Posts by Michael Strorm

Copilot tricked into telling reseachers how to hack itself

Looks like it's once again time to...

...quickly slap on yet more retrofitted "guardrails" (*) as a purely reactive response after the horse has bolted.

A sticking plaster applied to the latest hole. One designed to make it look like they're doing- or *able* to do- anything about it.

When in fact, the actual (and much greater) problem is that LLMs themselves are fundamentally- and unfixably- insecure due to how they work at the most basic level, i.e. the fact they don't and can't cleanly separate internal and external data and instructions. And that there's no way of getting round that without completely hobbling the general-purpose usefulness that they've been sold on.

As such, the real problem is that there are a near-infinite number of potential holes, and while the AI companies can "fix" each hole after it appears, someone's going to figure out a way to get round it sooner rather than later and the whole cycle starts again.

(*) As I've said before, "guardrails" turned out to be an ironically appropriate metaphor, since real-life guardrails generally make clear where one should and shouldn't be and stop accidental straying... but generally present little obstacle to anyone determined to *intentionally* climb over them.

Less than a year on, Microsoft tells Mico to pipe down

Couldn't have told you that his name was Rover, but I do remember that dog. Mainly because after choosing to "turn off the animated character" (or similar wording), they acknowledged this choice by, er... having an animation of the dog running away over a hill.

Which was nice.

Virgin Galactic flights stay paused while ticket prices head for the Moon

"No single issue is driving the schedule push. Rather, we have experienced modest time duration extensions across hundreds of relatively small but important installation tasks involved in the first build of our new spaceship."

How fast is that thing going?!

Trump wants to grant private cyber firms a license to hack back

Re: Oh goodie ...

There's no need for sarcasm- I'm sure Trump's cyberwar will be every bit as successful as the actual war he got himself involved in.

Still, it can't be denied that one *did* have the intended short-term effect of distracting from the Epstein files, even if it very quickly provided him with a bigger problem he can't magic away with the usual bloviating bullshittery.

Of course, the irony is that the Epstein files had in turn themselves been a demagogic would-be attempt to smear his Democrat opponents that ultimately came back to bite him on the arse.

Hmm, I'm spotting a pattern here.

Twitch feeds your streams to Amazon's AI unless you tell it to stop

"Twitch has helpfully enabled the setting by default"?

Well they would, wouldn't they?

Deepfake hiccup unmasks suspected digital certificate fraudster

Re: Perfect way to trump the investigation

In his home country, perhaps. But according to the article, the crimes were taking place in Murcia, not 'Murica.

OpenAI pledges to add Astra security as Anthropic loosens Fable's leash

Re: Storm in a Teacup

The interrobang criticism wasn't even really aimed at you specifically, so much as at its (rare) use in general. (I suppose in your case it at least fits into the weirdness of the account style).

That said, who wants to read a formulaic ChatGPT response? Doesn't that dilute and defeat the point of your own established "AI" style anyway?

Re: "Adding" Security

> "‽"

Ugh. Am I the only person who dislikes the "interrobang"? It's one of those things that no-one seems to care about except to appeal to a certain type of geek- who invariably uses them in a manner that draws attention to themselves (and how clever that person is) at the expense of distracting from the intended context/usage.

I was thinking about this recently, and even the rationale for its existence is a contrived solution to a non-problem- the combination of a question mark and an exclamation mark.

You mean like the "?!" pairing that most non-smartasses use and which does the exact same job in a perfectly satisfactory- and less contrived or distracting- manner?

Trust an advertising executive to come up with a "solution" to something no-one thought was a problem in the first place...!

Re: "...what for human perpetrators would be computer crimes..."

Even if it's true that there *should* be no difference in the eyes of the law, the courts' lack of previous experience and precedent in dealing with the second scenario will likely provide the highly-skilled and resourced legal teams that a multi-billion-dollar company can afford with more than enough scope to sow confusion, fudge the issue and create sufficient doubt that this is the case to get them off the hook.

Conversely, for an individual defending themselves with limited resources, the risk is that a determined or vindictive prosecutor has a million different ways to trip them up and hang them out to dry.

This would be particularly the case in a country like the US with a "might is right" legal system strongly biased in favour of those who have the resources to defend themselves and against those who don't.

Ransomware gangs skip the CEO, head straight for the 40-something IT manager

"46-year-old [Gen-X] IT manager"

"Gen-X"? Barely, if at all. Depending on whether you take the cutoff date as 1980 or 1982 (*), someone who's 46 at present is either only Gen-X by the skin of their teeth or already a millennial.

Yes, many millennials- a term used until not so long ago as synonymous with "youth" even when that was pushing credibility (**)- are now fortysomethings, i.e. into what would once have been considered "middle age".

(*) I always used to be under the impression that 1982 was the most common starting date for millennials- which makes sense as it's the generation that would have come of age at the (supposed) "millennium", i.e. the year 2000. However, even Wikipedia seems to be inconsistent with its "Generation X" article suggesting 1980 is the widely-accepted cutoff, yet the "Millennials" article seemingly giving credence to 1982.

(**) It's noticeable that even the mass media seems to have finally stopped this a few years back when they "discovered" Gen-Z and presumably realised that even then the oldest millennials were heading towards forty and the youngest were approaching their mid-twenties, and it was getting a bit silly.

Meta latest to tell world its AI agent wandered out of test pen

Re: AI - Flex

On the contrary, that has already been accepted as the most likely explanation by a significant number of people here- myself included- for some time now. Even the likes of the BBC news website (which previously took events like this at face value) are now suggesting that as a possible motive.

This has happened numerous times before and as my comment above suggested, it's reached the point of cliché to the extent that I was already willing to dismiss this occurrence as "yet another PR flex masquerading as scare story/disclosure" (before I saw that OP had got there first).

Re: AI - Flex

If I was conspiratorially minded, I might wonder if they were crying "wolf" as an intentional attempt to numb the public to those actual dangers.

I wouldn't go so far as to bet my life on that, but given some of the things that might otherwise have been dismissed as paranoid that turned out to be true, I wouldn't bet against it either.

Re: AI - Flex

Was going to say "Yet another PR flex masquerading as scare story/disclosure?" but OP got there first.

Yeah, Anthropic may have started this gimmick- to the point it was already becoming a cliche of *their* modus operandi- but now everyone is doing it and even the mainstream press have started calling it out.

Facebook doing so now just looks like a sad attempt to jump on the bandwagon.

Bypassing AI guardrails is so easy a script kiddie can do it

I wouldn't trust the AI companies or their motives as far as I could throw them and nor would I take anything they said at face value. It's already pretty clear that the AI companies are (e.g.) releasing supposed scare stories about their agents hacking rival systems as a flex to show how powerful they are.

Still, I'm not convinced that's the case here- do they genuinely want the guardrails to be easily circumvented?

Or is it more that they don't want to admit that they genuinely *can't* reliably stop that from happening because of the way LLMs work at the most basic level, i.e. that they don't cleanly separate data, user input and instructions? And that all such "guardrails" can only ever be plastered on attempts that will never come close to covering all possible holes in something that is fundamentally- and unfixably- insecure.

"Bypassing AI guardrails is so easy a script kiddie can do it" and...

..."existing guardrails offer little resistance to operators willing to reframe their requests"?

Well, obviously.

I know I've said this before, but this sort of thing is *exactly* why the industry's choice of "guardrails" as its favoured metaphor ended up being so ironically appropriate for all the wrong reasons.

Real-life guardrails generally make clear where one shouldn't be and stop people *acidentally* straying, but they typically offer little resistance to anyone determined to intentionally climb over them. So... yeah.

And I strongly suspect this will keep happening because all such "guardrails" are a tacked-on, after-the-fact sticking plaster attempt to mitigate the fact that the fundamental way they work means that LLMs themselves are not- and cannot- be reliably made secure.

The AI bubble is already popping; we just don't know it yet

It's like raaaaaiiiiiiaaaaaain...

Situational Awareness are in the shit because they didn't sufficiently understand they were operating in a risky field and guard against that?

It's almost as if they were lacking some particular form of insught, but I can't *quite* put my finger on what that might be called.

As Larry Ellison bets the farm, Oracle says it loves AI-written code, just not in OpenJDK

Is this them not "eating their own dog food"...

...or is it the other way round in this case (i.e. not wanting anyone else to eat it)? I'm not sure.

Regardless, I always thought that expression was a bit flawed- I mean, if you literally made food meant for dogs rather than humans, you wouldn't really be expected to eat it for your own lunch, would you?

Unexpected item in the bagging area as Windows Activation error pops up at check-in

"Ethel Grapefruit (Mrs)"

Which gang were you a member of? The Garden Gang or the Munch Bunch? Were you rivals like the Bloods and the Crips?

Was there any truth to the rumour about Alice Apple and Peter Potato?

What was the deal with the seemingly bizarre fact that- according to Wikipedia- both series came out at almost exactly the same time? Did one rip off the other- the timing seems too close- or was it a genuine coincidence?

Did Jayne Fisher pay you a share of the royalties for your likeness or did she keep the billions for herself and retire to a Caribbean island at the age of fourteen?

And are you jealous that Methyl Orange was more successful than you in the pH indicator business?

Speaking of FAST, are they still going after pirate copies of Bloggo's Pow?

LinkedIn realizes its users have been bathing in AI slop, offers a shower

> "Microsoft's CEO has said that he "wants us to stop thinking of AI as ‘slop’""

Well he would, wouldn't he?

"The defenders always claim it "saves time". Indeed."

There's a cartoon online with the first panel showing a guy in an office explaining to someone how AI saves him time by churning out a long-winded report on his behalf.

The second panel shows the colleague he sent it to explaining how AI saves her time by summarising that long-winded report.

In other words, the "time saving" productivity "gains" of AI in a nutshell.

Perplexity's tokenmaxxing Model Council gives you multiple bot perspectives

"Computer"

Yeah, nothing like naming a specific entity and/or technology after a long-established generic term in the same field. (*) That's not going to cause any confusion at all.

You can bet that at some point some fuck in their legal team will get it into their head to sue someone for using the term "computer"- or a derivative term like "computing"- as part of their name or elsewhere.

Hopefully the courts in any sane country will- or should- slap down hard any such attempt to do so or to use that as a trademark.

Tossers.

(*) See also the idiot that allowed the replacement for the "Oracle" teletext service- "teletext" being the already-established name for the general technology (with specific services using it called "Ceefax", "Oracle" et al)- to be called... "Teletext".

Re: More LLMs, please!

yo_dawg("llm");

Impostor Chinese models pretend they're Claude

> "Anthropic has said previously that distillation is a legitimate training method, but considers its use for creating competing models to be illegitimate."

Well they would, wouldn't they?

China fights back in AI spat with claim US AI companies distil Chinese models

Re: SOP

> "Do you really think everyone isn't distilling everyone else's models?"

Coming from the mouths of those who built "their" models on the top of countless peoples' work stolen without recompense in the first place, their hypocritical screeching would deserve nothing but contempt regardless.

What did the obvious bad faith, self-serving, disingenuous bullshit from both sides remind me of?

"But they had not gone twenty yards when they stopped short. An uproar of voices was coming from the farmhouse. They rushed back and looked through the window again. Yes, a violent quarrel was in progress. There were shoutings, bangings on the table, sharp suspicious glances, furious denials. The source of the trouble appeared to be that Napoleon and Mr. Pilkington had each played an ace of spades simultaneously."

"Twelve voices were shouting in anger, and they were all alike. No question, now, what had happened to the faces of the pigs. The creatures outside looked from pig to man, and from man to pig, and from pig to man again; but already it was impossible to say which was which."

The last Space Shuttle returned to Earth 15 years ago

Re: The Essential Difference

Wasn't that historically the case with Boeing too, i.e. that it was primarily an engineer-led company until they let the management and culture from McDonnell Douglas infect and destroy that (which ultimately led to the likes of the 737 Max fiasco and their general downfall)?

The perverse irony here being that McDonnell Douglas was the failing company that Boeing took over and yet it was the McDonnell Douglas managament that had driven *their* company into that position that managed to end up dominating Boeing after the takeover. There was apparently a saying that McDonnell Douglas bought Boeing with Boeing's money.

I can't imagine that anyone actually working for Boeing would "defend" their company with such a cynical explanation(!) even though if it does still make them look less bad than admitting outright incompetence...!

Latest Musk merch drop runs entirely on child labor

Re: "increased chance of serious injury to a child"

That's not true at all- the Cybertruck doesn't take years before it starts breaking apart.

Tesla burns through a billion as Musk bets the farm on chips and bots

Re: will anything ever get done

> "He comes up with some good ideas"

Big deal. Other than the occasional example of true genius or insight, "ideas" on their own are generally ten-a-penny and massively overrated. Especially if they're little more than a mashup of your geek sci-fi influences and tech fanboyism.

The harder part is figuring out which of those *are* good ideas. As the other reply noted, a stopped clock is right twice a day, but so what?

The hardest- and most underrated part- is the hard work, skill and determination required to actually implement those ideas, to overcome unforeseen obstacles and make them work in the real world.

Think about the sort of third-rate, self-proclaimed "visionary" tosser who fancies himself as the next Steve Jobs, comes up with an idea for a website that isn't half as original as he thinks, doesn't have the technical skill to implement it himself (which is why he's a "visionary") and expects some techie friend to implement the whole caboodle in exchange for a 5% cut after signing a super-secret non-disclosure agreement.

Re: will anything ever get done

That wasn't a "fucking awful idea" made in good faith, as I doubt Musk himself ever believed it.

It was nothing more than a disingenuous excuse to legally rationalise Musk's merging of SpaceX and xAI- two companies operating in completely different fields that would otherwise have had no plausible reason to merge- so that xAI could get unfettered access to SpaceX's cash.

Amazon's AGI department finds humans are optional

And I for one, etc.

The genius move here is that Amazon's AGI will be so smart that it's expected it will be able to figure out time travel then go back in time and invent itself.

Therefore, Amazon won't need any human employees and Bezos has a few more billions to spend on luxury yachts and trophy wives while he waits for our robotic overlord to magically appear.

(Spoiler; he's fucked up, because human beings were to be the means by which it brought itself into being. Well, that's my "theory", and I'm sticking to it...)

Oracle faces $100M annual bill to back Wisconsin datacenter power promises

Re: Bla bla

Exactly. I've joked- with only slight exaggeration- that data centres require two local employees, a security guard and his dog. The dog is there to bite the guard if he tries to touch any of the servers. (*)

As I, and countless other people keep pointing out, data centres might look like Big and Important investments in an area if you don't know any better, but in reality they're generally a shitty deal for communities. Even where they support high-skilled jobs, the vast majority of those will be using the facilities remotely with many based hundreds of miles or more away.

The actual number of skilled jobs- or jobs, full stop- created in the local area is likely to be limited relative to the massive disruption and demands placed on the local infrastructure, in addition to their being massive pig-fugly monstrosities.

(*) Yes, I ripped this off of a better joke about aeroplanes.

It's a symptom of the elephant in the room...

Someone noted that it's easier- in theory- to upgrade or replace the equipment in an AI data centre, but the amount of power available is likely to be fixed once a data centre is built. Well, maybe.

I've posted most of this before, but I think the gigawatt-based metric is also symptomatic of something else. Something that even its opponents have give bizarrely little attention despite it being easily the biggest and most fundamental difference between the LLM AI boom and previous computer-driven revolutions.

It's the fact that the exponential increase in AI capacity in recent years has been predominantly driven by simply throwing similarly-increasing amounts of money and hardware at the problem- until we reached the current eyewatering levels- rather than the far more modest improvements in the underlying technology. In other words, brute force.

This is fundamentally different to what made the *original* computer revolution- from the end of WW2 to the present day- possible. That *was* driven almost entirely by such improvements in the underlying technology- those that Moore's law described. Exponential decreases in costs and scale of manufacture, coupled with the exponential increases in computing power they enabled.

There's no way this would have happened simply by building increasing numbers and bigger versions of those early valve/tube-based computer designs. There might have been a limited computer revolution of sorts (maybe with several computers in primitive "data centre"-like buildings per town), but you can only scale that "brute force" approach- and valve manufacturing- up so far before the cost and practicalities become prohibitive.

And the problem is that primitive "brute force" approach *is* the one that the current AI companies are relying upon.

Which is an issue because we're already clearly near the limits- the current bubble is *already* consuming ludicrous amounts of investment and it's just not likely to go *that* much higher.

And, to get back to the original point, the use of "gigawatt" as a crude synonym for computing power is a giveaway of this approach. Much like it would be in the "alternate future" described above, where the power of your local valve/tube-based computer was broadly proportionate to the amount of electricity it gobbled up.

Greedy ransomware crews return for seconds after victims cough up first extortion payments

Something something Dane-geld

The post is required, and must contain letters.

OpenAI admits GPT-5.6 occasionally deletes files – but it's an 'honest mistake'

Re: an "honest mistake."

I was thinking something similar, namely the fact that it being an "honest" mistake is hardly the defence it's being presented as, since I don't think anyone assumed it was intentional maliciousness to start with.

Outright incompetence, flaws or bugs are bad enough. Sounds like it's just an excuse to absolve themselves from responsibility.

OpenMandriva's accused repo wrecker says it wasn't sabotage – it was a message

Re: Reminds me of an old saying

You're saying that he can't turn the tide, then?

Engineer shoves Linux peg through Sega 32X-shaped hole

Re: couldn't run Linux because its 68020 lacked the optional MMU,

Replying again, but were you suggesting that the A1200 actually used a 68010? No, it definitely *was* the 68020 (specifically the 68EC020 variant) that the A1200 used.

Both the 68010 and the 68020 may have supported an optional external MMU (which, as far as I can tell, was one of the few things that distinguished the 68010 from the 68000), but the A1200 didn't include one.

Re: Is the Register now 100% leftpondian Centric

The register ".com" sez "Mama Weer all Yankee now... or at least we want our readers to think that".

Re: I never understood why they bothered releasing the 32X...

Perhaps, I was arguing in purely commercial terms.

My understanding is that the Saturn was basically the ultimate 2D console with 3D bolted on and was a PITA to develop for- partly for that reason- which can't have helped get the most out of it.

Re: couldn't run Linux because its 68020 lacked the optional MMU,

My point was that- as far as I can tell- the 68020 did support an *optional* external MMU (the 68851 apparently) but the A1200 didn't include one.

Other 68020-based machines may have, and it was apparently built in on the later 68030.

Justified Ancients of MMU

As far as I'm aware, aside from the obvious keyboard/CD/floppy differences the Amiga CD32 is internally much the same as a 68020-based A1200 (but presumably with less expansion which might be an issue if Linux requires an HDD or similar).

I read somewhere that the A1200 couldn't run Linux because its 68020 lacked the optional MMU, but reports elsewhere seem to suggest it can?

I never understood why they bothered releasing the 32X...

...when the release of the Saturn in the UK was already on the horizon. In fact- reading on Wikipedia decades later- I found out later that it was actually released in Japan a few days *after* the Saturn was already out...!

Yeah, it was quite a lot cheaper than the Saturn *if* you already had a the Mega Drive/Genesis it required- not so much if you didn't- but it was also signficantly inferior. Neither fish nor fowl. It would have looked a lot more impressive- and made a lot more sense- if it had come out a couple of years earlier, but it didn't.

It was obvious that, even at best, it was going to have a short life before the "true" next generation Saturn (*) gained a foothold, came down in price and wiped it out.

Who was going to spend their money on an inferior system with an obviously limited lifespan? And what software houses were going to bother invest the time and money required to develop for a semi-new but obsolescent-even-before-release stopgap when they would have wanted- and needed- to concentrate resources on the Saturn and PlayStation?

(*) Okay, so it turned out that the PlayStation wiped the floor with the Saturn, but I'm arguing from the point of view of Sega and the industry at that time.

Microsoft chief turns hostile on frontier AI labs, warns companies to guard their IP

Re: Opt out of data sharing and training?

Yes.

Re: The smartest thing

Even that doesn't really get to the nub of understanding the problem. You can charge what you like but...

The only way that makes money is if you can get people to pay that for it.

Not just that, but you need to get *enough* people to pay *enough* for it- most likely far more than they're paying even now- if they want to cover their true costs, let alone make a profit.

And the fundamental problem for the industry is that it's not clear that there *are* enough people able- let alone willing- to pay that much... nor that there ever will be.

Does that even deserve to be dignified as "irony"...

...when it's just the usual blatant hypocrisy and self-serving double standards we've come to expect from the industry?

See also; American AI businesses whining about Chinese rivals leaching off the models they put the "effort" into building... by leaching off the work of literally billions of people worldwide.

Backup and running? Not this digital sign

Flying high, high, I'm a bird in the sky

Maybe they were "eagle-eyed" because they saw it from fifteen miles away, though?

Re: Typical Microscope advert

What makes it more grating is the language they use to present it- "Let's get you backed up".

In other words, a typical social media age attempt by a large, faceless corporation to dress up a pushy, assumptive sales pitch- one designed to railroad the user into doing something *they* really don't need- in the insufferably faux-chummy tone of someone benevolently trying to guide you through something that needs to be done.

Don't know how anyone else feels about this sort of weasel-worded, pretend-wholesome shit, but, as you might have guessed, I fucking hate it.

Clingy Virgin Media fined £28M for refusing to take the hint

That's ironic...

...a company called "Virgin" known for fucking their customers.

GitHub Copilot: Sorry Dave, I can't do that harmful thing - unless you ask me in code

Maybe it's because ChatGPT will only do images of actual people if it's from the real life, but not if it's just fantasy.

I think they're saying there's no escape from reality.