Posts by mihares 123 publicly visible posts • joined 7 May 2019
Going to die. Edge models can already execute on hardware very motivated consumer can buy and are already “not bad” if you are in that sort of things.
With another couple of generations nobody is going to bother paying the must-be-ever-increasing bill for cloud AI.
I predict that by 2031 most of the scum bAIgs will have kicked the bucket hard.
I've said it once and I say it again: you don't install your policing on my personal device.
It amounts to a constant, warrantless police search of my personal stuff.
To search me, una tantum, you need to convince a fsck'ing judge that it is necessary to a very specific end --you cannot weasel your way around it with a "think of the children" kind of law, not in any place of the world which aspires to be just and free. This is the kind of thing that China does.
Awww that reminds me of that time I tested a "data maintenance" script, freshly written, directly on the company NAS and not in the test directory I set up a couple of hours before.
Luckily that test directory could be used as a backup before anyone noticed that I nuked a whole research project...
> Another ally questions reliance on American AI
I question the reliance part. If at this point you are _reliant_ on (the latest hype spur of) AI, you're fsck'd three times over.
Note, since the consideration bubbled up also here:
a) If you can make your own software you are protected in the short and mid term: hardware doesn't just go bad or disappears by remote wish as software does.
b) ARM is not under the influence of yankees (and we have plenty of people capable of designing on silicon in EU... of course: painful and not immediate, but not impossible).
So they are looking to suck some money from the biotech and pharmaceutical big dicks, aren't they? Because the PR spin here is "hey! Look at me! I might sell you something to vibe gene editing here!"
I'm afraid that's the wrong PR for that. The right one is way more boring and is already being implemented in-house: it's how the hell do we bring together trials from all around the worlds done with poorly defined stats? Too bad LLMs suck _very hard_ at anything mathsy...
And not even original or new. Just the other day I was recalling a post from some random scumbAIg a few months ago which claimed "in a few weeks, you'll all be jobless"... now, apparently, people are rehiring those fired to make place to agentic AI (at a lower pay, I presume).
The deadline for the expected end of the world as we know it has passed so here we are: more doom posting along the lines "our product is so powerful and expected to become even more so that the world might end". Right. As for now, the single biggest threat that this whole AI thing is posing to the world is when the humongous financial bubble will burst and a similarly sized economic crisis will ensue.
If we want to pause something, let's do that with the money wasted on a product of many great promises, which has yet to prove it can profitably deliver on any of them.
Well, I only have 25-ish and one thing I learned is to own my fuckups (which implies the very strict policy of generating them by myself).
Furthermore, if your vibe coded but "manually reviewed" test suite is linked to regression and unexpected behaviour in enough use cases that you get asked to please don't, there might be two conclusions to be drawn:
1) despite your experience, your manual review stinks,
2) you machine generated "by the next most probable token" test suite stinks.
Suggestion: if you're getting flooded with machine generated bug reports, use the LLM there: have it cluster them, review the clusters and summarize them in whatever number can be handled _well_, do not have your requirements badly machine-translated into Python. That is a recipe for disaster --which, apparently, did happen.
Of course juniors‘ output is of less quality than experienced folks: nobody is born with knowledge of whatever process a random company is implementing.
The answer is formation and it is not achieved by hectoplasmic osmosis while sitting in the same office of more experienced people, it‘s made by investing time of the junior and a few seniors who show them how things are done —wether at the desk or by sharing the screen in a call (a technology that exists and is proven, don’t you know?).
Problem is: it’s an investment, you pay upfront to get more later. Many managers just don’t get that (and don’t get the hectoplasmic osmosis thing either) so they don’t like to hire juniors as a rule. Which is a small problem in a healthy economy, and a large one in an economy where you make your money on the number of extra minutes you steal from your workforce.
But I get why all those American fscktards have it in with the single biggest increase in worker’s well being and rights since the seventies.
People who make laws should be required to understand a bare minimum of what they regulate.
Even before considering any ethic implication or the moral ground on which it stands, this ain't gonna fly technically --except perhaps in the most bovine implementation of: giant library of banned files against which the current print should be compared. Which is stupid for copyrighted images and even more catastrophically idiotic when you consider STLs.
But the really poisonous aspect of it all is the will of the government to install policing onto private devices. If you want to do policing, hire agents and send them patrolling the streets and, if you want to take a look into my private stuff, get a fscking warrant and knock at the door. You don't install your snoops warrantless on my stuff. That's what China does, and China ain't no democracy.
Yet another thing that is going to make these AI sold by the token more profitable --even more than the money black hole that they are currently.
That said: interesting.
The mildly interesting part is that you can prune machine generated (text) input very aggressively. OK, let's say that this part does not surprise anyone and it's probably not even interesting. You can probably do that with the output as well.
The definitely interesting part is the concept of input compression for LLMs. I wonder what you could do if you'd implement compression/decompression directly in token space (of course locally: again, if they sell you AI by the token of course they aren't going to do it) --might be also an interesting compression algo even omitting the LLM completely.
BTW: there's at least one sure, obvious way to keep your token bill firmly at 0 unit of currency. Just try it, it's great.
I wouldn't pass such black and white judgements about the US winning races without specifying the context. If you are talking about rubbish consumer equipment, yeah, maybe.
But that aside, two considerations:
1) LLMs and diffusion models (image generation), alongside with all the previous iterations, are _far_ from useless. They are just not useful for _everything_. The way they are force fed to the people is: corporate is never going to be happy paying for their data to be used by these service providers and random users are going to use them until they are free of charge: the future of these tools is edge inference, full stop.
What about training? Well, we'll use aging ones until we're going to be able to train at home --fine tuning is already within reach.
2) The dumpster fire that is that very specific part of the internet which creates problem has become so _precisely_ because the regulator moved way too late: the GDPR came a good decade too late and the rest of the world does not have anything of the sort yet.
I do not want to even watch this race, if pirates the like of Altman and Musk are able to run it without rules.
> Earlier this week, executives from companies [...] publicly warned that Europe risked over-regulating itself out of the global AI race.
It would help to decide, first, whether it's a race worth running --and running the risk of having no or insufficient rules earlier on.
So far, the value produced by the whole thing is vanishingly insignificant compared to the amount of money moved around (very literally around, in a circle) and said value is surpassed by order of magnitude by the collateral damage done --if you don't want to think ethics, just think about RAM.
We've seen not so long ago layoffs at Microzozz and Meta justified in the same way: not even AI, but _investment_ in AI.
I don't think people intelligent enough to tie their own shoes could possibly trade revenue-generating workforce not for a finished, proven product which replaces them fully for a lower cost (or at least with a provably acceptable degradation of performance in the context of the savings), but for the hope that an unfinished, unproven toy which is now just a money sink will one day be able to replace the lost workforce.
So I am going to say: this is just AI-washing layoffs which the management wanted to do anyway, and now they have a shiny, stock market friendly excuse to do them.
That, or those managers' parents should really be blamed for the reckless over usage of child safe wall sockets.
I looked up quickly the ThinkPad T410 and it comes from an era where a) Nvidia cards were a common "plus" option and b) they just started that idiotic multiplexing with the inescapable Intel Iris GPU on your CPU die. Your experience tells me that you are probably plagued with one of those separate Nvidia cards.
I feel your pain. If you don't need the Nvidia card and can disable it for good in the BIOS or UEFI: do that. Getting the system to work is even today unnecessarily complicated --by Nvidia, not by any feature of Linux other than being GPL'd. Huge respect for the folks who develop Nouveau and their relentless war with green windmills, but _for that era_ it is just not viable.
Modern Nvidia GPUs play a lot nicer with Linux --out of need: if they can't get them working, they'd just do AI on AMD cards and Nvidia wants that money. Still, if you do other than compute, your life is going to be much better _without_ Nvidia.
~~~
That, and the usage paradigm that is fundamentally different: on GNU+Linux, you don't come forwards without talking to a shell. Just as on Windows you don't come forwards without pumping the mouse. The change isn't comfortable in either directions --personally, I prefer talking to a shell --> I am 100% *nix based outside of work, where I have to contend with the mouse pumping thing. And loading times which remind me of IDE disks of the '90s. And random crashes. And the lot (but the PowerShell does get a lot of love, fact).
We are talking about having an account on social media, I think we are well past the lack of care for privacy, freedom and possibly civilisation at large.
Since I do care and have done so since a while, I do not have any account on social media —and will teach my children why they should not have one either.
Identity verification (of which age verification is a rather small subset) is not hard.
Here in Germany you can do pretty much everything from ordering a SIM to open a finance online, you just have to go through a process called WebID which, brutally simplified, amounts to you sitting in front of a webcam waving your ID card at a remote person. If they like what they see, you get a (cryptographically) signed copy of your contract.
In the case of a SIM and a finance some sensitive personal info are bounced around and saved, but that’s rather a mute point because of the nature of what you’re doing. Arguably even more so in the case of the facial book, but in principle you can have the system giving you a signed token and forgetting instantly about your face and your ID.
Of course, it doesn’t scale well. Might need months processing all accounts.
Watch me giving a fuck.
> The most damning finding [...] is that "healthcare workers are pasting real patient data into AI chatbots, and that data is now a commercial database."
Well, *two* databases, since they got ingested by one or more of those scumbag companies _offering_ those AI chatbots. And that should also explain why, in some contexts, usage of known data-slupring services should be radioactively illegal.
I mean the user can perhaps protect themselves by (doing the sane thing and) not using those services with sensitive information, or even better: at all, but they cannot protect themselves by the lunacy of others, who may be authorized and effectively need/must know sensitive info.
Alas, that people have not even the residue of the concept that it may be not OK to give sensitive info online to whatever company is a sad, constructed and very well maintained reality.
> "This experience helps users quickly understand content, take action with fewer steps, and get more value from Copilot while extending productive browsing time in Edge."
This data scraping operation helps us quickly hoard user information, forcibly inflate Copilot’s user base, and pull the investrors‘ leg a little bit longer while figuring out how to swallow the huge losses we are making on this ridiculous AI system.
Use decent software and don’t be a part of this.
Again, not at all surprising --or, rather: it is surprising to see people "discovering" this, when there's a technique specifically though to make an LLM say what you want and not necessarily what the user asks.
I shall also most definitely *not* be surprised to find this kind of crap between user queries and Google-server LLM stuff to serve you ADs in the answer.
And it is very hard to solve: there is *nothing* that stops anybody to engineer the prompt you are giving to the LLM, and, in general, LLM input is very hard to sanitize --you're not dealing with a simple, well defined syntax like an SQL dialect, for example, you're dealing with unfiltered human language.
We are not good at sanitizing our own input (I bring social engineering as an example): LLMs are way worse. Don't trust them unless you don't need them if not for a speedup.
> That list, VMware claims, included plenty of software that Siemens had not licensed.
This is something that just does *not* happen, not in corporate.
And since when you see VMware you read Broadcom (who made a hellish mess with VMware product's licensing) I'm inclined to think that they are mostly trying to spill money out of Siemens.
In our case, what I think happened (and I am way too lazy to check whether this is actually true or not, it's just how I remember it a few years later) is that -O3 gives the permission to the compiler, among other things, to reorder the instructions as long as the result looks the same.
For the vast majority of *good* code, this is fine. There are cases where this is not fine even for good code, and for good reasons, hence the warnings about using -O3 in various documentations.
HEP (High Energy Physics) simulation code is most definitely *not* good code and I was most definitely not paid enough to fix it all --> I did not use -O3 to compile it.
That is not surprising to anyone who actually fine tuned any kind of neural network, I suppose.
What is interesting is how "little" it takes and how far it ripples through the model's output. Yet another indication that trusting the output of these things is very dangerous if you are not already well versed in the topic and can check the details easily.
“but this one comes gratis from a big, well-known name and it has existing management tools.”
That big, well known name enjoys part of its well-known-ness to massive data slurping.
Which, that well known name, says is the reason for many of its gratis offerings.
So… what about that?
Governments, even if it has been very recently proven that is not necessary, are pushing for mass scanning user devices. Because they see that it’s easily done —think of how many mischiefs they’d catch if they were allowed to do so!
They don’t give a damn about privacy if busting it is in theirs and not in Facebook’s or Apple’s interest.
This mechanism will be rolled out in the EU to implement the famed copyright filter, which has been already mandated by the Union and it’s waiting for uptake in the various states.
Apple was sick and tired of denying (or obliging) to requests along the lines “give us access or we’ll tell everyone you defend child molesters”. And this is the result.
The “good thing” is that the machinery is based on a feature extractor that happens to be on the device owned by its adversary, so it’s just a matter of time I hope before some hack finds a way to break it or, probably, DoS it with a class of adversarial examples.
Nonetheless, this was an answer to a request that is done all over the western world by the governments and that is: easy, gratuitous snooping on everyone’s device.
Contrary to China, most of us can and do elect their representatives: watch what you vote the next time because this shit was asked to Apple by them.
~~~
It comes down to trusting Apple: yes, of course. But it was already that way since the iSoftware is very much proprietary and very closed. This is a different but related problem: there are less mobile OSes than car brands in the DDR so the choice now is, effectively, having your pictures scanned by Apple or being key logged by Google.
Or send them all to the deuce and have a dumb mobile phone, a tablet PC running GNU+Linux and a desktop workstation doing the same.
I was kind of hoping for this kind of comment, because now I can unload further. Thanks!
In software engineering (engineering and life at large, if you will), there’s such a thing as the correct tool for the job.
There are things C sucks at, even before security concerns. Arguably, there are many of them. From my old field of expertise: data understanding and data analysis prototyping is one of them. If you are trying to make sense of a signal to noise ratio of 1e-7, the last thing you want is care about what exactly you’re doing with your memory. Failing to recognise this leads to monstrosities like CERN’s Root —interpreted, JIT-compiled C++ dialect… and you thought C was bad?
When it comes to system programming (from embedded to kernels), C is the way to go. Because it lets you free, unfettered range in exactly the things that matter (not even C++, C).
This also means that you are totally free to shoot yourself in the face with it if you’re not careful. So you need to be careful —or at least think about what you are writing.
People trained to connect their brains, have a think and then write are expensive, though, and there is a damn good reason why large corporations (one that does not do OSes, one that shamelessly leeches them and the other that has an effective monopoly on ElReg’s BORK!BORK!BORK!) are attracted to things that require less thinking —> less expensive people to deliver the same amount of risk.
So my point stands: if you do the Kernel, you need to be able to C well. If you can’t handle the latter, you’re unfit for the former. It’s as simple as that.
Waiting for Torvalds’ death ray to discharge in Cook’s direction (still hoping for an uncensored, un-PC version) I will say that:
0) C is not difficult, it’s the opposite of that —that is what makes it a very good choice, possibly the best choice, for low level system programming. If you come around annoying the fsck out of people claiming that C is too difficult to do things securely (or: well) the solution is that neither you nor who trained you should touch it. Or the Kernel. Let alone opinate about them.
1) There are generally two ways to think about security: the first is the wrong one, in which you do stuff not thinking about it, then you pass it on to someone who just thinks about it and ruins your work’s performance and usability because hey, a safe box with the door welded shut is the best!
The second is to think about what you’re doing in the first place and design it so that it is decently secure without someone welding it shut for you —AFAIK, this is the Kernel approach. Which is constantly broken by pillocks at Google, btw.
2) Should Google contribute 100 engineers to the Kernel, I want 200 people checking their work for damn backdoors, since Google’s anus’ track record for improper, surreptitious access is really not clean.
The very existence of this difficulty should ring a few bells, shouldn’t it?
Anyway, I think a sensible way to go about the business is to define it internally as a set of behaviours and then ask people if they experienced those, without leaking the definition or what you’re after —to be on the safe side, have the questionnaire written in 5 different ways so you can catch yourself suggesting answers.
Something you don’t do is stating that bullying may be a problem, you’re investigating it and then ask if the respondent has experienced it. Because that’s just harvesting meaningless answers that confirm whatever you wanted to “prove”.
Of course we don’t know whether any of that happened, because the survey hasn’t been published yet. We just know the results. And that it was self reporting —which frankly stinks of the paragraph above. And stinks because of self reporting. Self reporting always stinks.
