Posts by cyberdemon 3170 publicly visible posts • joined 26 Jan 2010
> AI uses HBM memory. That CAN NOT be socketed.
Indeed, but why exactly? Is it because of the wide (128-bit) data-bus, or because of impedance changes at connectors causing reflections in high-speed signals? If so, could it work at a lower speed, or with a fancier connector?
Would it be impossible to take some post-bubble surplus HBM chips and solder them onto a pluggable module (not necessarily compatible with DIMM) and invent a new PC form-factor?
They do more than regexes, it's more like a "principal component analysis" of the entire history of human text. It's very good at convincing people that it has intelligence, because it can combine the data in interesting ways not necessarily seen verbatim in the training data. But it cannot "innovate" entirely new concepts, and it loses the reliability and determinism of machines that came before it.
It's a "worst of both worlds" combination of Human ignorance and Machine obstinance. People still assume that "machines are reliable and meticulous", but LLMs are absolutely not. It is only useful for doing a bad job unreliably, but it is fast and cheap. (until the bubble pops)
Not "rationally" though.
They aren't "thinking it through" with any kind of philosophy. They are simply taking the same game-theory texts that you mentioned, along with fictional literature such as sci-fi and thriller books and film scripts, and predicting the next token. They are not intelligent, and are incapable of reasoning.
As such, they should never ever be put in charge of anything, least of all weapons of mass destruction.
Just switch the damn things off
At least the dotcom-era leftover dark-fibre was actually useful for future infrastructure after the market corrected itself, and cost very little to maintain.
Whereas "AI datacentres" despite the Trillions spent, will be no use to man nor beast when the bubble bursts, because it is useless without its insatiable energy demand.
Even a Soviet-era Lada factory has more economic value than one of these datacentres
This is a symptom of a wider malaise.
"Investors" (i.e. idiots and coke-snorting bankers) have become accustomed to the idea that "software" has an almost infinite eventual profit margin. Microsoft, IBM, Oracle et al have historically been printing money off the back of their shitty products, because any new license sale has a near zero "cost of goods sold".
Unfortunately for the AI mongers, energy is expensive, and "AI" burns a "shit-tonne" of it, scaling linearly with each new user and doubling exponentially with each new fatter model trained. So much so that each new license sold costs them more than they are able to charge for it, in perpetuity.
Buying the latest hardware for $Billions can perhaps halve their ongoing energy cost in the short term (if it is even available, given how they have already warped the markets), but they are bleeding from every orifice, so to speak.
Eventually the finance twats will wake up from their amphetamine-induced comas and scurry like rats on the RMS Titanic, but it will be too late.
Of course, Machine Learning, as has been around for decades will carry on, accelerated by edge NPUs etc as you put it (for better or for worse.. The first two applications of edge AI that spring to my doomer mind are face recognition and autonomous killer robots)
But LLMs and generative AI, (i.e. chatbots, coding helpers, image, audio and video generators) cannot feasibly be run on-device because the models are too large. They can be squashed down via quantisation and other techniques, but the output quality is best described as "dogshit" when compared to the mammoth models that can only ever run on a datacentre rack with Terabytes of RAM.
The "datacentres in space" crowd are up against some fundamental physics, not least of which is "How do you get rid of heat when you are in a vacuum". Surely they MUST be aware of that. But they also know that the only thing investors seem to care about these days is short-term gains, so they only need to find a bigger idiot. At this point it's nothing but a Ponzi scheme.
All of that would be fine (their investors fully deserve to lose their shirts), except that they have now resorted to taking out loans, and when the banks go bust it will be *us* paying the price, while the villains of this fiasco will have already cut and run
None of that matters to the likes of Sam Altman, because they've already built their doomsday bunkers "out of fear of WWIII".
Jeff Bezos apparently loved the "fallout" series of games so much that he made it into a TV series and plastered it all over his delivery vans. The piece of storyline that I don't remember being in the original games though, is that the tech billionaires actually started WWIII deliberately, to get rid of the excess human population and re-populate Earth in their own image.
This whole AI bubble is just so objectively grotesque that a Hugo Drax style evil plan is starting to look like a rational explanation for why the rich continue to pour their wealth and our resources into it. Maybe that's why Trump wants Greenland.
Thanks for the link.. Good grief, that is cringeworthy. He must have said "We're following the practices of the previous government. But ask me ANY QUESTION YOU LIKE about our relationship with China!" at least 10 times consecutively..
I do appreciate where you're coming from. It reminds me of a pub (long since closed) near my way called The Honest Lawyer. (in case the pic doesn't load: the pub sign is a barrister in a jail cell). Just opposite The Bent Brief (also sadly closed).
I.e. politics (and law) is and always has been a game of deceit. Which to most of us is depressing, as we were led to believe (especially when Starmer's labour was voted in) that it had cleaned up its act. So any politicians attempting to portray themselves as honest are perhaps making a rod for their own backs: How can you pretend to be honest, if you are playing a game of deceit.
While I agree that Milibean is "experienced", I disagree that he is "capable" (except of vassalising the country to powers far worse than the EU). He has demonstrated a lack of understanding about energy in his post as energy secretary, and a willingness to just accept whatever he is told by lobbyists e.g. those from Drax (or the default action, just copy the Tories)
The most disappointing fact about Starmer's Labour for me is that they are following exactly the policies of the previous Tory government. So much for "change". And by doing so, they are playing directly into the hands of Farage's neo-blackshirts.
> And I still think that "text completion tool" is a bit reductive. Not to mention incomplete - modern models can look up information on demand, delegate computation, persistently store information etc.
Not really, since the information looked-up on-demand is simply added to the context. (this is part of the reason why so-called 'advanced' models use up so much memory and energy..) and they soon run out of "context window" space and "forget" what was asked in the first place. Delegated computation is useless if the instructions given to the sub-program are wrong (what's worse, an idiot with a calculator in his pocket, or an idiot) and 'persistently-stored' info is notoriously unreliable (and subject to context-window limits as mentioned).
> If expert systems haven't taken over the world in the last half century, what evidence is there that they are going to be a short term silver bullet?
There isn't. But unfortunately, "magic silver bullets" don't exist. "There is no such thing as a free lunch."
> They have limited capacity to hold and update knowledge, yet egos often resist being driven by guidance (let alone algorithm), and are subject to all of the normal human biases (confirmation bias etc. etc.).
Er, unfortunately that applies to LLMs even more - they are text-predictors trained on human-human interaction, i.e. chatlogs snarfed from facebook, twatter, etc. An LLM can easily start emulating an "ego", even though it has no such thing. Biases, such as racial or gender bias, are acquired from the training data. Added "safeguards" such as "Please don't be racially or gender biased" in the system prompt, will easily backfire and lead to chaotic behaviour. It is perfectly capable of lying, manipulating, even 'gaslighting', because such perverse behaviours are present somewhere in the training data.
> If a "text completion tool" helps me as a patient live longer, I'll take it.
Actually, the "text completion tools" are far more likely to murder you than to save you, IMO.
It is extremely difficult (I won't say impossible) to make an "AI" that can usefully complement a good Doctor. Yet it is trivially easy to make an "AI" that can "usefully" complement a genocidal maniac. Drones with guns are a thing, and neither 'intelligence' nor 'self-awareness' are required to make a drone shoot at everyone with the wrong facial features. Nor are intelligence or self-awareness necessary to make a system that tries to kill anyone who "wants to shut it down". Remember the LLMs have been trained on a corpus of human-generated text, including all of our sci-fi fantasies about killer AIs. It only needs to get stuck in the wrong context-feedback-loop to go full Skynet if we empower it to, all without any emergent machine-consciousness required.
> Some truth in the above, but the implicit suggestion is that anything important should be the exclusive purview of humans
I didn't say that.
I would bring back logical, deterministic, computers. Not these pseudorandom wet-mass emulators.
> So rather than just trotting out the above trope, tell me: what, fundamentally, makes the mammalian biochemistry a better substrate for emergent “intelligence” in medication decisions that LLMs?
Why does a "medication decision" even need intelligence? How about a good old-fashioned "expert system" composed of if/then/else predicates?
But ultimately a human is required, because you need someone to take personal responsibility for the patient. An LLM has no responsibility. It is not alive, and has no fear of blame or punishment if it gets something wrong. It is, after all, a text completion tool.
But but! We are only a mere step away from completely usurping the entire 'justice system'. Anyone (cough, cough) not yet implicated in the redacted 'Epstein Files' for example, or who is guilty of or plans to be guilty of genocide, will very soon be able to claim that the evidence against them was fabricated, and generate counter-evidence against their accusers/enemies, etc.
Worth every Trillion, i'm sure..
You are missing the point. The LLM doesn't "claim" anything.
"prompts" are NOT instructions, they are statistical seeds. Outputs are not logically computed or deduced. They are pulled out of a hat of randomness with the prompt as a magnet.
What you are mistaking for a "claim" made by the LLM is just a sequence of tokens which fits the statistical regression of the training data in the supplied context. Of course it is going to be wrong sometimes, and it will be more likely to be wrong the more unusual the prompt.
Why anybody lets these things anywhere near anything important, let alone "trusts" them with it is frankly mind-boggling.
> The fix, he argues, involves recalibrating Gemini's RLHF to ensure that sycophancy can never override a safety boundary and that potential mental trauma is given equal weight to self-harm risks in the model's safety mechanisms.
No.
The Fix, involves throwing a giant switch to the 'off' position and a grovelling apology to the entire world for pretending that a LLM was ever 'intelligent' in any way whatsoever.
> Ever seen the wafers they make CPUs out of?
Unless you're talking about the lunatic asylum called Cerebras, silicon wafers are cut into much smaller chips which are far more durable than wafers. And then they are soon attached permanently to a package, and sandwiched with a big copper plate.
Whereas from the Reg's pic, it appears that Microsoft's glass wafers are stacked naked on shelves, to be handled by something similar to a tape-library robot.
So I'd say An_Old_Dog's second point is perfectly valid. How much data could I erase by sending in one madman armed with a clawhammer? Or more subtly, hack the tape-library robot and change the position index of the shelf to be in the middle of the platters instead of the edge
> If you believe that we're somewhere in the "irrational exuberance" phase of AI adoption, then Amazon is building the most expensive monument to hubris since Meta spent $46 billion trying to convince the world it wanted to attend meetings as a legless cartoon avatar.
This.
And on the pedestal these words appear: "My name is Ozymandias, King of Kings: Look on my works, ye Mighty, and despair!"
No thing beside remains. Round the decay Of that colossal wreck, boundless and bare. The lone and level sands stretch far away.
It's Your fault.
We warned you. But You said you'd had enough of "experts", and dismissed it all as " Project Fear"
And I didn't get the *referendum* that was promised. I distinctly remember being told that it would use the same threshold as a General Election, i.e. a "majority" would be required to prevent a second referendum. 48/52 on a 75% turnout is no majority. The people had not "spoken", they couldn't have been more undecided on the matter if they had tried.
Even St Farage himself was pushing for a second referendum when it was looking close, but when he edged ahead it was all "we won you lost get over it ner ner ner"
If the boot were on the other foot, 52/48 in favour of EU membership, there would have been a second referendum and I would have been happy with that.
(Apologies, btw, if your post was sarcastic. You should have used the Joke icon.)
Exactly. The charge should be blackmail. Not "hacking".
If he had instead sold it to a newspaper (if it was interesting enough to be newsworthy apart from the blunder itself) and then deleted it and told the cops that he had done so, then the bloke could've got his money and all we would be hearing about would be sacked police officers.
That doesn't sound like very much, for a dinnerplate-sized chip consuming God-knows how many watts..
Nvidia are a waste of silicon, but this Cerebras lot sound like a desperate attempt to produce a superintelligent artificial god which is doomed to failure.
Kopros ex Machina..
Nice.
That must involve some interesting non-standard wiring though. Here in the UK, the insurance company would probably condemn it.
For most people sadly, solar PV means having a grid-following inverter simply plugged in like a negative-power appliance.
> The amount of money being spent is frightening.
Yes indeed.. Stocks are down, gold is down, oil is down, crypto is down.. So where has all the money gone?? I fully expect a bank crash this year and I have no idea what anyone can really do to ride it out.. It is terrifying to think about.
> I'm going to pick up another 1kW of solar panels next week while they are still super cheap to match what I have.
I might do the same. I have 1kW of panels (realistically 600W, they are laid flat on a flat roof in sunny Britain). But from the grid's point of view, this is pure delinquency. It steals the lunch of the energy suppliers, while at the same time making blackouts much more likely. The inverters require a stable reference signal to operate, any transient disturbance will temporarily trip them off, and the combined effect of that would cause such a rapid change in load that any grid would struggle to cope with. On top of that, you have interesting resonant behaviours with a grid dominated by grid-following inverters causing oscillations in grid voltage and frequency.. Prime suspect in the Iberian blackout last year.
Bloody hell.
The CAT G3500K on order from Joule are 97.5 litre V20 engines running at 1500 RPM. 4 stroke, so that's somewhere in the region of 1 cubic metre per second of exhaust, each. Or more, since they are turbocharged. And they will need nearly 1000 of them for their 2GW. Apparently most hot-air balloons are 2800 m3, so that's somewhere of the order of 1 hot air balloon every 3 seconds of exhaust gases sent up into the atmosphere from this facility. I wouldn't want to be breathing Utah air when this monstrosity powers up.
It wouldn't be so bad if they were making something *useful* with all this.. But it's all for fucking AI which as far as I am concerned has negative value.
Fortunately though, they are nowhere near completion and my money says the AI bubble will go bang long before this bitbarn hits 2GW never mind 12GW.
The sad part is, when that bubble DOES go bang (and it will), everyone's banks and pension funds will be exposed to the fallout. The 2008 financial crash will look like a blip. It could end up being worse than the Great Depression.
And 2GW/reactor is not an average nuke. Most are 1GW or less (you don't want them to be too big or else they destabilise the grid when they trip offline), but of course nuke plants usually have at least two reactors.
I am oddly reminded of a Danish TV series..
https://en.wikipedia.org/wiki/Follow_the_Money_(TV_series)
There was a scene in that where they hoodwinked some Saudi investors into believing they had developed high temperature superconductors.. Or something
https://www.theguardian.com/tv-and-radio/2016/mar/26/follow-the-money-recap-episodes-three-and-four-superconductors-and-snogs
It was a great series, would recommend
Er, yeah.
"US Manufactured" means that they buy the glass/silicon bits from China and put their own, patriotic frame around them. Made with Chinese aluminium.
Even Germany no longer makes their own PV cells: https://www.spiegel.de/international/business/q-cells-bankruptcy-heralds-end-of-german-solar-cell-industry-a-825490.html
The Beeb needs to stay. They are one of the last bastions of proper journalism, and a haven for brilliant satire, like the News Quiz on Radio 4.
But they could save a lot of money and restore some public goodwill if they simply sacked the enforcement goons. Those who want to get TV without paying for it aren't worth the effort trying to force them, they will find a way. But the majority (so I believe - i have no evidence of course) of those who do pay the licence fee, do so willingly, i.e. honestly paying for a service. Most of us wouldn't start shoplifting if the shops stopped having CCTV. A small minority would. But if it's a small minority "stealing" TV, who cares?
Also, I have noticed quite a few coffee shops, bars etc with a small TV in the corner playing music channels. Apparently if you have a TV license and are playing music from a TV, then you don't need the PRS license.
Anything that requires bulk energy is being forced onto electricity.. Gas for heating & cooking is being decomissioned, petroleum for transportation (or heating) is being decomissioned, coal has *been* decomissioned, even wood-burning stoves are being banned.
All telecoms is being forced to go via the Internet, which itself relies on a functioning electricity grid.. Cash is being phased out too, all purchases increasingly require the Internet.
The old telephone exchanges had their own source of power, and independently supplied it to the telephone in your house, so your phone would work during a power outage. Not so anymore.
Frankly, it feels as if the Invisible Hand are preparing us for something Cataclysmic. The "Powers That Be" couldn't be this stupid by accident.
...
where did I put my sandwich board..
The real issue I take with doorbell cameras are the highly sensitive beamforming microphones on each one, capable of picking up every word, every fart, of passers by. A network of these with every house on the same system (ie Ring) would be able to reassemble a conversation between people strolling along a street.
People were very concerned when the Stasi did this in Soviet East Germany, but now, nobody seems to care.
But surely even the Cloud pushers have been hoisted by their own petard here..
What good is a pile of energy-guzzling coprocessors if there are no server CPUs, no RAM, no NAND, and no motherboards to assemble any actual servers...
It's as if Ancient Rome adopted a Perestroika policy to build millions of chariot wheels.. But then there were no carpenters left to build Chariots (never mind horses, or charioteers)
Yes! Let the bit barns run on solar, wind and batteries exclusively, BY ORDER. No gas and no grid connections. Nobody will shed a tear when the inverters splutter their last at 75% through a training checkpoint.
Burning finite global resources on this madness is, well, madness squared.
