Dave.

All current datacentre hardware is being run at power and density levels that hover somewhere between “operational” and “pile of glowing silicon slag”.

I’m sure some enterprising Chinese companies will take all the surplus hardware and create underclocked “consumer-capable” versions for inference. I’d happily pick up a couple of cards that are 75% of cutting edge performance if they’re a hundred bucks a pop.

I posted this the first time around:

Looking after your fridge is not a problem when you just have one fridge. You open it one day, oh shit it’s not working, you get it fixed. Done.

Companies have hundreds of stores each with 10 or more fridges in them. Nobody in the store is keeping tabs on whether the temperature is set correctly to meet food standards, or whether the condenser is blocking up and full of fluff, or a bunch of other maintenance and administrative stuff associated with owning a bunch of fridges.

Solution? Remote reporting and management. You can keep an eye on all 5000 of them 24/7, alert the store if there’s an alarm, call maintenance when there’s a failure , etc etc etc.

The only mistakes made here is that they didn’t put their fridges on a secured internal network, and they didn’t update the software when Danfoss said they should .

I work with embedded devices that use canbus. Last week it was happy to read the manufacturer’s published .DCF files for a couple of joystick panel interface boards. Not anymore, conversation flagged and halted any time I get near them.

Claude. You can download these files from the manufacturer’s website. They only describe how analog and digital inputs get converted into canbus packets. I am not trying to hack Adobe here. Nope.

Oh well, DeepSeek, Mistral, GLM, Beam et. al. are all happy to read these files and continue work on the project. Anthropic’s usefulness to me has ended, it was nice while it lasted.

But yeah for early game, it seems like a simple solution, and the answer is proper delegation instead of a free-for-all.

That’s the problem. Humans would each gravitate to what they do best and organise themselves into a structure relatively quickly.

LLMs get stuck in circular patterns because they don’t have the ability to do that. They can’t think outside the box. None of them arbitrarily becomes the boss to take charge. They’re all yes men stuck in endless board meetings where even just flipping a coin would break the deadlock. But they are incapable of realising that, because they are just statistical regurgitation machines flavoured with a bit of randomness on the most probable outputs.

I feel there has to be a “tinkering factor” in there, so that if that non-RF link is something constructed from spare parts and curious fingers it still counts to some extent depending on the home-brew component of the construction.

Also distances over the internet are immaterial. Perhaps shifting to a module/component count would be better. For example, 2 radios on each end 100%. A radio, a self-built software modem, ((the internet)) , another self-built modem, a radio is still 80% HAM.

haven’t gone through the trouble of registering copyrights with the government or with copyright policing services,"

This is country specific, of course. For example, in Australia the act of committing a work to media of any form is enough to register and assert copyright for that work.

So as an Australian resident I am quite within my rights to assert the following immediately, with the legal backing of the Commonwealth of Australia:

This post is copyright ©️ dgriffith. I grant the Fediverse and its users/instances permission to duplicate as necessary as part of the normal mechanisms of participation, all other rights are reserved.

Especially not with Finns. They like their forest, about 3/4 of the population goes and disappears into the forest during summer holidays.

And then, they are also quite strict on The Rules Being Followed To Maintain Polite Society, so a corporation both destroying forest and breaking The Rules, well, that’s not gonna end well.

I would encourage you to try get others in on the reviewing effort.

Mostly so future me doesn’t undo something I already argued with myself about. They’re in docs/adr if you’re curious.

Mostly so future LLMs don’t undo something they already argued with themselves about.

One way, I set the numbers local, can’t be changed otherwise.

Its simple (as am I), why does it need to be on the “cloud”???

That’s fine when you have one fridge.

Then you need to look after four thousand of them.

Oh, it’ll be fine, we’ll just set things once while installing them.

Oh look! New food, or efficiency, or store standards have dropped, now you need to update those settings in those 4 thousand fridges spread across the whole country. Well, we can do a couple of stores a day, I guess it’ll only take a month of a service tech’s time and travel costs to do it.

Orrrrrrr… we could hook these fridges up to the Internet and do it remotely, and update the settings on those 4000 fridges before lunch.

That’s the draw, convenient mass update. The main issue here is that they put the fridges on an unsecured network, and then on device auth was laughable, allowing any bored teenager with time on their hands to tinker with things.

Just because we stumbled through a whole bunch of world-ending things in the past mostly unscathed doesn’t mean we’ll continue to do so.

It seems that after the initial AI-generated patches there’s been good discussion amongst the kernel devs about sorting out the issues, and a few revisions already that have tidied things up a bit.

So if after all that there ends up being a good performance improvement directly from a revised AI-created patch, what do I do with my pitchfork? Still raised angrily? Half-mast? Do I just let the torch smoulder?

Nice! Although I do wish everyone would get with IPV6 at some stage. But it seems everyone is OK with the current coping mechanisms in IPV4, so maybe I should pin my hopes on IPV7 or 8.

So why should they apply air gapping to their agents?

Because if they don’t, they could face jail time personally or their company is fined to death? That’s how you apply corrective pressure to those thinking about “going rogue”.

~$ groups crank0271 docker

~$ docker run -rm…

No joke, this is exactly what LLMs will do if they can’t do something directly.

Most ISPs charge insane fees for static IPs

The joy of living in the US I guess. My ISP in Australia charges $5 a month for a static IPV4 address.

There may have to be fundamental engineering breakthroughs before what these systems do can be more than a very expensive and inefficient worldwide proof of concept.

Controversial thought after reading this sentence: It’s cryptocoins all over again.

What are you expecting to break down? A modern internal combustion engine plus continuously variable transmission is one the world’s most complex pieces of engineering in common use today.

An electric drive system is trivial in comparison.

WinXP ran quite nicely on 4GB of RAM. You could boot to the desktop in under 5 seconds if you had a solid state drive.

All the day to day tasks someone needs to do with their computers haven’t materially changed in the last twenty years since XP held the crown. In that time the speed of processors, storage, gpus, and ram has increased by literal orders of magnitude.

And yet, somehow, things have become much shittier since then.

Once there is enough adoption, expect token costs to skyrocket and advertising enshitifaction to accelerate

Models on local hardware that give adequate performance and the availability of non-US models mean that path is doomed to failure. If they crank up the price then people will look elsewhere. While they might not get frontier-level performance out of the alternatives, there’s plenty of stuff out there that’s good enough for 1/4 to 1/10th of the cost. Or just the cost of power, if you’re willing to pay for some hardware and are operating on a small scale.

The only play the AI companies have got now is twofold :

• Regulatory capture, hence all the press releases of basic and frankly embarrassing sandbox failures and “we need to slow down” talk to try and get that rolling, and
• Banning of those nasty, evil, non-US models in the name of national security.

And then what? The rest of the world will just move on to the next most useful thing on the list. There will be ~330 million US citizens to extract wealth out of, to service trillions in data centre debt. That shit’s not gonna fly.