Well, I can't speak for the US electoral roll breakdown, but it appears you may be in the 10% that can't add up. :)

When I had Comcast in the Bay Area, it randomly dropped one day because they installed the line to the house at a height that a semi driving through the neighborhood likely clipped the line. I was attempting to diagnose the issue from within the house, until I discovered the severed cable laying in the middle of the street.

That's over simplifying things a bit, but yeah, we have fundamental infrastructure issues in the UK with any proposed mass rollout of AI DCs. While we do have surplus generating capacity, to the point that we occasionally need to pay people to use it, and more generating/storage capacity (especially local domestic, which frees up baseload capacity for industrial uses) is being added all the time, there are some other more fundamental UK issues. Firstly, the national power distribution grid is in need of some serious upgrades to support some of this. The South West in particular is on a very thin high tension thread that currently has insufficient capacity for the the planned relocation of the UK's nuclear deterrent subs to Devonport and associated growth, for instance. Secondly, and this seldom gets mentioned in the context of all these new AI DCs that are being planned or the "shouty kids/let's save water" adverts currently being run, is that we have some severe water supply pressures as a result of climate change combined with a decades long lack of investment in new reservoirs and upgrading leaky Victorian era distribution networks. Care which part of the country is currently experiencing the brunt of a "once in several decades (but is now expected to happen every few years)" severe drought event? Yup, it's the home counties, where most of the planned DCs are supposedly going to get built!

Those all have technical/construction solutions though, the third does not, and that is that the home counties in particular are notorious for the NIMBYism, although they are far from alone in that - just a few days ago a massive planned solar/storage facility in Scotland got pushed back, and possibly nixed, because of it. Sure, there are totally valid arguments against concreting over large expanses of what is currently greenbelt to some extent, even before you get to water/power shortages, pollution (including noise), additional traffic, and all the other impacts, but there have to be some limits. One solution I like for this (albeit not without some quite complex zoning issues due to suitable areas for building supply not being equal to where the supply is needed) is to make supply, or cost of supply, proportionate to acceptance of local capacity build outs. If you want your power to be generated in someone else's backyard, then you'll pay more for the privilege while they get a rebate. If you object to that proposed new local reservoir, then you don't get to complain if you need to start buying potable water in bulk during the next drought. If, as in the case of DCs, you are putting in disproportionately high demands for supply, then you either pay - up front - for a suitable proportion of the required upgrades, or commit to some other form of reimbursement with suitable early termination clauses if you want a longer timescale.

Then again, this *could* just be Microsoft wriggling off the hook because they're starting to think the AI bubble might be somewhat overinflated and therefore they might not be needing all that capacity they've been saying they are going to build afterall, but don't want to start the stampede for the exits until they're already through the door.

Damnit! UCT != CUT either, which is yet *another* standard timezone, using the same three letters. Cue oblig. XKCD on standards...

Above should have used CUT, not UCT.

It's more likely to be much ado about nothing for almost everyone apart from a few astronomers, rocket scientists, and similar groups, all of which you'd expect would have the braincells to figure it out, but will no doubt have a few in the mix that don't. Other standard timezones exist (the various UTCs are != UCT), so if your system is working fine on one of those then it will continue to do so.

UCT also moves backwards *and* forwards continually, mostly due to gravitational variances due to other celestial objects, changes in the Earth's spin, and all the other things that effect planetary orbits; some years Earth (and all the other things going around the sun) will orbit a little faster, others a little slower, so it's not a cumulative year on year increase to the delta, although there can be a run in one direction or the other for a while. Assuming it averages out OK, then there's a distinct possibility that by the time an adjustment really is required no one will be left around to care about it, let alone do it.

The general expectation in industry is that the number of bugs must be finite and, even though every change might introduce new ones, as the AI models find new ones and they get fixed the total number of yet to be be found bugs will come down. As with Microsoft's recent run of consecutive record breaking Patch Tuesdays, the conclusion being drawn is that sooner or later the number of bugs being found and fixed after each analysis session will start to fall and, subsequent to that, normal patching service, albeit on a hopefully much more robust codebase, will resume.

Where it gets a little more sticky is that, while finite, the number of available bugs to find and fix could still be extremely large, so the time required to regress back to previous, or even better than previous, bugfix development volumes is unknown. If your bet is that duration is quite short (less than a year, say), then why would you want to encumber your business with a raft of new full time employees all wanting things like health benefits and severance protections, when you can just bring in some contractors and dump them at will when they are no longer needed? If you're thinking it's a long game, then at least some full time employees is probably a better option because eventually they will be cheaper than using contractors, with how long that is dependant on your typical salaries for each. BoA is presumably betting on the former scenario based on what they understand of their codebase and whatever additional volume of bugs AI has found in it so far.

Using AI, especially. One thing that struck me in the last few days was in relation to comments on the story about the person who has used AI to vibe-code clones of part of the Adobe graphical tools suite and has since (within a few days!) gone on to close the core Microsoft Office apps in the same way. Apparently (I have not checked it personally) this software, albeit still in "beta", runs like a bag of spanners and while, it might superficially look like the tools it's emulating, lacks the depth of many features and some are missing entirely. The code is also impenetrable, so the "it's open source, the community can check for malware" argument is even less practicable than it usually is. It's an interesting demo of where things are currently at, but ready to go, drop in replacements, they are not.

Sure, there was the obvious flame war about the pros and cons of "AI", implications for copyright, potential to be sued, etc., but what was almost completely absent was any kind of reality check about how coding works in practice when a PHB is in charge. You know, the kind that totally sees the OpEx of some license or other going out, has read about this AI-driven vibe-coding stuff, is *totally* on board with the concept of "ship it and let the users find the bugs", joins the dots, and tasks some hapless AI-literate (that's what the CV said, anyway) intern on vibe-coding a replacement. You just know that's either going to start happening Real Soon Now, or already is, in all sorts of places that should know a LOT better than that, especially given the issues with the code shat out in the examples above.

Many of them, and quite likely many of their clients/users too, are going to get quite badly burnt, I expect.

And if a rogue AI agent actually *causes* one of those more damaging attacks, what then? "To err is human, but to really fsck things up takes a computer" and all that?

What seems to be missing here isn't just any clarity in legislation about liability for AI agents, especially labs running frontier models who seem to get free legal passes that would absolutely not be afforded to regular plebs who - potentially entirely by accident - cause an agent to hack someone else, but any serious discussion about putting those legal frameworks in place is also lacking. Seriously, WTF is up with that? How many incidents that are clearly in breach of existing computer misuse legislation do we need to get some action here, either under existing laws (including, potentially, civil law) or to start working on tweaks to existing laws to provide some specific clarity relating who is liable for any misuse of AI agents?

I get the AI labs almost certainly have a tacit "don't sue us, and we won't sue you" agreement in place for things like this to try and avoid regulation they definitely do not want because their actions are quite obviously not aligned with the regulation they are claiming they want put into place. Third parties like South Korean banks are under no such terms however, and given they are in the financial sector (or health sector in another recent example) probably have legal obligations of their own as well, so again - why are we not even hearing of any talk of legal action?

Maybe we could just cut to the chase and go straight to Plaid Intelligence? It's not like anyone outside 1600 Pennsylvania is taking this "rebranding" seriously anyway, so why not have some fun with mocking it while we can?

At least the voters are not talking about things like Epstein, Iran, and the price of gas, which is likely the whole point; look at all the stupid, forget about the serious, and (for ~50% of the them) be sure to tick the name with "R" next to it next month like you're programmed to.

That was a new one for me too. My guess was what we used to refer to as "red tops" in the UK; the tabloid / gutter press because they typically have a red banner with the name of the paper on the front page, although some of the worst modern examples of that (the Mail & Express) actually don't have a red banner. In the mostly online world, that obviously needs to cover online-only and broadcast media forms too so "red top" seems to be falling out of use.

Apparently though, it's actually a reference to totally fake or low-quality news sites, masquerading as legitimate outlets purely to push misinformation and is named for the "Mystery Meat" by-product that resembles pink slime used as filler in some processed meats (mmm, yummy!) Defining characteristics include:

• The content is generally produced by low-wage employees, automated content production (e.g. AI), and templates. It might just list information without any analysis or context.
• Purport to cover local or hyperlocal news and to some extent are taking advantage of a decline in traditional local news.
• Many outlets are part of networks financed by partisans to push their narrative or point-of-view.

If you're thinking that mostly means the likes of Reach's portfolio of garbage and so-called "independent journalist" sites, take a look at the default Microsoft Edge homepage on MSN, which would totally qualify. It's clearly largely AI-generated, highly formulaic/template based, seldom provides any analysis or context, pretends to be local (because clickbait sites that include a country's name in them are *totally* relevant to another country entirely), and, yup, is financed by someone with a definite narrative to push. It clearly only exists to get people to doomscroll and generate clicks and ad-revenue. Bingo! Full house!

Heads *need* to roll though, but I'm not sure anyone with a say in the matter at the Pentagon realises that, let alone why is really does need to be the case. The US' own judiciary has been pretty clear on this; whenever a lawyer has been caught using AI-generated submissions that contain hallucinations some form of censure has ensured PDQ. It's similar in industry as well; multiple instances of employees using AI without checking the facts that have subsequently been either disciplined, fired, or (if they are senior enough and the screw-up warrants it) issued a golden parachute.

Now, look at the situation on the ground. This school has been there for around a decade, and it is *very* obviously a school - pink and blue painted play areas, with assembly areas and sports facilities suitable for kids; that is totally NOT a military PT facility within the larger complex. Sure, it's surrounded by IRGC facilities, but absolutely no military anywhere else in the world co-locates their military's family accommodation blocks and support facilities close to the operations blocks and facilities to make it easier for their staff & troops, right? Oh, wait, yes, they do, don't they? Not always, sure, but often enough that the US response here matters, because if they try and sweep this under the rug "because AI", then they just made all their similar facilities viable targets "because AI". There's a reason everyone else, including US allies, were quick to publically condemn this, and even say it might be a war crime because of failure to properly assess the strike's potential for collateral damage, and that's to make it absolutely clear they do not consider this have acceptable levels of collateral damage, let alone be a legitimate target.

Don't forget, the IRGC tend to be fanatics even by Iranian standards, and are often either willing to strap on a suicide vest or can motivate others to do it for them. 123 dead kids means a *LOT* of bereaved parents, brothers, sisters, and other family members, many of whom are probably already linked to the IRGC, who are going to need a lot less motivation to do something a little more radical than just march along the street waving flags and chanting "Death to America!" the next time they are asked to do something for their country. What the US does here is totally going to be interpreted as setting precedent by Iran, and quite possibly by other countries and groups within their axis, should it come to a conflict or just an opportunistic attempt at some terrorism.

I think there's a bit of a difference though. Regardless of hat colour, most hackers have going to have at least some sense of what the likely repercussions of their actions might be and, even when those standards are very low indeed, are probably not going to think doing something that could cause serious loss of life "for the Lulz" is a good life option. They might make a mistake here and there, but how many times have you read about some random hacker *deliberately* doing something that would definitely cause loss of life in full knowledge that it was going to happen outside of on the periphery of an actual conflict?

GANs on the other hand have no such moral compass, sense of self-preservation, or keen interest in avoiding lengthy spells in government run institutions where they'll in all likelihood get to make some interesting new "friends" (unless it's a web forum they hijacked in Germany, apparently). They'll just carry on trying to complete whatever poorly defined objective they have been set, within whatever limited / technically absent guardrails they can't circumvent - the pattern here seems very much that meeting the objective has more weight in the models than any guardrails. If a GAN somehow computes that global thermonuclear war is a required step to its objective of having a nice game of chess, then you'd better hope it warms up with a LOT of Tic-Tac-Toe.

My late wife did undergraduate research in photosystem II where there was quantum effects back in 2002. She did her PhD thesis on proton coupled electron transfer (PCET) in soybean lipoxogenase, then had a long career developing models for Fortran code that did quantum simulation of proteins (she was behind Tagdock in Amber MD). This is nothing new; several biological systems have quantum effects that are central to their mechanism.

Understood, and yeah single tasking seems like a good workaround given the situation. Or switching to an LLM with a larger context window, if possible/practical. I was able to get about 7X more context window "runway" using the AST approach, but it's a relatively small code base to begin with.

Sounds like you're running up against the context window of whatever LLM you're using. Try converting your code to a minified abstract syntax tree first and pass that to your LLM, and ask it what full text it needs in addition to that. It should stay within the context window much better that way.