to, tube back out to Heathrow, the reverse on the way back. Hope the travel agent was *really* happy!!!

Re: Paris...

> I saw daylight in the car from the hotel to the workplace

The hotel would like to apologise that they could not supply a car with blacked-out windows, as requested by Accounts when they made the booking. They said something about "trolls"?

Re: All your QWERTY belong to us...

You noticed he only said "servers", not desktops?

You know, servers, the caged beasts that are kept alive by only a small, select, group of almost priestly status?

A chosen cadre who are each others' support lines and may even be shuffled around across the EU to help troubleshoot and fill in gaps?

And who may eventually find that consistency in their machine chapels outweighs the frustration of being forced to hunt and peck? Who knows, going slowly may help them reflect: is this really supposed to be "rm *"?

Rocky Linux claims to have found 'path forward' from CentOS source purge

Re: Popcorn time?

PS Yes, a nowadays a printout does count as "machine readable" - at least, RH would probably argue that and demonstrate using a recent model document capture device; or they could be really mean[1] and print out in a bar code, like we used to get along the edges of computer magazines.

[1] 'cos the hardware is rare as hen's teeth

Sod RH but GPL needs testing in court

and not just in the US.

You may have spotted from some of the above comments that I am not at all convinced that RH's stance is breaking the GPL as it stands. Clearly, many disagree - I'd love to disagree with myself on this, but so far none of the argument[1] has been convincing. RH have followed the Capitalist Dream, a loophole that they have driven a railway spike into.

Aside from walking away from RH, the only other way this can progress is to test it in court - but has the GPL already been tested enough in court that taking on this specific pont of interpretation and "in the spirit of" will be *the* point challenged or does it risk upsetting the whole apple cart? IANAL and hopefully have missed reports of GPL being demonstrated to be robust in court.

Walking away from RH poses - interesting - challenges. As individuals, we aren't RH's customers (not any more - I did buy a boxed copy in the 90s to run on the 486, but never even upgraded it).

How about discussing[2] positive ways of weening their actual customers off RHEL? Some comments have come from people in the depths of such companies - what would affect their choice to renew with RHEL? Anything better than passing FUD up the chain of command would be preferable...

[1] yes, argument, singular.

[2] yeah, this comment is over a day late and won't be spotted by many now, ah well

Re: Disingenuous

Welcome to Capitalism. Sorry.

Re: Popcorn time?

RH do not have to allow you access to the portal in any situation - if they decide not to like you for any reason, they can require you to make a formal request and include a "nominal" fee (including media costs and postage) before they send you a copy of the sources.

If they are feeling really pissy, you may get some of the material as printout or eight-track tape.

They are also under no obligation to provide sources for non-copyleft binaries.

Re: Popcorn time?

*sigh* It places no restriction - positive or negative - on your use of the material you already have.

The licence has nothing whatsoever to do with anything else other than *that* copy.

What happens with respect to anything else, such as whether you are able to get an new copy, is never once mentioned in the licence and is out of scope.

Until you can get a court to agree with your extended interpretation. So go to it. It needs testing.

Re: What they've achieved

> I, for one, am not going to bother learning Red Hat specific stuff

> I know that some software vendors insist on a particular OS, but we can now say "no, that's not going in our network" - they can either lose a sale or agree to installing it on another OS.

Absolutely, go for it.

RH isn't breaking GPL but that does *not* mean we have to like them and continue to buy from them.

>> Gpl2 has nothing to say about programs other than the one you received. It does not refer to , or claim mastery over different programs (which would be absurd).

> But it does say about the distribution of the source code to that program, and the rights you were given to use that source code.

Be careful with the word "that".

Try phrasing it as (very clunkily):

> But it does say about the distribution of the source code to [the copy you already have of the] program, and the rights you were given to use that source code [matching the copy you already have].

Otherwise we could misread your "that" to be referring to the "future versions", which I'm sure you never intended us to do.

Now, you have full control over the copy you already have and nobody can take that away from you, nor is anybody trying to take that away from you.

But unless you have a contract, totally separate from the GPL, you can not coerce RH into giving you a coy of an executable, let alone sources, for any *future* version.

> If Red Hat violate the GPL, which clearly they are

Clear as mud.

Maybe we do need to hope for a court case and the disagreement over that point will be settled; at least in one jurisdiction.

> how can RH do this and not have invalidated their licence to huge chunks of their distribution?

Go back over the sections you have quoted. If you have a copy of executables for v1.2 then you have the right to get a copy of the sources for v1.2[1]. GPL satisfied.

Nowhere does it say that RH is then required to supply you with executables for v1.3; even if RH create v1.3 by modifying the GPLed sources for v1.2, until you get a copy of the 1.3 executable from RH they do not need to let you have the 1.3 sources.

And if you no longer have a subscription to RHEL there is no requirement for RH to supply the executables of v1.3 to you.

RH have changed the pricing and are not making friends, but their product (the subscription, that is), their pricing. If enough customers stop paying then RH may change their minds (or not) but inaccurate arguments about licensing help nobody.

[1] but only of the GPLed bits and any scripts needed to build them - you don't have a right to, say, a free copy of the compiler used to make the executable - I can send you all the bits for a GPLed Windows executable that I built with VC6.. Nor do you any rights to sources from RH for non-copyleft programs, whether you have the executables or not.

> if it modifies the code the resulting derivative is still covered by GPL, must provide the modified code if required and should not add further restrictions.

True.

And the definition of "if required" by GPL is if you have a copy of the executable of their modified form. Which you don't because you lost the subscription and RH never sent you a copy of those executables.

What RH has done isn't making them any friends but it also isn't breaching GPL.

Re: Popcorn time?

> The question here is whether the business relationship under which RH provide GPLed code implying such a restriction on their customers.

It doesn't restrict what you do with the copy you already have, do the GPL terms are satisfied.

It does restrict whether RH wants to let you have a copy of the *next* release - which is nothing to do with the GPL, it is entirely about business relationships.

> I'm not a lawyer but as I read it the GPL requires you not to place any restrictions on the right to further distribute the source code.

They don't. You have the sources to v1.2, go right ahead and distribute them

RH will then choose to not send you the executables for v1.3 - as is their right - and without those executables you don't have a right to get the sources for v1.3 from RH. The End.

But you still have v1.2, executables and sources, and can distribute it at will.

BTW if you didn't get your copy of the v1.2 executables direct from RH then they have no liability to provide you with the sources.

Re: Ignoring the big issue

All of the above bears repeating, many times. So I have :-)

Please excuse if my paraphrases are not as neat and noninflammatory as your post.

Re: Ignoring the big issue

> GPL licences, etc. need to be amended to stop future foul play

So the licences need to say that if you managed to get a copy of version 1.2 of a program from a specific supplier, that supplier is then bound to give you a copy of v1.3?

If you violate the RHEL subscription terms, you have decided to forego your copy of v1.3 - you still have v1.2 in all its glory.

What licence has ever said that you get the next version? If you want to say "GPL" then please quote the clause (you can easily get a copy to quote from): nope, it days that if you already have a copy of the v1.3 executable you can get the sources (for a reasonable fee - which can be via a subscription) but it doesn't say the vendor has to give you that executable in the first place.

Although I do agree that a court case would be good, as that does tend to make things clear to everyone (even if it also makes the lawyers fatter and more prone gout or diabetes).

Re: Dangers of speed reading before I had to dash out

> You don't have freedoms 2 or 3 because your subscription will be terminated if you exercise them.

Nonsense.

You have those freedoms and keep them forever.

If your subscription is terminated then you still have that version and can continue to use and distribute it at will.

What you lose when your subscription goes is the automatic provision of an updated copy from RHEL.

There is absolutely NOTHING in the GPL that gives you the right to receive any source code mods if you happen to gave an older copy of the sources or executable.

IF you get an executable THEN you the right to the GPLed sources.

BUT you voluntarily[1] gave up the subscription that would get you that executable. The End.

Your access to Freedom 0 is also almost totally unlimited: you are free to take a copy of a GPLed executable out of the RHEL distro and use it anywhere. The only way this is being limited is because you are voluntary choosing to run the *whole* RHEL distro, which is *not* under the GPL: the subscription management software and any number of non-GPL code.

You don't *want* to take the GPL exes out and try to use them without the entirety of the RHEL distro because that is too costly for your purposes. That is your choice to determine and has no bearing on whether you would be within your licenced rights to do so.

[1] yes, voluntarily: you knew you were breaking the subscription rules and chose to do so.

Experts scoff at UK Lords' suggestion that AI could one day make battlefield decisions

Re: What about AI camouflage?

I believe we have a recording of the last time that sort of subterfuge was attempted:

Henry Crun: It's much too dark to see, strike a light.

Seagoon: Not allowed in blackout.

Minnie Bannister: Strike a dark light.

Seagoon: No madam! Madam we daren't. Why, only twenty eight miles across the Channel the Germans are watching this coast.

Henry Crun: Don't you be a silly pilly policeman.

Minnie Bannister: Bravo Henry.

Henry Crun: Pittle Poo.

Minnie Bannister: Pittle Poo. They can't see a match being struck.

Seagoon: Oh, all right.

FX: [Striking match - bomb whistle - explosion]

Seagoon: Any questions?

Henry Crun: Yes, where are my legs?

Minnie Bannister: Where are my legs?

Seagoon: Now are you aware of the danger of German long range guns?

Henry Crun: Mnk ahh I have it! I've got it, I've got the answer. Just by chance I happen to have on me a box of German matches.

Seagoon: Wonderful! Strike one. Ha, they won't dare fire at their own matches.

Henry Crun: Of course not. Now...

FX: [Striking match - bomb whistle - explosion]

Henry Crun: ...Curse... The British, the British!!!

Chinese balloon that US shot down was 'crammed' with American hardware

Re: Red Spy Balloon Phoning Home

> there's a REALLY big balloon in the way

Put the silver patch on top of the balloon!

Come on people, don't let the nay-sayers win! To the garden sheds of the world and we can figure out how to make this gasbag heliograph work.

Re: Notice how....

> 'crammed with American components' because it had a Raspberry Pi or something similar in it.

*American* components?

Re: Notice how....

Send in Lady Constance de Coverlet, she'll keep them occupied and unable to cause any more trouble.

Google accused of ripping off advertisers with video ads no one saw. Now, the expert view

Facebook cutting off news

This has me baffled: what difference would that make?

I joined FB when lockdowns came in (Concellation and family messaging kept me sane - ish), so that is - years ago now. Yikes.

Just where is FB supposed to be providing news, other than what is going on on their own site or paid "announcements"? Seriously, I asked the family and none of us have been shown any actual real-world news by FB.

And yet we hear claims like "people aren't reading the news sites, they get all their news from FB" - do they mean "they are only paying attention to echo chambers of groups & individuals on FB and calling that news"? Or is there simply something obvious that I and all the other Corners have missed about we are "supposed" to use FaceBook?

Microsoft's GitHub under fire for DDoSing crucial open source project website

> "Oh, but why don't they just host their project on github!"

Microsoft just *love* people who complain like that!

> Some would rather host their projects themselves

Or even just use another third-party project hosting site.

Come on World, what is wrong with you, *every* project should be on Github! What Is Good For Microsoft Is Good For America etc etc.

The 100s of architectures refers to the ffmpeg builds, not specifically GMP.

It *may* be the case that the only reason ffmpeg "needs" to be rebuilt for each architecture is only because that allows the GMP build to optimise itself, but that seems unlikely. From GMP's point of view, many of these "architectures" may well be the same, just some other part of ffmpeg wants to know the difference (e.g. forget about x86 variants, where does this OS variant store this file?).

> In many CI setups caching is standard and, indeed pretty much a requirement for anything using images from Docker hub. GitHub doesn't seem to have this and it also doesn't seem to have any kind of rate limiting

True about the lack of rate limiting.

But whether or not Github provides caching would not be relevant in this case - if you read the (even updated) build script that the article points at, the guy was explicitly deleting files and then fetching them again using a command line Mercurial client, which he replaced with explicit use of a wget to an FTP site. Those would bypass any Github caching.

Re: Just mirror it on github and pull locally

He was managing to run a Mercurial client to get the sources, which rather implies he could have been using it properly (just checking for updates instead of explicitly chucking away the local copy every time the build run). Even pushed from his local copy into a git repo and made the script use that (hence all the copies of the script...).

Saying "Github doesn't do Mercurial" - well, it is true that Github doesn't provide all the tools to work with a Mercurial repo in the same way as it does a git repo, but this guy already knows enough to use a command line Mercurial client to pull the library down, it can't be beyond him to add in use of the git command line tools as well.

Re: This isn't exactly Microsoft's fault.

> maintaining an attractive nuisance

Loving your characterisation of Github users as having the awareness of seven year olds. That does strike at the heart of the problem for this ffmpeg builds guy (including his response of "you should change the cron timings".

Re: Whatever happened to local caching?

> I would frequently come back to a project from 6 months ago a

There is a teeny, tiny, itty bitty difference between cleaning out every six months and starting from scratch every time a batch file enters a loop to trigger another build.

And complain to whoever wrote the stuff that is changing the environment (or requires you to change it) in a way that can affect another project: that is just bad work on their part. You really should not need to have to resort to Docker or VMs or anything else to keep a working environment. I know a lot of projects piddle about like that, but maybe we can get them to grow up one day.

Whatever happened to local caching?

Why are builds like this continually grabbing from *any* server?

> This build was configured to run parallel simultaneous tests on 100 different types of computers/architectures

And all of these architectures need their own copy of the GMP code? Sure they aren't just using the same sources with the usual flurry of #ifdef's?

For that matter, what about the rest of the ffmpeg and dependencies source code? Are these being grabbed afresh for every single architecture? That certainly seems appears to be the case, if I'm reading the updated script correctly[1].

The DDOS aspect not only appears to be built in, but why aren't all these copycat builders pointing out that it is wasting *their* resources as well (time, bandwidth and electricity)?[2]

[1] and the author of that update wants to have a careful think - any guesses whether running a fetch from *any* version control server - in this case, a Mercury server - is going to more costly for both ends than a wget of a single tarball from an FTP server? Or could a version control client perhaps pull down *just* recent changes - oops, no, we deleted the local copy, can't do that!

[2] oh, it is just Github's bandwidth and electricity? Well, waste away!

Mummy and Daddy Musk think Elon's cage fight against Zuck is a terrible idea

Get Musk to schedule the fight

When has he ever hit one of his own announced timescales?

"By the end of 2023 you will be able to buy a ticket for this match!"

"We expect to be able to land not one, but two, tickets by third quarter 2026"

"It will be financially irresponsible not to own one of the tickets in 2036"

...

"Speak up! Tickets? Fight? We did that last Thursday, didn't we? Who are you?"

Five billion phones are dead in drawers – carriers want to mine them

Re: Mine my drawers?

They used strip-mining...

Re: Mine my drawers?

Well, they've mined out the loincloths, bloomers and clouts...

Bosses face losing 'key' workers after forcing a return to office

Re: employees were happy (31%), motivated (30%) and excited (27%) to be in the office

Indeed - it doesn't matter if those are "percentages of all the employees are happy..." or are supposed to be "of the companies surveyed, this percentage said their employees were..."[1] it doesn't sound like many employees were "excited" to be in the office.[2]

[1] don't you just love stats reporting? Precise numbers, waffly descriptions.

[2] who is excited to get into the office? Is it Taco Tuesday? Actually, that *does* fit the figure quoted (with 7% left over for those who - are nipping into the supply cupboard with Brenda from Accounts?)

Microsoft, OpenAI sued for $3B after allegedly trampling privacy with ChatGPT

Re: Have I got LLMs all wrong?

You are correct in your original thinking[1].

This claim of just dumping out copied text en masse keeps appearing: it seems to be very appealing to some (or maybe they just can't their heads around things like MCs in the first place?).

There are plenty of ways that you *could* get "accurate" personal data out, starting with pure chance: these models are being queried many, many times and someone who actually lives at 23 Railway Cuttings, East Cheam is one day going to see that appear in a response. Or maybe your personal info is actually spread around the Internet so much that the model has pushed the correlation of "Fred" followed by "Bloggs" followed by "Mastercard 1234 5678..." up to near certainty!

You can also get *inaccurate* personal data out: oh look, ChatGPT "claims" a radio show host is an embezzler. The LLM gets it in the neck both coming and going.

[1] Okay, the process is more involved than a Markov Chain (and a lot messier and less comprehensible - less explanatory power - but MCs are a good starting point)

light on specific instances of harm.

Great. The *one* thing that is *required* to demonstrate that there is a claim for damages is replaced by page after page of "maybe". Do they *want* this case to fail?

You could almost imagine that these pseudonymous plaintiffs are really named " OpenAI" and "Microsoft".

After decades contributing to science, John Goodenough powers down

Re: Really sure he's absolutely gone?

Just put to one side for a day or so to rest then try again, you can get a few percent extra life depending upon the chemistry. Worth a try.

EU launches 4 testbeds to put AI tech through its paces before it goes to market

> Providers of AI-driven solutions get the opportunity to test their products in real environments to assess if they meet the customer needs

Shouldn't the "solutions providers" have done that already? Verified against Requirements Specs and so on?

> TEFs support the new legislation, which called for regulatory sandboxes. Through rigorous testing, developers can ensure their technology is trustworthy and complies with laws so it can be deployed and made commercially available

Ok, that bit is more sensible - hopefully the TEFs do what all (good) test labs do and issue a certification (with the very carefully worded list of what tests performed and results obtained).

But possibly not a bottom line of "therefore this is safe to release" just in case it goes all Skynet after 18 months and they don't want to shoulder any part of the blame ('cos who can predict what it'll do with unpredictable User inputs?)