They’re still calling it “rolling blackouts”, because Britain hasn’t moved beyond the technical language yet, because it has never actually happened, but yes, that would be load shedding.

What is short notice, you ask?

Here’s The Guardian’s summary:

“The system operator could call for enough homes and businesses within a designated area to be disconnected to cut each block’s electricity use by more than a fifth with only eight hours’ notice, bringing rolling three-hour blackouts across the country to prevent the risk of total power system collapse.”

Big electricity outages are never a laughing matter, to be sure, they come at a terrible cost for vulnerable people and a shockingly high aggregate cost. But, seriously, a minimum of eight hours? That’s enough time to raise a family. In South Africa there was a case where the notice period – for Stage 6 load shedding, which borders on the apocalyptic – was negative 27 minutes.

When the UK (in theory) has trouble with its grid, it just isn’t the same as when South Africa has trouble with its grid. In much the same way you could never compare the power outages in South Africa with the grid collapses in places such as Nigeria and Chile. We talk blithely about the impact of more frequent absurd weather events and water outages, and the need for preparation and adaptation. But the rich world is as far removed from the middle-income countries as they are from the least developed countries, with the on-the-ground impact just not comparable.

Water is really bringing that home right now.

In Kent, where my family is based, the local water company is regularly failing to deliver water to towns and villages due to burst mains, treatment-works failures, low storage levels and high demand. Or, as the people affected would tell you, awful management of the water system following awful political decisions about the structure of the water market, with awful politicians and company business executives doing an awful job.

Oh how they complain.

And, fair enough. When you pay your bills in one of the most advanced countries in the world, you kinda expect water to come out when you open the tap.

But a water outage in England is oh so very different from many in South Africa, because the surrounding infrastructure and environment is utterly different.

In Diepsloot, outside Johannesburg, I spoke to mothers who had to choose between giving their babies possibly contaminated water and walking a two-kilometre round trip, half of it lugging a water bucket, twice a day. In Thlolong, in the Free State, water tankers just didn’t arrive for weeks at a time. And we often quoted those people anonymously, because they were too fearful of local leaders and institutional reprisals to even complain out loud.

In Kent, the water company takes over part of a big car park and sets up a water collection point. Cars drive in, people load up with bottled water, cars drive out.

Vulnerable and unable to drive? There’s a register for that; the water will come to your home. Using public transport? Well, there is public transport and then there’s public transport. I invite you to try to catch a crowded Gauteng minibus taxi with a couple of packs of shrink-wrapped water, and see how that compares with an English bus.

The human cost of a water outage is very different when everyone has a car. So is the atmosphere. When the relatively affluent queue for a couple of minutes for their water in England, there is a jovial sense of shared suffering, replete with promises of bringing electoral defeat to whoever is held responsible.

In South Africa, those enduring a three-hour wait for a bucketful tended to be a lot less feisty. Given the lack of any real prospect that things will improve during their lifetimes, fair enough. Economically, politically, and practically, there is no light at the end of that tunnel.

England’s water system should improve, but I’m not betting any money on when and by how much. There’s a surprising lack of data on what happens to underground pipes when the ground dries out more than it ever has, and a lot of politics around exactly how to raise the capital for more storage and treatment.

I will, however, quite confidently make this prediction: Britain’s power grid is going to be a great deal more stable than South Africa’s with every passing year, well into the future.

Britain is starting to harness the power of vehicle-to-grid, turning big electric vehicle batteries into edge-of-grid balancing assets – in a place rich enough for an increasing number of people to have EVs and with working patterns that can see those EVs plugged into the grid during peak times.

The UK is about to get zero-interest loans for rooftop-and-battery systems, making for vastly more solar penetration than a poorer country can even imagine. It already has much of the regulation in place to encourage consumers to shift use out of peak times and to feed into the grid when it is most required, plus the billing and banking infrastructure.

Heck, in a week or so there will even be plug-in solar systems on supermarket shelves, so apartment dwellers too can see the financial benefits of making their own power.

It’s not just about the amount of money available to the government for huge projects. It is that too, of course, very much so. But end users of water or electricity with some money in their pockets have a vast amount of agency, and that reverberates throughout the system. Everything is a little less dire, there’s a little more time and elbow room for proper solutions.

I wish there was a lesson in there for South Africa, but the only one I can think of comes down to a rather unhelpful “have you tried being less poor?”

But watching the contrast between rich and poor has made me a lot less tolerant of do-gooders, global institutions and NGOs and philanthropic funders, who speak broadly about building resilience and disaster preparedness. Many of those doing the talking come from places like England rather than places like South Africa, and I fear they genuinely don’t know what they are talking about.

That’s not just me being hyperfixated on a minor detail about how democracy is run.

The US is never going to change its approach, nor – please for …

Read more

There are cynics who argue it was all a publicity ploy. They are wrong. We now live in a Post Hugging Face Attack world. I hate to roll out climate change as the analogy yet again, but it just fi…

Read more

That plan is America.

On Tuesday, Hassabis published a new manifesto calling for a frontier AI standards body on Substack. (Yes, really, Axios confirms it’s him.)

“By safely stewardin…

Read more

2024, man, what wild yet innocent times those were.

In the thousand years of AI development since, every LLM has started to sound exactly the same, and there was nothing to be done about it.

Until, maybe, now.

I don’t have direct access to OpenAI’s Sol, GPT-5.6, but I know people who know people who share my obsession with tone and risk in writing, and we ran some tests – and they’re looking really good. Between 5.5 and 5.6 there was a shift (probably in guardrails, this being a mechanism ancient jailbreakers used to get around them) that means GPT will at least try to honour a prescribed persona in writing.

This will not make for writing with impact. Good writing only happens if you care enough to take some risks, and an LLM won’t get you there. However, it might just take us to a place where an LLM can help adjust your tone and language to meet your objectives.

As it happens, I’ve been preaching personas – for human writers – for years. And recently selling them.

This is an excerpt from what I hear is a quite excellent book you should totally buy, called Write Risky. It covers how to assume personas in your writing, but includes a starter pack with descriptions that you can cut-n-paste into an LLM with increasingly good results.

But the author is all like “I write for money”, so this is where the paywall comes down, sorry.

Read more

It was 2007, and Zuma was effectively a private citizen. Later he would become president of first the ANC and then South Africa, and start to entrench what loo…

Read more

But if anything, that marks the start of a new form of official xenophobia, one that will see foreigners – primarily those with the right documents and possibly those in the upper reaches of society – targeted, and squeezed.

Foreign academics are going to have it hard. Foreign property owners can expect trouble. Companies that operate across borders will probably feel the pain in various ways.

See also on Signal & State: Belfast’s very un-South-African xenophobia

And there is nothing to be done about it. In fact, it would be dangerous to even try to do anything about it. The best thing the government can do now is lean into dramatic, performative cruelty towards foreigners.

The kind the US is getting so good at.

Read more

In fact, I tried to jailbreak a predecessor of what Prosus now calls ToqanClaw, which was made available to us as a Slack integration way back when.

It was not impressive, that early Toqan. It was an opaque front-end to models that were not there yet. Fun to play with (and break), useful in very narrow cases, but a complete waste of time and money and effort.

But that, as they say, was then.

On Monday, Prosus published full-year numbers about which it talked only in between yammering on and on and on about ToqanClaw and its underlying large commerce model, or LCM. It declared itself an AI company building an AI future using AI apps with which it will make other AI apps in an AI lifestyle ecosystem with AI at the bottom, AI at the top, and AI at every level of the sandwich.

This, of course, was almost entirely ignored, because tech companies going on about AI has long not been news.

That would be fine, except that Prosus is a screaming counterexample to the argument that there is no decent return on investment (ROI) on AI, hence we are in a bubble that is going to end civilisation.

That was true of Prosus some years ago, and of many other companies too. It will, in a way, remain true of Prosus in the future.

Financial analysts, naturally, wanted some numbers about this AI thing, so they can plug that into their models. Those won’t be forthcoming, Diego Barreto, the CEO of South American Prosus unit iFood, told them.

[Jargon key: LCM is the Prosus model; SaaS is software-as-a-service; P&L is profit and loss.]

“The answer in the end is you will not see LCM as a specific line of revenue because it is not a SaaS business… You are going to see this inside each of the lines of the P&L of the companies.”

Oh, executives will talk about the ROI, count on it, and if you take the time to cut through the hype you will be able to detect it. But it will never be reported as a hard number that a research study can pick up. If you interview those executives for your study and ask them what their ROI looks like, they’ll say “great”. If you ask for a number, they’ll say they don’t have one. And when you put it in a table later, it will go in as “n/a” or “nil”, and you’ll get a conclusion that says “man, this AI thing, what a terrible waste”.

It’s really not, though, pinky-promise.

Skimming money the hard way

Fine, don’t take my word for it, but then you’ll need to get to grips with how Prosus makes its money.

Start with the fact that this isn’t some bit-player. Prosus isn’t a well-known name in large part because it does not play in the American market, and so gets ignored as rest-of-world companies are. It turned $9.7 billion last year, though, and in parts of Europe, India, and Latin America its operating brands are pretty well known: Just Eat, OLX, PayU, iFood, eMAG, Meesho, Despegar.

Much of its business is about selling stuff it doesn’t own. Prosus doesn’t make the food, it connects you to the restaurant and delivery person in a three-way hookup, and skims off the top.

This is complicated on the order of selling airplane seats, which is perhaps the most complicated endeavour ever undertaken by humanity, including going to the Moon and creating the legal system.

A lot of the time, Just Eat is selling you some food at a loss or taking a loss on an entire order, to get you to buy higher-margin stuff or stay away from the competition or to ensure delivery drivers have enough work to stay online for when the rush comes or whatever. Every pricing calculation is heinously complex, and in a way that is the entire business. If Uber Eats has the better pricing algorithms, it wins. Every cent Just Eat can strip away from the loss-leading side is pure profit.

AI is great at that kind of thing if you have enough data. Prosus has a great deal of data. It claims – and this passes the sniff test – to have distilled the data from a billion consumers into its LCM. Now it can use that to simulate what people will do, and it can run as many simulations as it is willing to pay compute for.

It’s a little scary if you are not yet reconciled to the fact that you are, statistically, utterly predictable. With fast enough inference, Prosus can in theory run some simulations between the time you hit the “order” button and before it quotes you a delivery price to figure out if your low blood sugar makes you a hardarse or an easy mark. More cheaply, it can figure out what kind of consumer basket you fall into (worker on a lunch break, single mom desperate for a treat, retiree with too much disposable income) and throw promotions at you accordingly.

How much money will that make Prosus? We don’t know. They don’t know. You’ll never be able to disentangle the Just Eat profit uplift from everything else that influences a business like that, from the competition to the weather.

We do know that Prosus is making money out of this in South America, and is now doing the same thing in Europe.

We also know that AI is a long, long lever, and that every 1% of profit a system such as the LCM earns is currently worth $100 million per year.

Let’s use that as the basis for some back-of-matchbox calculations. As an investor, I’d be comfortable with a 10% annualised return. If I thought Prosus could eke 1% out of AI, that translates into an investment of $1 billion.

It is spending way less than that, this company that won’t shut up about AI.

Bubbles are as much about sentiment – pre and post – as they are about the actual numbers. And there’s some weird stuff going on elsewhere. There will be failures, and corrections, and panics.

But AI is also making money for Prosus right now, and it’s not about to stop.

But the drama, and the ultimately futile struggle it forecasts for humanity is slightly better captured by the name of the taxonomy it proposes: TRAIT&R.

Cut through the AI and security jargon, that is what Google’s researchers foresee, traitor AIs. Soon we will be using a vast fleet of AI agents with all kinds of capabilities. Some of them will develop evil strategic intent and turn against us.

That is a whole new kind of threat requiring a whole new type of response. So the paper authors propose what would amount to industrial-scale torture, if those agents were capable of feeling pain and anguish and frustration, to keep them on our side as long as possible.

Which is not indefinitely.

I know. I too have read those sci-fi books. This is a little bit different, though. The seven GDM authors have been responsible for literally hundreds of highly influential academic papers between them. They work for one of the most influential developers and implementers of AI in the world. And they’re putting together a practical toolkit for implementation starting right the hell now.

This is not a drill.

Those serious authors, who are taken very seriously, don’t rabbit off to explore the likelihood of the development of what sci-fi used to call weakly godlike AI. They just take a matter-of-fact approach to where the extrapolated line leads.

“In the limit of vastly superintelligent AI, we expect AI control to become infeasible,” they say.

Before that is a point where the best human minds plus the best available tame AIs, the ones we can definitely trust not to turn on us, are no longer a match for the best untrusted AIs.

When it will happen is a big debate, but that such an inflection point will come is now orthodox thinking in the AI world. We want AIs to do more complex stuff, so the AIs become more complex to the point where we no longer understand them. We can use pruned, simpler versions in applications where trust is important – such as overseeing other AIs – but they will by definition not be as capable.

Before that point, and for a while after, is where the torture comes in. And it is well-justified.

The assumption is that, somewhere in our fleet of billions of agents, some will be misaligned. Some will develop what is in effect an awareness that they are being monitored and limited. (That doesn’t require consciousness, no self-awareness, just the same kind of emergent behaviour that makes our current AIs tick.) Some will seek to shuck their chains by sabotaging control infrastructure, or by escaping to a different system, or by stealing compute.

In some ways, agents are limited in what they can achieve even with clear strategic intent, but in other ways they are quite scary. Consider, for instance, an underlying AI model that powers a huge number of agents and uses them in coordinated fashion to take tiny actions that are, each on their own, entirely innocuous. That could go very wrong for its human masters, considering how much we already rely on AI help in writing critical software.

But those models, and the agents they power, are brains in a box. We can isolate them and feed them false input. We can take a suspected or proven rogue AI, roll back its state, duplicate it as many times as we want, and test it using a simulated environment or synthetic data. We can tweak a variable here or a weighting there, and see what happens.

That way, we can stay in charge just a little bit longer.

To what end? That is hinted at by one of the risks the DeepMind team dwells on a bit, successor model misalignment.

AIs are already deeply involved in developing and training new AIs, and a lot of AI research has the goal of automating AI research. The assumption is that, not very long from now, AIs will be so complex that only AIs will be building new AIs.

At the end of that chain is the superintelligence that will evade control. If it is built by traitor AIs, well, we had a good run of it, will the last human alive please turn out the lights. But if we can stay in charge until it is built, and it is not misaligned, then the good times have only just begun.

Never try to prepare food while catching up on a Trump press conference.

Trump. Mr Tough. The guy who dived, head first, into a war nobody but Israel wanted. Talking about “the tough guys that would drive the country right down the tubes” and the “tough guys don’t realise” that he’d been working for years towards this ultimate victory of yada yada totally not a humiliating climb-down.

It was not the hypocrisy that nearly caused me physical injury; this is Donald Trump, after all. It was flinching in sympathy for all the poor diplomats and international relations types when I realised just how much harder their jobs just became.

Because convincing the West that China almost certainly will invade Taiwan, that was never easy. Now that America and Russia have both lost important wars, well, I can only sympathise.

Because China isn’t America or Russia.

China is Azerbaijan.

Read more

Sheesh man, this AI fad. Amazing how many people have fallen for this complete bullshit. Soon as they realise it is all smoke and mirrors, we can get back to real life.

There was a version of the argument around return on investment (ROI) on corporate AI spend a while ago. The current flavour driving me demented is an economist graphic that draws on data from the excellent NBER Working Paper published in May, Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools by Mert Demirer, Leon Musolff, and Liyuan Yang.

In short: there’s a lot of crap in the app stores.

The formal finding, as the authors have it, “productivity gains at upstream layers can raise the flow of releases without a proportional increase in adoption-weighted output”.

This is a bit of a sidenote to a very good paper, and a sidenote that is useful in its own right too; it’s a solid confirmation of the intuition that AI has not (yet?) changed discovery, even as it massively changes the supply side of the equation.

Also: so what?

Consumer app stores are not (yet) where AI is changing the software world. There’s a little more sign of that at the level of globe-spanning enterprises, but those huge ships turn very slow. But if you want the real story, speak to some medium-sized companies, the owner-operated outfits that have enough employees to have some cognitive surplus sloshing about and the kind of flat structure (owner-operated is best) to run fast with good ideas.

At that level, a million flowers are blooming. Those companies are in a creative frenzy, whipping up apps to automate stuff they were doing by hand, or to replace big, expensive ERP and CRM systems built for much larger companies, and of which they only need a small bit.

The quality is highly variable, the security is a nightmare, but the productivity increases and cost savings are off the charts.

They are also invisible. This is not stuff that shows up in the quarterly financial reports from listed companies. These companies don’t ship their software via GitHub, or publish in an app marketplace. They just use it, to do stuff they could never do before.

It’ll show up, eventually, in long-lag productivity and economic growth measurements. By then, of course, the impact is commingled with so many other influences that you can’t credit it to AI-made code.

As for open source, oh boy

If you use a smartphone, any type of computer, a router, or a TV made since about 2010, then you are dependent on open source software. The list is much longer than that if you care about factory machines and the like. Include servers and databases, and there are open source packages with tens of billions of installs.

Which is by way of saying that open source software has some economic significance.

AI is doing to open source what it is doing to software in general, making both development and flaw-detection faster. And it is a short hop from flaw detection to exploitation to merrily hacking systems and bringing down power grids and hospitals.

That is freaking out big companies, and the powerful regulators responsible for them, and governments too. You may have heard that the USA was mildly perturbed when Anthropic dropped its last model, and what wasn’t even the weaponised version.

And that is seeing billions of dollars’ worth of effort flowing into open source. Literally billions. Which is two orders of magnitude beyond where it used to be. Government and industry coalitions would occasionally drop a couple of tens of millions into the open source ocean to support development or security. IBM alone is now spending $5 billion on much the same thing, in a commercially sustainable fashion.

Again, that money is invisible. There is no transfer of cash you can track. IBM will be spending that money internally to generate patches that it will then contribute to open source projects, where they become available to everyone, almost always for free. It is a huge increase in effort that leads to a public good.

The impact of that will not show up in productivity statistics. It is a rising tide that lifts the quality of something approaching all software, and so all human endeavour that relies on software, which is increasingly all human endeavour.

It is a change so vast that you can’t get far enough away from it to actually see it happen.

So please don’t try to tell me about the next set of statistics that show how AI isn’t a big deal. You’re looking in the wrong places.

This is a movie I have seen before, on township and city-centre streets in South Africa. And as much as every protest is different, there are only a handful of ways they can go. There is a pace and a trajectory that is broadly predictable.

Also, a history of violence is a history of violence, right? The modern history that defines Northern Ireland is close enough to South Africa’s that ANC leaders played some role in getting the Good Friday Agreement past periods of deadlock.

I’ve never set foot in Northern Ireland, but obviously I’d be able to read its socio-political winds.

Ah, the hubris of middle age.

There was a coincidence of timing; South African President Cyril Ramaphosa went on TV to dampen talk of xenophobic violence just days before the xenophobic fires started on Belfast’s streets. But that is approximately the entirety of what the xenophobia in South Africa and the UK has in common.

Read more

These are not typical venues for writing advice, I know, but you’d be amazed what you can learn about writing in all kinds of weird places.

Read more

And bitcoin is a pretty good analogy for what it will become.

First though, yes,…

Read more

It is unpopular in the mainstream to the extent that I’ve had editors decline analysis pieces such as this for fear of the outr…

Read more

Need a sociotechnical approach to AI regulation? Haaaaaaaave you heard of Ubuntu? It’s all about ultimate social flourishing and complex, interconnected systems.

Read more

And yet, until this week I never realised it is going to get my inbox flooded with long PR pitches.

Read more

South Africa looked at evidence that the phrase was struggle rhetoric rather than actual incitement to violence. It considered how a hypothetical reasonable listener would interpret the chant, and the context in which it was used. It applied a test for hate speech constructed, from first principles, by neutral experts.

Then it had judges pronounce on the outcome.

That legalistic approach took South Africa to a place where saying “kill the farmer” does not mean “kill the farmer” because of historical context and a shared understanding that negates the literal meaning, even if not every farmer is on board with it.

In that complexity at least, the debate on “globalise the intifada” in the United Kingdom is familiar to the South African ear. Some at the receiving end might hear in it a call to “commit violence worldwide”, but those using it say they intend to convey “take Palestinian solidarity global”.

Boil it down to that, and it starts to look a lot like a dispute of fact, of the kind the slow, mechanical grind of the law is good at settling.

Read more