Grid Power
On the plus side, at least they're drawing power from the grid, and not installing a load of "temporary" behind-the-meter gas turbines in populated areas and claiming nobody is being harmed because they're "temporary".
469 publicly visible posts • joined 14 May 2007
So we know that just training on bigger and bigger datasets isn't going to make things much better: if that were the case then we wouldn't have invented RAG and Post Training weights and all the other layers of post-processing after each turn has been through the regression mill. But we have, because even if you plagiarise the whole of human endeavour as training data, a layered neural network is never going to be "intelligent" (first rule of AI: what you get out has to be less than you put in - essentially a version of the second law of thermodynamics).
Which kinda stuffs up the "it has to be bigger" sales line. The domain specific use cases that use a fraction of the weights and a bespoke RAG database aren't going to be improved by bigger. And we're talking orders of magnitude smaller weight sets: if you can run your internal AI Assistant on a couple of half million pound boxes using a distillation from a set of open weights then ... follow the money.
And that's where the business model based on "we're going to get so many more customers than we have now, and each customer's going to pay us much more money" seems a little overoptimistic. Small may be beautiful, but more importantly Small is Cheap.
How are we assuming that it's a good thing to be churning out ever more PCs? We've long been past the point where everyone who needs/wants a PC has one, so shipments *should* be going down.
We don't have to worship at the altar of Line Go Up. Of course, the glut when we find there are no datacentres for all these machine being built ... will be its own problem.
Most software vendors recognise the simple rule: if you have a disruptive migration to keep customers using your product, they will use the disruption to go elsewhere.
Hell, I remember being stopped from migrating a customer's Solaris box when they said "If we notify customers of the migration then they'll realise they're still paying for it and cancel."
Absolutely. Boeing is the poster child for companies ruined by MBAs squeezing everything to generate cash to buy their own shares.
The truth is Western industry has been hollowed out from the inside, not that "they stole it" as the orange gibbon claims. Willfully saddling a company with debt for non investment purposes is egregiously wicked.
So they've been roped in by OpenAI to a datacentre deal that will never happen: find someone who wants something for nothing and give them nothing for something.
But they're fiddling round the edges, a round of sackings because everyone else is doing that just now. Justify them "because AI" despite the shitshow that's OpenAI whose existence depends on much more AI that nobody wants. The circularity of it makes my head hurt.
If only there were a word to describe how I'll feel when it all falls apart: a sort of happiness at the misfortune, maybe "pity-happiness" if we were able to just ram words together to make new ones.
They've for a while been an "Installed base only" company: anyone who was going to buy in as a customer has already done so, newer organisations wouldn't choose to buy in (see VMware, Oracle et al). So they will only grow by increasing the amount that locked in customers pay them (I mean you, Larry) forever being the toolkit that someone has foisted on them.
Wind forward a few years and they'll be one of those companies that, when mentioned, occasions the reply "Oh, are they still going". Like BMC today.
IBM, like Oracle, has for a long time used the sales tactic "We've got them by the balls, so we can squeeze as hard as we like".
So while the Claude Code story is the usual self promoting bollocks, the market reaction might just be "your days of abusing your customers like this may be numbered".
And that, boys and girls, is why we don't use static credentials. Perfectly good mechanisms exist for dynamically creating credentials when they're needed and revoking them after use. We've now reached the point where using static credentials for privileged operations could be considered professionally negligent.
I use a couple of keyboard remappings (put the Ctrl key to the left of A where God meant it to be; swap Alt and Windows because my keyboard's broken). The upgrade lost these remappings so I had to reapply them.
Thanks, as ever, to the Mint team. Having retired last April, I've been able to go Microslop free, which is delightful.
Indeed, I had a similar conversation with Deepseek about whether Edge of Darkness was broadcast before or after Chernobyl. I quote the response verbatim
----
The timeline is clear: The original Edge of Darkness was made and broadcast after the Chernobyl disaster.
Here are the key dates:
Chernobyl Disaster: April 26, 1986
Original Edge of Darkness (BBC TV series): Broadcast in the UK from November 4 to December 9, 1985.
----
We then had a long discussion about the relative merits of RAG queries, where this sort of thing happens far less often because there's much more processing and less stochastic parroting.
Welcome to greedy bastard land.
In the red corner you've got world's biggest manipulative retailer. Everything is based on screwing suppliers, manipulating customers and abusing staff.
In the blue corner you've got the AI grifters who are in the early stages of inserting themselves in the purchasing chain. Assume this is so that later they can sell promoted positions on their godawful LLM damaged browsers.
And they'll end up bickering about who gets the upper hand in abusing customers for their own enrichment. A pox on them all.
^^^^ This.
The same as bundling Crappilot with O365, they have to come up with a ruse which will put some sales in the "AI" bucket so they can puff it to the markets.
Unfortunately, people not wanting this shite is immaterial, it simply has to be done to justify the tens of billions that's gone down the AI drain. The next step will be to enshittify non-Crappilot W11 by taking out some feature everyone uses.
Remember how earlier in the year Microsoft jacked up the prices for O365 and bundled Copilot? And we learned that, in order to not pay for Copilot, you have to follow the workflow to cancel O365, at which point you get offered the old price without Copilot?
So how much of this AI revenue is actually coming from tricking your customers into paying for something they don't want or need?
A well run company is concerned with the long term, and with risk management.
A well run company will consider its supply chains long into the future. A well run company will consider the risk of external shocks and manage itself to provide resilience. A well run company develops staff that will be its future skilled workforce. If a well run company has a bumper year, it will either invest or retain its windfall allowing either growth or resilience. The share price of a well run company should have a boring increase as it grows organically.
If you take a well run company (for example, Boeing "classic") you can squeeze money out of it for many years before things start going wrong. You can lengthen your supply chains, remove margins of error and use increased operating profit to buy back your own shares. But in the process you are hollowing the company out, year after year, and making it weaker. Replicate this across the market and you have today's late-stage capitalism: execs are incentivised on this quarter's profit, or share price, so that's how they manage the company.
And that, my friends, is the story of the 21st century in "the west". If you want long term corporate thinking, you have to look east.
IMHO it's because any investment decision in China, whether financial or political, is made with long term considerations to the fore. The very furthest political thinking in the west is 4-5 years, so if a minister proposed (for example) building out STEM university education they'd have to give a "how will this get us re-elected" justification.
https://www.financecharts.com/stocks/BA/cash-flow/repurchase-of-capital-stock
Turn the time window to 20 years and see what it shows: the squeezing of costs and cutting of corners hasn't been to keep the company afloat, or finance new products. Instead it results in the generation of cash, and with nothing "worthwhile" to invest in within the company, it gets used to make shareholders happy.
Lots of tech companies do the same, but Boeing is the poster child for chronically weakening a company for repeated short term financial gain through share buy backs. And it's a very addictive drug, once you've started the market will punish you if you stop or postpone buy backs.
Alternative narrative: Alphabet and Meta have had a duopoly on teenage video sharing with YouTube and Instagram. They have been slurping personal data then charging advertisers for "targeted audience" for years.
Along comes TikTok and provides competition for the first time. "That money should be ours" they cry "It's not fair that they're doing a better job than we are. Do something Uncle Sam or our line might stop going up".
"Time to deploy our lobbyist network" they think. And they only need to use two words: "But China".
They collect creepy amounts of tracking data in order to profile American teenagers? True, "But China". They develop algorithms to maximise scrolling time rather than social engagement? True, "But China".
Their lobbying campaign achieves two objectives: it removes a competitor and distracts attention from their own creepy and controlling practices.
Lovely inversion of the problem. "We'd be building facilities that would suck in gigawatts of power if only you'd supply it to us".
Our grid works really well for "normal" loads, including running an increasing number of cars (well over a million now). What it's not so great for is supporting the AI grift where datacentres come at 400MVA a time (https://www.theregister.com/2024/09/13/hyperscale_customer_to_take_massive/) - say the equivalent of a Coventry or Leicester wanting to plug in just-like-that.
Alternatively, Chinese industry moves the way Western industry used to: they spend many years improving products that people want to buy making them at sensible cost. And not having MBA w**kers spending all day d**king around with power point, sacrificing the future of their companies for this quarter's figures/bonus by deskilling, slash an burn. Then whinging about "unfair competition" when someone else takes a longer term view than they have.
Yeah, in the case of non-software patents (things that can be made), then inventing it yourself doesn't mean someone else didn't invent it first then patent it. Even if you invent it first, the only defence against someone else patenting it later is to either:
1. Patent it yourself, or
2. Publish a paper describing your invention
In most of the world the system works quite well: patents get examined properly and are granted for 20 years when something is inventive. The problems in the US are:
1. They allow software patents
2. They tend to grant dubious patents and leave it to later litigation to determine the patentability
3. If you lose in the later litigation you're subject to the infamous triple damages
4. There are patent troll friendly areas, and it's very difficult to get verdicts overturned by higher courts (hence when NetApp, CA based, sued Sun Microsystems, CA based, the suit was filed in East Texas)
If you look at nVidia's forecasts, and the new GPU production facilities being built by TSMC it all points in one direction: our energy efficiency savings being cancelled out by endless GPU barns so that the Powerpoint jockeys can bore on about how Aye Eye is, like, shifting the paradigm of everything.
We should follow Ireland's example and start turning applications for these latter-day gas guzzlers. IMHO.
Autocomplete on steroids is a pretty accurate summary.
Parent commentard's observations are also mine - LLMs confidently produce absolute shite when asked about something specific, will sometimes do a reverse ferret when challenged or, as Lee says, double down producing a pyramid of piffle.
When asked to produce a piece of code, what comes back is typically broken: the main benefit over an incompetent human developer is that it screws it up faster and more cheaply and doesn't argue about what you said in the first place. Try asking ChatGPT for something in OpenSCAD - it won't say it's beyond its capabilities it will just produce poorly parameterised code that (at best) is a broken object. As described, it doesn't care about solving the problem just about producing a plausible looking reply.
And this isn't a harmless bit of nonsense, it is:
a) draining investment out of productive economic activity
b) crowding substantive issues out of political and management discourse, and worst of all
c) an energy guzzling catastrophe, reversing years of savings from energy efficient computing
Always my beef with the new-improved...
I'm not concerned with how it works when new, I'm concerned with all of the failure modes over time. BUT, if the concrete coming out is chemically the same, or similar, to the stuff we get from big carbon producing kilns, I'm all for it.
And yet, I wonder if humanity has leapt into overconsumption of concrete, as it has with plastics, and the better solution is to stop using the stuff.
GEC was quite the outlier with its own cocktail of disfunction.
Credit control was hugely damaging - the rule was that suppliers weren't paid until they put you on stop and you needed something. So projects would grind to a halt because components hadn't turned up. The opening position with any new supplier was "we'll require six months credit".
Acquisitions would be bled dry and then shut down. I remember an early 1980s robotics company that Weinstock bought, moved from Watford to Rugby, then starved of development cash until its products were old-hat, at which point it was shut down.
The end result was the infamous cash mountain which meant that GEC became a takeover target. The good thing was that after one junk-bond financed takeover scare, Weinstock panicked and merged swathes of the company with Alstom, at which point some of the financial fuckwittery lifted.
This was all before the Harvard MBA/McKinsey types got into their stride and started hollowing out business after business. And so now we have long-view Chinese companies that make long term business decision and, when they pay off, the MBAs try and put up trade barriers rather than relearning how to run engineering businesses.
Not a new sentiment - Harry Truman used to refer to Chaing Kai Shek as "Cash my cheque". But that was back when Taiwan had China's seat in the UN Security Council and it wasn't clear Mao Ze Dong's regime would last.
Now, if all the Harvard MBAs and McKinsey consultant's hadn't recommended the offshoring of chip manufacture, the world would be different. But then they wouldn't have been able to charge those juicy fees. It really grinds my gears when bean counters sacrifice supply chain security for cheapness then whinge that their supply chains aren't secure.
I remember in the 1990s going down to a Novell office to see their take on a SGML based word processor. Their internal markup approach lent itself very well to morphing into generalised markup (I'd always used WP with reveal codes on up to that point) and when pointed at the right DTD and stylesheet it gave you WYSIWIG SGML.
It never saw the light of day, or if it did then it wasn't properly promoted. The world wasn't interested in proper document management when they could just produce great big hairballs of MS-Word.
Quite simply, anything developed for the public sector should be public code. Councils do about the same thing as each other. Hospitals have the same IT requirements.
Basic public service software: general ledger for a council; appointments booking system for hospitals --- should be developed by central government and made freely available, and UK public sector bodies should have *&^% good reasons for not using it. Councils are not that complicated, the legislation is the same for everyone, and to fuck up a basic ledger system is so inadequate.
https://publiccode.eu/en/
Once again, the problem is stated as being data going to a Chinese company, when the real problem is the data going to *any* company.
I'd love to see this article rewritten without the "China" angle, stating the real concern: user data is being scraped from UK government websites and shared with commercial third parties.
I've written to my MP about this several times. At its core, the Horizon scandal wasn't about mistakes, it was about wrongdoing. It was about Fujitsu management being told by their development teams "Yes, this could happen" and then swearing on oath that it couldn't.
Miscarriages of justice work both ways: getting away with flagrant wrongdoing, whether at an individual or corporate level, is as obnoxious as banging up an innocent person.